LongBench v2 / 66f53892821e116aacb33225 / Which of the following methods is used in this article to calculate the…

BERTSCORE: EVALUATING TEXT GENERATION WITH
BERT
ABSTRACT
We propose BERTSCORE, an automatic evaluation metric for text generation.
Analogously to common metrics, BERTSCORE computes a similarity score for
each token in the candidate sentence with each token in the reference sentence.
However, instead of exact matches, we compute token similarity using contextual
embeddings. We evaluate using the outputs of 363 machine translation and image
captioning systems. BERTSCORE correlates better with human judgments and
provides stronger model selection performance than existing metrics. Finally, we
use an adversarial paraphrase detection task to show that BERTSCORE is more
robust to challenging examples when compared to existing metrics.
1
INTRODUCTION
Automatic evaluation of natural language generation, for example in machine translation and caption
generation, requires comparing candidate sentences to annotated references. The goal is to evaluate
semantic equivalence. However, commonly used methods rely on surface-form similarity only. For
example, BLEU (Papineni et al., 2002), the most common machine translation metric, simply counts
n-gram overlap between the candidate and the reference. While this provides a simple and general
measure, it fails to account for meaning-preserving lexical and compositional diversity.
In this paper, we introduce BERTSCORE, a language generation evaluation metric based on pre-
trained BERT contextual embeddings (Devlin et al., 2019). BERTSCORE computes the similarity
of two sentences as a sum of cosine similarities between their tokens’ embeddings.
BERTSCORE addresses two common pitfalls in n-gram-based metrics (Banerjee & Lavie, 2005).
First, such methods often fail to robustly match paraphrases. For example, given the reference peo-
ple like foreign cars, BLEU and METEOR (Banerjee & Lavie, 2005) incorrectly give a higher score
to people like visiting places abroad compared to consumers prefer imported cars. This leads to
performance underestimation when semantically-correct phrases are penalized because they differ
from the surface form of the reference. In contrast to string matching (e.g., in BLEU) or matching
heuristics (e.g., in METEOR), we compute similarity using contextualized token embeddings, which
have been shown to be effective for paraphrase detection (Devlin et al., 2019). Second, n-gram mod-
els fail to capture distant dependencies and penalize semantically-critical ordering changes (Isozaki
et al., 2010). For example, given a small window of size two, BLEU will only mildly penalize
swapping of cause and effect clauses (e.g. A because B instead of B because A), especially when
the arguments A and B are long phrases. In contrast, contextualized embeddings are trained to
effectively capture distant dependencies and ordering.
We experiment with BERTSCORE on machine translation and image captioning tasks using the
outputs of 363 systems by correlating BERTSCORE and related metrics to available human judg-
ments. Our experiments demonstrate that BERTSCORE correlates highly with human evaluations.
In machine translation, BERTSCORE shows stronger system-level and segment-level correlations
with human judgments than existing metrics on multiple common benchmarks and demonstrates
∗Equal contribution. † Work done at Cornell.
1
arXiv:1904.09675v3  [cs.CL]  24 Feb 2020


Published as a conference paper at ICLR 2020
strong model selection performance compared to BLEU.
We also show that BERTSCORE is
well-correlated with human annotators for image captioning, surpassing SPICE, a popular task-
specific metric (Anderson et al., 2016).
Finally, we test the robustness of BERTSCORE on
the adversarial paraphrase dataset PAWS (Zhang et al., 2019), and show that it is more ro-
bust to adversarial examples than other metrics.
The code for BERTSCORE is available at
https://github.com/Tiiiger/bert_score.
2
PROBLEM STATEMENT AND PRIOR METRICS
Natural language text generation is commonly evaluated using annotated reference sentences. Given
a reference sentence x tokenized to k tokens ⟨x1, . . . , xk⟩and a candidate ˆ
x tokenized to l tokens
⟨ˆ
x1, . . . , ˆ
xl⟩, a generation evaluation metric is a function f(x, ˆ
x) ∈R. Better metrics have a higher
correlation with human judgments. Existing metrics can be broadly categorized into using n-gram
matching, edit distance, embedding matching, or learned functions.
2.1
n-GRAM MATCHING APPROACHES
The most commonly used metrics for generation count the number of n-grams that occur in the
reference x and candidate ˆ
x. The higher the n is, the more the metric is able to capture word order,
but it also becomes more restrictive and constrained to the exact form of the reference.
Formally, let Sn
x and Sn
ˆ
x be the lists of token n-grams (n ∈Z+) in the reference x and candidate
ˆ
x sentences. The number of matched n-grams is P
w∈Sn
ˆ
x I[w ∈Sn
x], where I[·] is an indicator
function. The exact match precision (Exact-Pn) and recall (Exact-Rn) scores are:
Exact-Pn =
P
w∈Sn
ˆ
x I[w ∈Sn
x]
|Sn
ˆ
x|
and
Exact-Rn =
P
w∈Sn
x I[w ∈Sn
ˆ
x]
|Sn
x|
.
Several popular metrics build upon one or both of these exact matching scores.
BLEU
The most widely used metric in machine translation is BLEU (Papineni et al., 2002), which
includes three modifications to Exact-Pn. First, each n-gram in the reference can be matched at
most once. Second, the number of exact matches is accumulated for all reference-candidate pairs in
the corpus and divided by the total number of n-grams in all candidate sentences. Finally, very short
candidates are discouraged using a brevity penalty. Typically, BLEU is computed for multiple values
of n (e.g. n = 1, 2, 3, 4) and the scores are averaged geometrically. A smoothed variant, SENT-
BLEU (Koehn et al., 2007) is computed at the sentence level. In contrast to BLEU, BERTSCORE is
not restricted to maximum n-gram length, but instead relies on contextualized embeddings that are
able to capture dependencies of potentially unbounded length.
METEOR
METEOR (Banerjee & Lavie, 2005) computes Exact-P1 and Exact-R1 while allowing
backing-off from exact unigram matching to matching word stems, synonyms, and paraphrases. For
example, running may match run if no exact match is possible. Non-exact matching uses an external
stemmer, a synonym lexicon, and a paraphrase table. METEOR 1.5 (Denkowski & Lavie, 2014)
weighs content and function words differently, and also applies importance weighting to different
matching types. The more recent METEOR++ 2.0 (Guo & Hu, 2019) further incorporates a learned
external paraphrase resource. Because METEOR requires external resources, only five languages
are supported with the full feature set, and eleven are partially supported. Similar to METEOR,
BERTSCORE allows relaxed matches, but relies on BERT embeddings that are trained on large
amounts of raw text and are currently available for 104 languages. BERTSCORE also supports
importance weighting, which we estimate with simple corpus statistics.
Other Related Metrics
NIST (Doddington, 2002) is a revised version of BLEU that weighs each
n-gram differently and uses an alternative brevity penalty. ∆BLEU (Galley et al., 2015) modifies
multi-reference BLEU by including human annotated negative reference sentences. CHRF (Popovi´
c,
2015) compares character n-grams in the reference and candidate sentences. CHRF++ (Popovi´
c,
2017) extends CHRF to include word bigram matching. ROUGE (Lin, 2004) is a commonly used
metric for summarization evaluation. ROUGE-n (Lin, 2004) computes Exact-Rn (usually n = 1, 2),
while ROUGE-L is a variant of Exact-R1 with the numerator replaced by the length of the longest
common subsequence. CIDER (Vedantam et al., 2015) is an image captioning metric that computes
2


Published as a conference paper at ICLR 2020
cosine similarity between tf–idf weighted n-grams. We adopt a similar approach to weigh tokens
differently. Finally, Chaganty et al. (2018) and Hashimoto et al. (2019) combine automatic metrics
with human judgments for text generation evaluation.
2.2
EDIT-DISTANCE-BASED METRICS
Several methods use word edit distance or word error rate (Levenshtein, 1966), which quantify
similarity using the number of edit operations required to get from the candidate to the refer-
ence. TER (Snover et al., 2006) normalizes edit distance by the number of reference words, and
ITER (Panja & Naskar, 2018) adds stem matching and better normalization. PER (Tillmann et al.,
1997) computes position independent error rate, CDER (Leusch et al., 2006) models block reorder-
ing as an edit operation. CHARACTER (Wang et al., 2016) and EED (Stanchev et al., 2019) operate
on the character level and achieve higher correlation with human judgements on some languages.
2.3
EMBEDDING-BASED METRICS
Word embeddings (Mikolov et al., 2013; Pennington et al., 2014; Grave et al., 2018; Nguyen et al.,
2017; Athiwaratkun et al., 2018) are learned dense token representations. MEANT 2.0 (Lo, 2017)
uses word embeddings and shallow semantic parses to compute lexical and structural similarity.
YISI-1 (Lo et al., 2018) is similar to MEANT 2.0, but makes the use of semantic parses optional.
Both methods use a relatively simple similarity computation, which inspires our approach, including
using greedy matching (Corley & Mihalcea, 2005) and experimenting with a similar importance
weighting to YISI-1. However, we use contextual embeddings, which capture the specific use of
a token in a sentence, and potentially capture sequence information. We do not use external tools
to generate linguistic structures, which makes our approach relatively simple and portable to new
languages. Instead of greedy matching, WMD (Kusner et al., 2015), WMDO (Chow et al., 2019), and
SMS (Clark et al., 2019) propose to use optimal matching based on earth mover’s distance (Rubner
et al., 1998). The tradeoff1 between greedy and optimal matching was studied by Rus & Lintean
(2012). Sharma et al. (2018) compute similarity with sentence-level representations. In contrast, our
token-level computation allows us to weigh tokens differently according to their importance.
2.4
LEARNED METRICS
Various metrics are trained to optimize correlation with human judgments. BEER (Stanojevi´
c &
Sima’an, 2014) uses a regression model based on character n-grams and word bigrams. BLEND (Ma
et al., 2017) uses regression to combine 29 existing metrics. RUSE (Shimanaka et al., 2018) com-
bines three pre-trained sentence embedding models. All these methods require costly human judg-
ments as supervision for each dataset, and risk poor generalization to new domains, even within a
known language and task (Chaganty et al., 2018). Cui et al. (2018) and Lowe et al. (2017) train a
neural model to predict if the input text is human-generated. This approach also has the risk of being
optimized to existing data and generalizing poorly to new data. In contrast, the model underlying
BERTSCORE is not optimized for any specific evaluation task.
3
BERTSCORE
Given a reference sentence x = ⟨x1, . . . , xk⟩and a candidate sentence ˆ
x = ⟨ˆ
x1, . . . , ˆ
xl⟩, we use
contextual embeddings to represent the tokens, and compute matching using cosine similarity, op-
tionally weighted with inverse document frequency scores. Figure 1 illustrates the computation.
Token Representation
We use contextual embeddings to represent the tokens in the input sen-
tences x and ˆ
x. In contrast to prior word embeddings (Mikolov et al., 2013; Pennington et al.,
2014), contextual embeddings, such as BERT (Devlin et al., 2019) and ELMO (Peters et al., 2018),
can generate different vector representations for the same word in different sentences depending on
the surrounding words, which form the context of the target word. The models used to generate
these embeddings are most commonly trained using various language modeling objectives, such as
masked word prediction (Devlin et al., 2019).
1We provide an ablation study of this design choice in Appendix C.
3


Published as a conference paper at ICLR 2020
Reference
the weather is 
cold today
Candidate
it is freezing today
Candidate
Contextual
Embedding
Pairwise Cosine
Similarity
RBERT = (0.713⇥1.27)+(0.515⇥7.94)+...
1.27+7.94+1.82+7.90+8.88
<latexit sha1_base64="OJyoKlmBAgUA0KDtUcsH/di5BlI=">ACSHicbZDLatAFIaPnLRJ3JvTLrsZYgoJAqFxGqwsCqal0FVJQ5wELCNG41EyZHRh5ijECL1EnqAv02X2eUZsumipXR6Mj2Ipf+MPDznXM4Z/64UNKg7187raXlR49XVtfaT54+e/6is/7y0OSl5mLIc5Xr45gZoWQmhihRieNC5bGShzFZx+a+tG50
Ebm2QFOCzFO2UkmE8kZWhR1ov2oClFcYPX+4/5BXZN3JEw049Wm7/XpdogyFYZQr9fci3aoTsL1Pd23265oZrkaOqaXAb5FIv6DXOdwMvCOqo0/U9fyby0NCF6Q52/15+BYC9qHMVTnJepiJDrpgxI+oXOK6YRsmVqNthaUTB+Bk7ESNrM2aPGVezIGryxpIJSXJtX4ZkRm9PVCw1ZprGtjNleGru1xr4v9qoxCQYVzIrShQZny9KSkUwJ02qZCK14Kim1jCupb2V8FNmc0SbfduGQO9/+aE57HnU9+gX2h18hrlW4TVswCZQ6MAPsEeDIHDN7iBn/DL+e78cH47f+atLWcx8wruqNX6B8dUrVw=</latexit>
<latexit sha1_base64="RInTcZkWiVBnf/ncBstCvatCtG4=">ACSHicbZDPShxBEMZ7Nproxugaj14al4AyMEyvyoyHwGIQPImKq8LOMvT09mhjzx+6a0KWYV4iL5EnySXH3HwGLx4U8SDYs7sHo/mg4eNXVT1F+VSaHDda6vxbmb2/Ye5+ebHhU+LS63lz6c6KxTjPZbJTJ1HVHMpUt4DAZKf54rTJL8Lr6VtfPvnOlR
ZaewCjng4RepCIWjIJBYSs8DsA+A8od/eOT6oKf8VBrCgr13HI5sBiIRrTJyOt2EbtE2p8hzdrY27EAOM9BVWTfYNbKJ43dq59q+4/tV2Gq7jsWfmvI1LS7O08/f3nLi4dh628wzFiR8BSYpFr3iZvDoKQKBJO8agaF5jlV/SC941NqTlmUI6DqPAXQ4Y4zpR5KeAxfTlR0kTrURKZzoTCpX5dq+H/av0CYn9QijQvgKdsiguJIYM16nioVCcgRwZQ5kS5lbMLqnJEUz2TRMCef3lt+a04xDXIUek3T1AE82hVbSG1hFBHuqifXSIeoih3+gG3aF76491az1Yj5PWhjWdWUH/qNF4BkPYrbk=</latexit>
<latexit sha1_base64="RInTcZkWiVBnf/ncBstCvatCtG4=">ACSHicbZDPShxBEMZ7Nproxugaj14al4AyMEyvyoyHwGIQPImKq8LOMvT09mhjzx+6a0KWYV4iL5EnySXH3HwGLx4U8SDYs7sHo/mg4eNXVT1F+VSaHDda6vxbmb2/Ye5+ebHhU+LS63lz6c6KxTjPZbJTJ1HVHMpUt4DAZKf54rTJL8Lr6VtfPvnOlR
ZaewCjng4RepCIWjIJBYSs8DsA+A8od/eOT6oKf8VBrCgr13HI5sBiIRrTJyOt2EbtE2p8hzdrY27EAOM9BVWTfYNbKJ43dq59q+4/tV2Gq7jsWfmvI1LS7O08/f3nLi4dh628wzFiR8BSYpFr3iZvDoKQKBJO8agaF5jlV/SC941NqTlmUI6DqPAXQ4Y4zpR5KeAxfTlR0kTrURKZzoTCpX5dq+H/av0CYn9QijQvgKdsiguJIYM16nioVCcgRwZQ5kS5lbMLqnJEUz2TRMCef3lt+a04xDXIUek3T1AE82hVbSG1hFBHuqifXSIeoih3+gG3aF76491az1Yj5PWhjWdWUH/qNF4BkPYrbk=</latexit>
<latexit sha1_base64="fGWl4NCvlvtMu17rjLtk25oWpdc=">ACSHicbZBLS+RAFIUrPT7bVzsu3RQ2ghIqVbpuBgQRZiVqNgqdJpQqa5oYeVB1Y1ME/Lz3Lic3fwGNy6UwZ2VNgtfBwoO372Xe+uEmRQaXPef1fgxMTk1PTPbnJtfWFxqLf812muGO+xVKbqMqSaS5HwHgiQ/DJTnMah5BfhzUFVv7jlSos0OYNRxgcxv
UpEJBgFg4JWcBoUPvA/UOwfnp6VJf6F/UhRVmy4Tpds+SBirjFxOt1N26AdslOjrO7vWn7cpiCLouqwa6QTRyvUznX9hzPK4NW23XcsfBXQ2rTRrWOg9Zf5iyPOYJMEm17hM3g0FBFQgmedn0c80zym7oFe8bm1BzKAYB1HidUOGOEqVeQngMX0/UdBY61Ecms6YwrX+XKvgd7V+DpE3KES5cAT9rYoyiWGFep4qFQnIEcGUOZEuZWzK6pyRFM9k0TAvn85a/mvOMQ1yEnpL13VMcxg1bRGtpABHXRHvqNjlEPMXSHtATerburUfrv/Xy1tqw6pkV9EGNxisxMKq0</latexit>
1.27
7.94
1.82
7.90
8.88
idf 
weights
Importance Weighting
(Optional)
Maximum Similarity
x
<latexit sha1_base64="f2yzimwbR/Dgjzp6tZ
360fHRqNI=">AB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cW7Ae0oWy2k3btZhN2N2IJ/QV
ePCji1Z/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZLGLVCahGwSU
2DTcCO4lCGgUC28H4dua3H1FpHst7M0nQj+hQ8pAzaqzUeOqXK27VnYOsEi8nFchR75e/eoOYpRFKw
TVu5ifEzqgxnAqelXqoxoWxMh9i1VNItZ/ND52SM6sMSBgrW9KQufp7IqOR1pMosJ0RNSO97M3E/
7xuasJrP+MySQ1KtlgUpoKYmMy+JgOukBkxsYQyxe2thI2oszYbEo2BG/5VXSuqh6btVrXFZqN3kcR
TiBUzgHD6gBndQhyYwQHiGV3hzHpwX5935WLQWnHzmGP7A+fwB5jmM/A=</latexit>
<latexit sha1_base64="f2yzimwbR/Dgjzp6tZ
360fHRqNI=">AB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cW7Ae0oWy2k3btZhN2N2IJ/QV
ePCji1Z/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZLGLVCahGwSU
2DTcCO4lCGgUC28H4dua3H1FpHst7M0nQj+hQ8pAzaqzUeOqXK27VnYOsEi8nFchR75e/eoOYpRFKw
TVu5ifEzqgxnAqelXqoxoWxMh9i1VNItZ/ND52SM6sMSBgrW9KQufp7IqOR1pMosJ0RNSO97M3E/
7xuasJrP+MySQ1KtlgUpoKYmMy+JgOukBkxsYQyxe2thI2oszYbEo2BG/5VXSuqh6btVrXFZqN3kcR
TiBUzgHD6gBndQhyYwQHiGV3hzHpwX5935WLQWnHzmGP7A+fwB5jmM/A=</latexit>
<latexit sha1_base64="f2yzimwbR/Dgjzp6tZ
360fHRqNI=">AB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cW7Ae0oWy2k3btZhN2N2IJ/QV
ePCji1Z/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZLGLVCahGwSU
2DTcCO4lCGgUC28H4dua3H1FpHst7M0nQj+hQ8pAzaqzUeOqXK27VnYOsEi8nFchR75e/eoOYpRFKw
TVu5ifEzqgxnAqelXqoxoWxMh9i1VNItZ/ND52SM6sMSBgrW9KQufp7IqOR1pMosJ0RNSO97M3E/
7xuasJrP+MySQ1KtlgUpoKYmMy+JgOukBkxsYQyxe2thI2oszYbEo2BG/5VXSuqh6btVrXFZqN3kcR
TiBUzgHD6gBndQhyYwQHiGV3hzHpwX5935WLQWnHzmGP7A+fwB5jmM/A=</latexit>
<latexit sha1_base64="f2yzimwbR/Dgjzp6tZ
360fHRqNI=">AB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cW7Ae0oWy2k3btZhN2N2IJ/QV
ePCji1Z/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZLGLVCahGwSU
2DTcCO4lCGgUC28H4dua3H1FpHst7M0nQj+hQ8pAzaqzUeOqXK27VnYOsEi8nFchR75e/eoOYpRFKw
TVu5ifEzqgxnAqelXqoxoWxMh9i1VNItZ/ND52SM6sMSBgrW9KQufp7IqOR1pMosJ0RNSO97M3E/
7xuasJrP+MySQ1KtlgUpoKYmMy+JgOukBkxsYQyxe2thI2oszYbEo2BG/5VXSuqh6btVrXFZqN3kcR
TiBUzgHD6gBndQhyYwQHiGV3hzHpwX5935WLQWnHzmGP7A+fwB5jmM/A=</latexit>
ˆ
x
<latexit sha1_base64="5QTnVRVSrnyznVU7d5bF
5u03Iw=">AB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDbTbt0swm7E7GE/gvHhTx6u/
x5r9x0+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAo
k7wST29zvPHJtRKwecJpwP6IjJULBKFqp0x9TzJ5mg2rNrbtzkFXiFaQGBZqD6ld/GLM04gqZpMb0PDdBP6M
aBZN8VumnhieUTeiI9yxVNOLGz+bnzsiZVYkjLUthWSu/p7IaGTMNApsZ0RxbJa9XPzP6UYXvuZUEmKXLH
FojCVBGOS/06GQnOGcmoJZVrYWwkbU0Z2oQqNgRv+eV0r6oe27du7+sNW6KOMpwAqdwDh5cQPuoAktYDCB
Z3iFNydxXpx352PRWnKmWP4A+fzB7A8j8k=</latexit>
<latexit sha1_base64="5QTnVRVSrnyznVU7d5bF
5u03Iw=">AB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDbTbt0swm7E7GE/gvHhTx6u/
x5r9x0+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAo
k7wST29zvPHJtRKwecJpwP6IjJULBKFqp0x9TzJ5mg2rNrbtzkFXiFaQGBZqD6ld/GLM04gqZpMb0PDdBP6M
aBZN8VumnhieUTeiI9yxVNOLGz+bnzsiZVYkjLUthWSu/p7IaGTMNApsZ0RxbJa9XPzP6UYXvuZUEmKXLH
FojCVBGOS/06GQnOGcmoJZVrYWwkbU0Z2oQqNgRv+eV0r6oe27du7+sNW6KOMpwAqdwDh5cQPuoAktYDCB
Z3iFNydxXpx352PRWnKmWP4A+fzB7A8j8k=</latexit>
<latexit sha1_base64="5QTnVRVSrnyznVU7d5bF
5u03Iw=">AB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDbTbt0swm7E7GE/gvHhTx6u/
x5r9x0+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAo
k7wST29zvPHJtRKwecJpwP6IjJULBKFqp0x9TzJ5mg2rNrbtzkFXiFaQGBZqD6ld/GLM04gqZpMb0PDdBP6M
aBZN8VumnhieUTeiI9yxVNOLGz+bnzsiZVYkjLUthWSu/p7IaGTMNApsZ0RxbJa9XPzP6UYXvuZUEmKXLH
FojCVBGOS/06GQnOGcmoJZVrYWwkbU0Z2oQqNgRv+eV0r6oe27du7+sNW6KOMpwAqdwDh5cQPuoAktYDCB
Z3iFNydxXpx352PRWnKmWP4A+fzB7A8j8k=</latexit>
<latexit sha1_base64="5QTnVRVSrnyznVU7d5bF
5u03Iw=">AB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDbTbt0swm7E7GE/gvHhTx6u/
x5r9x0+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAo
k7wST29zvPHJtRKwecJpwP6IjJULBKFqp0x9TzJ5mg2rNrbtzkFXiFaQGBZqD6ld/GLM04gqZpMb0PDdBP6M
aBZN8VumnhieUTeiI9yxVNOLGz+bnzsiZVYkjLUthWSu/p7IaGTMNApsZ0RxbJa9XPzP6UYXvuZUEmKXLH
FojCVBGOS/06GQnOGcmoJZVrYWwkbU0Z2oQqNgRv+eV0r6oe27du7+sNW6KOMpwAqdwDh5cQPuoAktYDCB
Z3iFNydxXpx352PRWnKmWP4A+fzB7A8j8k=</latexit>
Reference
Figure 1: Illustration of the computation of the recall metric RBERT. Given the reference x and
candidate ˆ
x, we compute BERT embeddings and pairwise cosine similarity. We highlight the greedy
matching in red, and include the optional idf importance weighting.
We experiment with different models (Section 4), using the tokenizer provided with each model.
Given a tokenized reference sentence x = ⟨x1, . . . , xk⟩, the embedding model generates a se-
quence of vectors ⟨x1, . . . , xk⟩. Similarly, the tokenized candidate ˆ
x = ⟨ˆ
x1, . . . , ˆ
xm⟩is mapped
to ⟨ˆ
x1, . . . , ˆ
xl⟩. The main model we use is BERT, which tokenizes the input text into a sequence
of word pieces (Wu et al., 2016), where unknown words are split into several commonly observed
sequences of characters. The representation for each word piece is computed with a Transformer
encoder (Vaswani et al., 2017) by repeatedly applying self-attention and nonlinear transformations
in an alternating fashion. BERT embeddings have been shown to benefit various NLP tasks (Devlin
et al., 2019; Liu, 2019; Huang et al., 2019; Yang et al., 2019a).
Similarity Measure
The vector representation allows for a soft measure of similarity instead of
exact-string (Papineni et al., 2002) or heuristic (Banerjee & Lavie, 2005) matching. The cosine
similarity of a reference token xi and a candidate token ˆ
xj is
x⊤
i ˆ
xj
∥xi∥∥ˆ
xj∥. We use pre-normalized
vectors, which reduces this calculation to the inner product x⊤
i ˆ
xj. While this measure considers
tokens in isolation, the contextual embeddings contain information from the rest of the sentence.
BERTSCORE
The complete score matches each token in x to a token in ˆ
x to compute recall,
and each token in ˆ
x to a token in x to compute precision. We use greedy matching to maximize
the matching similarity score,2 where each token is matched to the most similar token in the other
sentence. We combine precision and recall to compute an F1 measure. For a reference x and
candidate ˆ
x, the recall, precision, and F1 scores are:
RBERT = 1
|x|
X
xi∈x
max
ˆ
xj∈ˆ
x x⊤
i ˆ
xj ,
PBERT = 1
|ˆ
x|
X
ˆ
xj∈ˆ
x
max
xi∈x x⊤
i ˆ
xj ,
FBERT = 2 PBERT · RBERT
PBERT + RBERT
.
Importance Weighting
Previous work on similarity measures demonstrated that rare words can
be more indicative for sentence similarity than common words (Banerjee & Lavie, 2005; Vedantam
et al., 2015). BERTSCORE enables us to easily incorporate importance weighting. We experiment
with inverse document frequency (idf) scores computed from the test corpus. Given M reference
sentences {x(i)}M
i=1, the idf score of a word-piece token w is
idf(w) = −log 1
M
M
X
i=1
I[w ∈x(i)] ,
where I[·] is an indicator function. We do not use the full tf-idf measure because we process single
sentences, where the term frequency (tf) is likely 1. For example, recall with idf weighting is
RBERT =
P
xi∈x idf(xi) maxˆ
xj∈ˆ
x x⊤
i ˆ
xj
P
xi∈x idf(xi)
.
Because we use reference sentences to compute idf, the idf scores remain the same for all systems
evaluated on a specific test set. We apply plus-one smoothing to handle unknown word pieces.
2We compare greedy matching with optimal assignment in Appendix C.
4


Published as a conference paper at ICLR 2020
Baseline Rescaling
Because we use pre-normalized vectors, our computed scores have the same
numerical range of cosine similarity (between −1 and 1). However, in practice we observe scores
in a more limited range, potentially because of the learned geometry of contextual embeddings.
While this characteristic does not impact BERTSCORE’s capability to rank text generation systems,
it makes the actual score less readable. We address this by rescaling BERTSCORE with respect
to its empirical lower bound b as a baseline. We compute b using Common Crawl monolingual
datasets.3 For each language and contextual embedding model, we create 1M candidate-reference
pairs by grouping two random sentences. Because of the random pairing and the corpus diversity,
each pair has very low lexical and semantic overlapping.4 We compute b by averaging BERTSCORE
computed on these sentence pairs. Equipped with baseline b, we rescale BERTSCORE linearly. For
example, the rescaled value ˆ
RBERT of RBERT is:
ˆ
RBERT = RBERT −b
1 −b
.
After this operation ˆ
RBERT is typically between 0 and 1. We apply the same rescaling procedure
for PBERT and FBERT. This method does not affect the ranking ability and human correlation of
BERTSCORE, and is intended solely to increase the score readability.
4
EXPERIMENTAL SETUP
We evaluate our approach on machine translation and image captioning.
Contextual Embedding Models
We evaluate twelve pre-trained contextual embedding models,
including variants of BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019b), XLNet (Yang et al.,
2019b), and XLM (Lample & Conneau, 2019). We present the best-performing models in Section 5.
We use the 24-layer RoBERTalarge model5 for English tasks, 12-layer BERTchinese model for Chi-
nese tasks, and the 12-layer cased multilingual BERTmulti model for other languages.6 We show the
performance of all other models in Appendix F. Contextual embedding models generate embedding
representations at every layer in the encoder network. Past work has shown that intermediate layers
produce more effective representations for semantic tasks (Liu et al., 2019a). We use the WMT16
dataset (Bojar et al., 2016) as a validation set to select the best layer of each model (Appendix B).
Machine Translation
Our main evaluation corpus is the WMT18 metric evaluation dataset (Ma
et al., 2018), which contains predictions of 149 translation systems across 14 language pairs, gold
references, and two types of human judgment scores. Segment-level human judgments assign a score
to each reference-candidate pair. System-level human judgments associate each system with a single
score based on all pairs in the test set. WMT18 includes translations from English to Czech, German,
Estonian, Finnish, Russian, and Turkish, and from the same set of languages to English. We follow
the WMT18 standard practice and use absolute Pearson correlation |ρ| and Kendall rank correlation
τ to evaluate metric quality, and compute significance with the Williams test (Williams, 1959) for |ρ|
and bootstrap re-sampling for τ as suggested by Graham & Baldwin (2014). We compute system-
level scores by averaging BERTSCORE for every reference-candidate pair. We also experiment with
hybrid systems by randomly sampling one candidate sentence from one of the available systems for
each reference sentence (Graham & Liu, 2016). This enables system-level experiments with a higher
number of systems. Human judgments of each hybrid system are created by averaging the WMT18
segment-level human judgments for the corresponding sentences in the sampled data. We compare
BERTSCOREs to one canonical metric for each category introduced in Section 2, and include the
comparison with all other participating metrics from WMT18 in Appendix F.
In addition to the standard evaluation, we design model selection experiments. We use 10K hybrid
systems super-sampled from WMT18. We randomly select 100 out of 10K hybrid systems, and rank
them using the automatic metrics. We repeat this process 100K times. We report the percentage of
the metric ranking agreeing with the human ranking on the best system (Hits@1). In Tables 23-28,
3https://commoncrawl.org/
4BLEU computed on these pairs is around zero.
5We use the tokenizer provided with each model. For all Hugging Face models that use the GPT-2 tokenizer,
at the time of our experiments, the tokenizer adds a space to the beginning of each sentence.
6All the models used are from https://github.com/huggingface/pytorch-transformers.
5


Published as a conference paper at ICLR 2020
Metric
en↔cs
en↔de
en↔et
en↔fi
en↔ru
en↔tr
en↔zh
(5/5)
(16/16)
(14/14)
(9/12)
(8/9)
(5/8)
(14/14)
BLEU
.970/.995
.971/.981
.986/.975
.973/.962
.979/.983
.657/.826
.978/.947
ITER
.975/.915
.990/.984
.975/.981
.996/.973
.937/.975
.861/.865
.980/ –
RUSE
.981/ –
.997/ –
.990/ –
.991/ –
.988/ –
.853/ –
.981/ –
YiSi-1
.950/.987
.992/.985
.979/.979
.973/.940
.991/.992
.958/.976
.951/.963
PBERT
.980/.994
.998/.988
.990/.981
.995/.957
.982/.990
.791/.935
.981/.954
RBERT
.998/.997
.997/.990
.986/.980
.997/.980
.995/.989
.054/.879
.990/.976
FBERT
.990/.997
.999/.989
.990/.982
.998/.972
.990/.990
.499/.908
.988/.967
FBERT (idf)
.985/.995
.999/.990
.992/.981
.992/.972
.991/.991
.826/.941
.989/.973
Table 1: Absolute Pearson correlations with system-level human judgments on WMT18. For each
language pair, the left number is the to-English correlation, and the right is the from-English. We
bold correlations of metrics not significantly outperformed by any other metric under Williams Test
for that language pair and direction. The numbers in parenthesis are the number of systems used for
each language pair and direction.
Metric
en↔cs
en↔de
en↔et
en↔fi
en↔ru
en↔tr
en↔zh
BLEU
.956/.993
.969/.977
.981/.971
.962/.958
.972/.977
.586/.796
.968/.941
ITER
.966/.865
.990/.978
.975/.982
.989/.966
.943/.965
.742/.872
.978/ –
RUSE
.974/ –
.996/ –
.988/ –
.983/ –
.982/ –
.780/ –
.973/ –
YiSi-1
.942/.985
.991/.983
.976/.976
.964/.938
.985/.989
.881/.942
.943/.957
PBERT
.965/.989
.995/.983
.990/.970
.976/.951
.976/.988
.846/.936
.975/.950
RBERT
.989/.995
.997/.991
.982/.979
.989/.977
.988/.989
.540/.872
.981/.980
FBERT
.978/.993
.998/.988
.989/.978
.983/.969
.985/.989
.760/.910
.981/.969
FBERT (idf)
.982/.995
.998/.988
.988/.979
.989/.969
.983/.987
.453/.877
.980/.963
Table 2: Absolute Pearson correlations with system-level human judgments on WMT18. We use
10K hybrid super-sampled systems for each language pair and direction. For each language pair, the
left number is the to-English correlation, and the right is the from-English. Bolding criteria is the
same as in Table 1.
we include two additional measures to the model selection study: (a) the mean reciprocal rank of the
top metric-rated system according to the human ranking, and (b) the difference between the human
score of the top human-rated system and that of the top metric-rated system.
Additionally, we report the same study on the WMT17 (Bojar et al., 2017) and the WMT16 (Bojar
et al., 2016) datasests in Appendix F.7 This adds 202 systems to our evaluation.
Image Captioning
We use the human judgments of twelve submission entries from the COCO
2015 Captioning Challenge. Each participating system generates a caption for each image in the
COCO validation set (Lin et al., 2014), and each image has approximately five reference cap-
tions.
Following Cui et al. (2018), we compute the Pearson correlation with two system-level
metrics: the percentage of captions that are evaluated as better or equal to human captions (M1)
and the percentage of captions that are indistinguishable from human captions (M2). We compute
BERTSCORE with multiple references by scoring the candidate with each available reference and
returning the highest score. We compare with eight task-agnostic metrics: BLEU (Papineni et al.,
2002), METEOR (Banerjee & Lavie, 2005), ROUGE-L (Lin, 2004), CIDER (Vedantam et al., 2015),
BEER (Stanojevi´
c & Sima’an, 2014), EED (Stanchev et al., 2019), CHRF++ (Popovi´
c, 2017), and
CHARACTER (Wang et al., 2016). We also compare with two task-specific metrics: SPICE (Ander-
son et al., 2016) and LEIC (Cui et al., 2018). SPICE is computed using the similarity of scene graphs
parsed from the reference and candidate captions. LEIC is trained to predict if a caption is written
by a human given the image.
7For WMT16, we only conduct segment-level experiments on to-English pairs due to errors in the dataset.
6


Published as a conference paper at ICLR 2020
Metric
en↔cs
en↔de
en↔et
en↔fi
en↔ru
en↔tr
en↔zh
BLEU
.134/.151
.803/.610
.756/.618
.461/.088
.228/.519
.095/.029
.658/.515
ITER
.154/.000
.814/.692
.742/.733
.475/.111
.234/.532
.102/.030
.673/ –
RUSE
.214/ –
.823/ –
.785/ –
.487/ –
.248/ –
.109/ –
.670/ –
YiSi-1
.159/.178
.809/.671
.749/.671
.467/.230
.248/.544
.108/.398
.613/.594
PBERT
.173/.180
.706/.663
.764/.771
.498/.078
.255/.545
.140/.372
.661/.551
RBERT
.163/.184
.804/.730
.770/.722
.494/.148
.260/.542
.005/.030
.677/.657
FBERT
.175/.184
.824/.703
.769/.763
.501/.082
.262/.544
.142/.031
.673/.629
FBERT (idf)
.179/.178
.824/.722
.760/.764
.503/.082
.265/.539
.004/.030
.678/.595
Table 3: Model selection accuracies (Hits@1) on WMT18 hybrid systems. We report the average of
100K samples and the 0.95 confidence intervals are below 10−3. We bold the highest numbers for
each language pair and direction.
Metric
en↔cs
en↔de
en↔et
en↔fi
en↔ru
en↔tr
en↔zh
(5k/5k)
(78k/ 20k)
(57k/32k)
(16k/10k)
(10k/22k)
(9k/1k)
(33k/29k)
BLEU
.233/.389
.415/.620
.285/.414
.154/.355
.228/.330
.145/.261
.178/.311
ITER
.198/.333
.396/.610
.235/.392
.128/.311
.139/.291
-.029/.236
.144/ –
RUSE
.347/ –
.498/ –
.368/ –
.273/ –
.311/ –
.259/ –
.218/ –
YiSi-1
.319/.496
.488/.691
.351/.546
.231/.504
.300/.407
.234/.418
.211/.323
PBERT
.387/.541
.541/.715
.389/.549
.283/.486
.345/.414
.280/.328
.248/.337
RBERT
.388/.570
.546/.728
.391/.594
.304/.565
.343/.420
.290/.411
.255/.367
FBERT
.404/.562
.550/.728
.397/.586
.296/.546
.353/.423
.292/.399
.264/.364
FBERT (idf)
.408/.553
.550/.721
.395/585
.293/.537
.346/.425
.296/.406
.260/.366
Table 4: Kendall correlations with segment-level human judgments on WMT18. For each language
pair, the left number is the to-English correlation, and the right is the from-English. We bold corre-
lations of metrics not significantly outperformed by any other metric under bootstrap sampling for
that language pair and direction. The numbers in parenthesis are the number of candidate-reference
sentence pairs for each language pair and direction.
5
RESULTS
Machine Translation
Tables 1–3 show system-level correlation to human judgements, correla-
tions on hybrid systems, and model selection performance. We observe that BERTSCORE is con-
sistently a top performer. In to-English results, RUSE (Shimanaka et al., 2018) shows competitive
performance. However, RUSE is a supervised method trained on WMT16 and WMT15 human
judgment data. In cases where RUSE models were not made available, such as for our from-English
experiments, it is not possible to use RUSE without additional data and training. Table 4 shows
segment-level correlations. We see that BERTSCORE exhibits significantly higher performance
compared to the other metrics. The large improvement over BLEU stands out, making BERTSCORE
particularly suitable to analyze specific examples, where SENTBLEU is less reliable. In Appendix A,
we provide qualitative examples to illustrate the segment-level performance difference between
SENTBLEU and BERTSCORE. At the segment-level, BERTSCORE even significantly outperforms
RUSE. Overall, we find that applying importance weighting using idf at times provides small bene-
fit, but in other cases does not help. Understanding better when such importance weighting is likely
to help is an important direction for future work, and likely depends on the domain of the text and
the available test data. We continue without idf weighting for the rest of our experiments. While
recall RBERT, precision PBERT, and F1 FBERT alternate as the best measure in different setting, F1
FBERT performs reliably well across all the different settings. Our overall recommendation is there-
fore to use F1. We present additional results using the full set of 351 systems and evaluation metrics
in Tables 12–28 in the appendix, including for experiments with idf importance weighting, different
contextual embedding models, and model selection.
Image Captioning
Table 5 shows correlation results for the COCO Captioning Challenge.
BERTSCORE outperforms all task-agnostic baselines by large margins. Image captioning presents a
challenging evaluation scenario, and metrics based on strict n-gram matching, including BLEU and
ROUGE, show weak correlations with human judgments. idf importance weighting shows signifi-
7


Published as a conference paper at ICLR 2020
Metric
M1
M2
BLEU
-0.019∗
-0.005∗
METEOR
0.606∗
0.594∗
ROUGE-L
0.090∗
0.096∗
CIDER
0.438∗
0.440∗
SPICE
0.759∗
0.750∗
LEIC
0.939∗
0.949∗
BEER
0.491
0.562
EED
0.545
0.599
CHRF++
0.702
0.729
CHARACTER
0.800
0.801
PBERT
-0.105
-0.041
RBERT
0.888
0.863
FBERT
0.322
0.350
RBERT (idf)
0.917
0.889
Table 5:
Pearson correlation on the
2015 COCO Captioning Challenge.
The M1 and M2 measures are described
in Section 4. LEIC uses images as addi-
tional inputs. Numbers with ∗are cited
from Cui et al. (2018).
We bold the
highest correlations of task-specific and
task-agnostic metrics.
Type
Method
QQP
PAWSQQP
Trained on QQP
(supervised)
DecAtt
0.939∗
0.263
DIIN
0.952∗
0.324
BERT
0.963∗
0.351
Trained on QQP
+ PAWSQQP
(supervised)
DecAtt
-
0.511
DIIN
-
0.778
BERT
-
0.831
Metric
(Not trained
on QQP or
PAWSQQP)
BLEU
0.707
0.527
METEOR
0.755
0.532
ROUGE-L
0.740
0.536
CHRF++
0.577
0.608
BEER
0.741
0.564
EED
0.743
0.611
CHARACTER
0.698
0.650
PBERT
0.757
0.687
RBERT
0.744
0.685
FBERT
0.761
0.685
FBERT (idf)
0.777
0.693
Table 6:
Area under ROC curve (AUC) on QQP
and PAWSQQP datasets.
The scores of trained De-
cATT (Parikh et al., 2016), DIIN (Gong et al., 2018),
and fine-tuned BERT are reported by Zhang et al.
(2019). Numbers with ∗are scores on the held-out test
set of QQP. We bold the highest correlations of task-
specific and task-agnostic metrics.
cant benefit for this task, suggesting people attribute higher importance to content words. Finally,
LEIC (Cui et al., 2018), a trained metric that takes images as additional inputs and is optimized
specifically for the COCO data and this set of systems, outperforms all other methods.
Speed
Despite the use of a large pre-trained model, computing BERTSCORE is relatively fast. We
are able to process 192.5 candidate-reference pairs/second using a GTX-1080Ti GPU. The complete
WMT18 en-de test set, which includes 2,998 sentences, takes 15.6sec to process, compared to 5.4sec
with SacreBLEU (Post, 2018), a common BLEU implementation. Given the sizes of commonly used
test and validation sets, the increase in processing time is relatively marginal, and BERTSCORE is
a good fit for using during validation (e.g., for stopping) and testing, especially when compared to
the time costs of other development stages.
6
ROBUSTNESS ANALYSIS
We test the robustness of BERTSCORE using adversarial paraphrase classification. We use the
Quora Question Pair corpus (QQP; Iyer et al., 2017) and the adversarial paraphrases from the Para-
phrase Adversaries from Word Scrambling dataset (PAWS; Zhang et al., 2019). Both datasets con-
tain pairs of sentences labeled to indicate whether they are paraphrases or not. Positive examples
in QQP are real duplicate questions, while negative examples are related, but different questions.
Sentence pairs in PAWS are generated through word swapping. For example, in PAWS, Flights from
New York to Florida may be changed to Flights from Florida to New York and a good classifier
should identify that these two sentences are not paraphrases. PAWS includes two parts: PAWSQQP,
which is based on the QQP data, and PAWSWiki. We use the PAWSQQP development set which
contains 667 sentences. For the automatic metrics, we use no paraphrase detection training data.
We expect that pairs with higher scores are more likely to be paraphrases. To evaluate the automatic
metrics on QQA, we use the first 5,000 sentences in the training set instead of the the test set because
the test labels are not available. We treat the first sentence as the reference and the second sentence
as the candidate.
Table 6 reports the area under ROC curve (AUC) for existing models and automatic metrics. We
observe that supervised classifiers trained on QQP perform worse than random guess on PAWSQQP,
which shows these models predict the adversarial examples are more likely to be paraphrases. When
8


Published as a conference paper at ICLR 2020
adversarial examples are provided in training, state-of-the-art models like DIIN (Gong et al., 2018)
and fine-tuned BERT are able to identify the adversarial examples but their performance still de-
creases significantly from their performance on QQP. Most metrics have decent performance on
QQP, but show a significant performance drop on PAWSQQP, almost down to chance performance.
This suggests these metrics fail to to distinguish the harder adversarial examples. In contrast, the
performance of BERTSCORE drops only slightly, showing more robustness than the other metrics.
7
DISCUSSION
We propose BERTSCORE, a new metric for evaluating generated text against gold standard refer-
ences. BERTSCORE is purposely designed to be simple, task agnostic, and easy to use. Our analysis
illustrates how BERTSCORE resolves some of the limitations of commonly used metrics, especially
on challenging adversarial examples. We conduct extensive experiments with various configuration
choices for BERTSCORE, including the contextual embedding model used and the use of impor-
tance weighting. Overall, our extensive experiments, including the ones in the appendix, show that
BERTSCORE achieves better correlation than common metrics, and is effective for model selec-
tion. However, there is no one configuration of BERTSCORE that clearly outperforms all others.
While the differences between the top configurations are often small, it is important for the user to
be aware of the different trade-offs, and consider the domain and languages when selecting the exact
configuration to use. In general, for machine translation evaluation, we suggest using FBERT, which
we find the most reliable. For evaluating text generation in English, we recommend using the 24-
layer RoBERTalarge model to compute BERTSCORE. For non-English language, the multilingual
BERTmulti is a suitable choice although BERTSCORE computed with this model has less stable
performance on low-resource languages. We report the optimal hyperparameter for all models we
experimented with in Appendix B
Briefly following our initial preprint publication, Zhao et al. (2019) published a concurrently devel-
oped method related to ours, but with a focus on integrating contextual word embeddings with earth
mover’s distance (EMD; Rubner et al., 1998) rather than our simple matching process. They also
propose various improvements compared to our use of contextualized embeddings. We study these
improvements in Appendix C and show that integrating them into BERTSCORE makes it equivalent
or better than the EMD-based approach. Largely though, the effect of the different improvements
on BERTSCORE is more modest compared to their method. Shortly after our initial publication,
YiSi-1 was updated to use BERT embeddings, showing improved performance (Lo, 2019). This
further corroborates our findings. Other recent related work includes training a model on top of
BERT to maximize the correlation with human judgments (Mathur et al., 2019) and evaluating gen-
eration with a BERT model fine-tuned on paraphrasing (Yoshimura et al., 2019). More recent work
shows the potential of using BERTSCORE for training a summarization system (Li et al., 2019)
and for domain-specific evaluation using SciBERT (Beltagy et al., 2019) to evaluate abstractive text
summarization (Gabriel et al., 2019).
In future work, we look forward to designing new task-specific metrics that use BERTSCORE as a
subroutine and accommodate task-specific needs, similar to how Wieting et al. (2019) suggests to use
semantic similarity for machine translation training. Because BERTSCORE is fully differentiable,
it also can be incorporated into a training procedure to compute a learning loss that reduces the
mismatch between optimization and evaluation objectives.
ACKNOWLEDGEMENT
This research is supported in part by grants from the National Science Foundation (III-1618134, III-
1526012, IIS1149882, IIS-1724282, TRIPODS-1740822, CAREER-1750499), the Office of Naval
Research DOD (N00014-17-1-2175), and the Bill and Melinda Gates Foundation, SAP, Zillow,
Workday, and Facebook Research. We thank Graham Neubig and David Grangier for for their
insightful comments. We thank the Cornell NLP community including but not limited to Claire
Cardie, Tianze Shi, Alexandra Schofield, Gregory Yauney, and Rishi Bommasani. We thank Yin
Cui and Guandao Yang for their help with the COCO 2015 dataset.
9


Published as a conference paper at ICLR 2020
REFERENCES
Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. SPICE: Semantic proposi-
tional image caption evaluation. In ECCV, 2016.
Ben Athiwaratkun, Andrew Wilson, and Anima Anandkumar. Probabilistic fasttext for multi-sense
word embeddings. In ACL, 2018.
Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for mt evaluation with im-
proved correlation with human judgments. In IEEvaluation@ACL, 2005.
Iz Beltagy, Kyle Lo, and Arman Cohan. SciBERT: A pretrained language model for scientific text.
ArXiv, 2019.
Ondˇ
rej Bojar, Yvette Graham, Amir Kamran, and Miloš Stanojevi´
c. Results of the WMT16 metrics
shared task. In WMT, 2016.
Ondˇ
rej Bojar, Yvette Graham, and Amir Kamran. Results of the WMT17 metrics shared task. In
WMT, 2017.
Arun Chaganty, Stephen Mussmann, and Percy Liang. The price of debiasing automatic metrics in
natural language evalaution. In ACL, 2018.
Julian Chow, Lucia Specia, and Pranava Madhyastha. WMDO: Fluency-based word mover’s dis-
tance for machine translation evaluation. In WMT, 2019.
Elizabeth Clark, Asli Celikyilmaz, and Noah A. Smith. Sentence mover’s similarity: Automatic
evaluation for multi-sentence texts. In ACL, 2019.
Courtney Corley and Rada Mihalcea. Measuring the semantic similarity of texts. In ACL Workshop,
EMSEE ’05, 2005.
Yin Cui, Guandao Yang, Andreas Veit, Xun Huang, and Serge J. Belongie. Learning to evaluate
image captioning. In CVPR, 2018.
Michael Denkowski and Alon Lavie. Meteor universal: Language specific translation evaluation for
any target language. In WMT@ACL, 2014.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep
bidirectional transformers for language understanding. In NAACL-HLT, 2019.
George Doddington.
Automatic evaluation of machine translation quality using n-gram co-
occurrence statistics. In HLT, 2002.
William B Dolan and Chris Brockett. Automatically constructing a corpus of sentential paraphrases.
In IWP, 2005.
Saadia Gabriel, Antoine Bosselut, Ari Holtzman, Kyle Lo, Asli Çelikyilmaz, and Yejin Choi. Co-
operative generator-discriminator networks for abstractive summarization with narrative flow.
ArXiv, 2019.
Michel Galley, Chris Brockett, Alessandro Sordoni, Yangfeng Ji, Michael Auli, Chris Quirk, Mar-
garet Mitchell, Jianfeng Gao, and William B. Dolan. deltaBLEU: A discriminative metric for
generation tasks with intrinsically diverse targets. In ACL, 2015.
Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional
sequence to sequence learning. In ICML, 2017.
Yichen Gong, Heng Luo, and Jian Zhang. Natural language inference over interaction space. In
ICLR, 2018.
Yvette Graham and Timothy Baldwin. Testing for significance of increased correlation with human
judgment. In EMNLP, 2014.
10


Published as a conference paper at ICLR 2020
Yvette Graham and Qun Liu. Achieving accurate conclusions in evaluation of automatic machine
translation metrics. In NAACL, 2016.
Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Armand Joulin, and Tomas Mikolov. Learning
word vectors for 157 languages. arXiv preprint arXiv:1802.06893, 2018.
Yinuo Guo and Junfeng Hu. Meteor++ 2.0: Adopt syntactic level paraphrase knowledge into ma-
chine translation evaluation. In WMT, 2019.
Tatsu Hashimoto, Hugh Zhang, and Percy Liang. Unifying human and statistical evaluation for
natural language generation. In NAACL-HLT, 2019.
Chenyang Huang, Amine Trabelsi, and Osmar R Zaïane. ANA at semeval-2019 task 3: Contex-
tual emotion detection in conversations through hierarchical LSTMs and BERT. arXiv preprint
arXiv:1904.00132, 2019.
Hideki Isozaki, Tsutomu Hirao, Kevin Duh, Katsuhito Sudoh, and Hajime Tsukada. Automatic
evaluation of translation quality for distant language pairs. In EMNLP, 2010.
Shankar Iyer, Nikhil Dandekar, and Kornel Csernai. First quora dataset release: Question pairs.
https://tinyurl.com/y2y8u5ed, 2017.
Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola
Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar,
Alexandra Constantin, and Evan Herbst.
Moses: Open source toolkit for statistical machine
translation. In ACL, 2007.
Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Weinberger. From word embeddings to document
distances. In ICML, 2015.
Guillaume Lample and Alexis Conneau. Cross-lingual language model pretraining. arXiv, 2019.
Gregor Leusch, Nicola Ueffing, and Hermann Ney. CDER: Efficient MT evaluation using block
movements. In EACL, 2006.
Vladimir Iosifovich Levenshtein. Binary Codes Capable of Correcting Deletions, Insertions and
Rever sals. Soviet Physics Doklady, 10, 1966.
Siyao Li, Deren Lei, Pengda Qin, and William Yang Wang. Deep reinforcement learning with
distributional semantic rewards for abstractive summarization. In EMNLP-IJCNLP, 2019.
Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In ACL, 2004.
Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James
Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO:
Common objects in context. In ECCV, 2014.
Nelson F. Liu, Matt Gardner, Yonatan Belinkov, Matthew E. Peters, and Noah A. Smith. Linguistic
knowledge and transferability of contextual representations. arXiv preprint arXiv:1903.08855,
2019a.
Yang Liu. Fine-tune BERT for extractive summarization. arXiv preprint arXiv:1903.10318, 2019.
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike
Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretrain-
ing approach. arXiv, abs/1907.11692, 2019b.
Chi-kiu Lo. MEANT 2.0: Accurate semantic mt evaluation for any output language. In WMT, 2017.
Chi-kiu Lo. YiSi - a unified semantic MT quality evaluation and estimation metric for languages
with different levels of available resources. In WMT, 2019.
11


Published as a conference paper at ICLR 2020
Chi-kiu Lo, Michel Simard, Darlene Stewart, Samuel Larkin, Cyril Goutte, and Patrick Littell. Ac-
curate semantic textual similarity for cleaning noisy parallel corpora using semantic machine
translation evaluation metric: The NRC supervised submissions to the parallel corpus filtering
task. In WMT, 2018.
Ryan Lowe, Michael Noseworthy, Iulian Vlad Serban, Nicolas Angelard-Gontier, Yoshua Bengio,
and Joelle Pineau. Towards an automatic Turing test: Learning to evaluate dialogue responses.
In ACL, 2017.
Qingsong Ma, Yvette Graham, Shugen Wang, and Qun Liu. Blend: a novel combined mt metric
based on direct assessment – casict-dcu submission to WMT17 metrics task. In WMT, 2017.
Qingsong Ma, Ondrej Bojar, and Yvette Graham. Results of the WMT18 metrics shared task: Both
characters and embeddings achieve good performance. In WMT, 2018.
Nitika Mathur, Timothy Baldwin, and Trevor Cohn. Putting evaluation in context: Contextual em-
beddings improve machine translation evaluation. In ACL, 2019.
Tomas Mikolov, Ilya Sutskever, Kai Chen, Gregory S. Corrado, and Jeffrey Dean.
Distributed
representations of words and phrases and their compositionality. In NIPS, 2013.
Dai Quoc Nguyen, Dat Quoc Nguyen, Ashutosh Modi, Stefan Thater, and Manfred Pinkal.
A
mixture model for learning multi-sense word embeddings. In ACL, 2017.
Myle Ott, Sergey Edunov, David Grangier, and Michael Auli. Scaling neural machine translation.
In WMT, 2018.
Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier,
and Michael Auli. fairseq: A fast, extensible toolkit for sequence modeling. arXiv preprint
arXiv:1904.01038, 2019.
Joybrata Panja and Sudip Kumar Naskar. Iter: Improving translation edit rate through optimizable
edit costs. In WMT, 2018.
Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic
evaluation of machine translation. In ACL, 2002.
Ankur Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention
model for natural language inference. In EMNLP, 2016.
Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word
representation. In EMNLP, 2014.
Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and
Luke S. Zettlemoyer. Deep contextualized word representations. In NAACL-HLT, 2018.
Maja Popovi´
c. chrf: character n-gram f-score for automatic mt evaluation. In WMT@ACL, 2015.
Maja Popovi´
c. chrf++: words helping character n-grams. In WMT, 2017.
Matt Post. A call for clarity in reporting BLEU scores. In WMT, 2018.
Nils Reimers and Iryna Gurevych. Alternative weighting schemes for elmo embeddings. arXiv
preprint arXiv:1904.02954, 2019.
Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. A metric for distributions with applications to
image databases. In ICCV. IEEE, 1998.
Vasile Rus and Mihai Lintean. A comparison of greedy and optimal assessment of natural language
student input using word-to-word similarity metrics. In Proceedings of the Seventh Workshop on
Building Educational Applications Using NLP. ACL, 2012.
12


Published as a conference paper at ICLR 2020
Andreas Rücklé, Steffen Eger, Maxime Peyrard, and Iryna Gurevych. Concatenated power mean
word embeddings as universal cross-lingual sentence representations. arXiv, 2018.
Shikhar Sharma, Layla El Asri, Hannes Schulz, and Jeremie Zumer. Relevance of unsupervised
metrics in task-oriented dialogue for evaluating natural language generation.
arXiv preprint
arXiv:1706.09799, 2018.
Hiroki Shimanaka, Tomoyuki Kajiwara, and Mamoru Komachi. Ruse: Regressor using sentence
embeddings for automatic machine translation evaluation. In WMT, 2018.
Matthew Snover, Bonnie Dorr, Richard Schwartz, Linnea Micciulla, and John Makhoul. A study of
translation edit rate with targeted human annotation. In AMTA, 2006.
Peter Stanchev, Weiyue Wang, and Hermann Ney. EED: Extended edit distance measure for machine
translation. In WMT, 2019.
Miloš Stanojevi´
c and Khalil Sima’an. Beer: Better evaluation as ranking. In WMT, 2014.
Christoph Tillmann, Stephan Vogel, Hermann Ney, Arkaitz Zubiaga, and Hassan Sawaf. Accelerated
dp based search for statistical translation. In EUROSPEECH, 1997.
Kristina Toutanova, Chris Brockett, Ke M Tran, and Saleema Amershi. A dataset and evaluation
metrics for abstractive compression of sentences and short paragraphs. In EMNLP, 2016.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,
Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017.
Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image
description evaluation. In CVPR, 2015.
Weiyue Wang, Jan-Thorsten Peter, Hendrik Rosendahl, and Hermann Ney. Character: Translation
edit rate on character level. In WMT, 2016.
John Wieting, Taylor Berg-Kirkpatrick, Kevin Gimpel, and Graham Neubig. Beyond BLEU:training
neural machine translation with semantic similarity. In ACL, 2019.
Adina Williams, Nikita Nangia, and Samuel Bowman. A broad-coverage challenge corpus for sen-
tence understanding through inference. In ACL, 2018.
Evan James Williams. Regression analysis. wiley, 1959.
Felix Wu, Angela Fan, Alexei Baevski, Yann Dauphin, and Michael Auli. Pay less attention with
lightweight and dynamic convolutions. In ICLR, 2019.
Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey,
Maxim Krikun, Yuan Cao, Qin Gao, Jeff Klingner, Apurva Shah, Melvin Johnson, Xiaobing Liu,
Lukasz Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto Kazawa, Keith Stevens,
George Kurian, Nishant Patil, Wei Wang, Cliff Young, Jason Smith, Jason Riesa, Alex Rudnick,
Oriol Vinyals, Gregory S. Corrado, Macduff Hughes, and Jeffrey Dean. Google’s neural machine
translation system: Bridging the gap between human and machine translation. arXiv preprint
arXiv:1609.08144, 2016.
Wei Yang, Haotian Zhang, and Jimmy Lin. Simple applications of BERT for ad hoc document
retrieval. arXiv preprint arXiv:1903.10972, 2019a.
Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V. Le.
XLNet: Generalized autoregressive pretraining for language understanding. arXiv, 2019b.
Ryoma Yoshimura, Hiroki Shimanaka, Yukio Matsumura, Hayahide Yamagishi, and Mamoru Ko-
machi. Filtering pseudo-references by paraphrasing for automatic evaluation of machine transla-
tion. In WMT, 2019.
13


Published as a conference paper at ICLR 2020
Yuan Zhang, Jason Baldridge, and Luheng He. PAWS: Paraphrase adversaries from word scram-
bling. arXiv preprint arXiv:1904.01130, 2019.
Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Christian M. Meyer, and Steffen Eger. Moverscore:
Text generation evaluating with contextualized embeddings and earth mover distance. In EMNLP,
2019.
14


Published as a conference paper at ICLR 2020
Case
No.
Reference and Candidate Pairs
Human
FBERT
BLEU
FBERT > BLEU
1.
x: At the same time Kingfisher is closing 60 B&Q outlets across the country
38
125
530
ˆ
x: At the same time, Kingfisher will close 60 B & Q stores nationwide
2.
x: Hewlett-Packard to cut up to 30,000 jobs
119
39
441
ˆ
x: Hewlett-Packard will reduce jobs up to 30.000
3.
x: According to opinion in Hungary, Serbia is “a safe third country".
23
96
465
ˆ
x: According to Hungarian view, Serbia is a “safe third country."
4.
x: Experts believe November’s Black Friday could be holding back spending.
73
147
492
ˆ
x: Experts believe that the Black Friday in November has put the brakes on spending
5.
x: And it’s from this perspective that I will watch him die.
37
111
414
ˆ
x: And from this perspective, I will see him die.
BLEU > FBERT
6.
x: In their view the human dignity of the man had been violated.
500
470
115
ˆ
x: Look at the human dignity of the man injured.
8.
x: For example when he steered a shot from Ideye over the crossbar in the 56th minute.
516
524
185
ˆ
x: So, for example, when he steered a shot of Ideye over the latte (56th).
7.
x: A good prank is funny, but takes moments to reverse.
495
424
152
ˆ
x: A good prank is funny, but it takes only moments before he becomes a boomerang.
9.
x: I will put the pressure on them and onus on them to make a decision.
507
471
220
ˆ
x: I will exert the pressure on it and her urge to make a decision.
10.
x: Transport for London is not amused by this flyposting "vandalism."
527
527
246
ˆ
x: Transport for London is the Plaka animal "vandalism" is not funny.
FBERT > Human
11.
x: One big obstacle to access to the jobs market is the lack of knowledge of the German language.
558
131
313
ˆ
x: A major hurdle for access to the labour market are a lack of knowledge of English.
12.
x: On Monday night Hungary closed its 175 km long border with Serbia.
413
135
55
ˆ
x: Hungary had in the night of Tuesday closed its 175 km long border with Serbia.
13.
x: They got nothing, but they were allowed to keep the clothes.
428
174
318
ˆ
x: You got nothing, but could keep the clothes.
14.
x: A majority of Republicans don’t see Trump’s temperament as a problem.
290
34
134
ˆ
x: A majority of Republicans see Trump’s temperament is not a problem.
15.
x:His car was still running in the driveway.
299
49
71
ˆ
x: His car was still in the driveway.
Human > FBERT
16.
x: Currently the majority of staff are men.
77
525
553
ˆ
x: At the moment the men predominate among the staff.
17.
x: There are, indeed, multiple variables at play.
30
446
552
ˆ
x: In fact, several variables play a role.
18.
x: One was a man of about 5ft 11in tall.
124
551
528
ˆ
x: One of the men was about 1,80 metres in size.
19.
x: All that stuff sure does take a toll.
90
454
547
ˆ
x: All of this certainly exacts its toll.
20.
x: Wage gains have shown signs of picking up.
140
464
514
ˆ
x: Increases of wages showed signs of a recovery.
Table 7: Examples sentences where similarity ranks assigned by Human, FBERT, and BLEU differ
significantly on WMT16 German-to-English evaluation task. x: gold reference, ˆ
x: candidate outputs
of MT systems. Rankings assigned by Human, FBERT, and BLEU are shown in the right three
columns. The sentences are ranked by the similarity, i.e. rank 1 is the most similar pair assigned by
a score. An ideal metric should rank similar to humans.
A
QUALITATIVE ANALYSIS
We study BERTSCORE and SENTBLEU using WMT16 German-to-English (Bojar et al., 2016). We
rank all 560 candidate-reference pairs by human score, BERTSCORE, or SENTBLEU from most
similar to least similar. Ideally, the ranking assigned by BERTSCORE and SENTBLEU should be
similar to the ranking assigned by the human score.
Table 7 first shows examples where BERTSCORE and SENTBLEU scores disagree about the ranking
for a candidate-reference pair by a large number. We observe that BERTSCORE is effectively able
to capture synonyms and changes in word order. For example, the reference and candidate sentences
in pair 3 are almost identical except that the candidate replaces opinion in Hungary with Hungarian
view and switches the order of the quotation mark (“) and a. While BERTSCORE ranks the pair
relatively high, SENTBLEU judges the pair as dissimilar, because it cannot match synonyms and is
sensitive to the small word order changes. Pair 5 shows a set of changes that preserve the semantic
meaning: replacing to cut with will reduce and swapping the order of 30,000 and jobs. BERTSCORE
ranks the candidate translation similar to the human judgment, whereas SENTBLEU ranks it much
lower. We also see that SENTBLEU potentially over-rewards n-gram overlap, even when phrases
are used very differently. In pair 6, both the candidate and the reference contain the human dignity
of the man. Yet the two sentences convey very different meaning. BERTSCORE agrees with the
human judgment and ranks the pair low. In contrast, SENTBLEU considers the pair as relatively
similar because of the significant word overlap.
15


Published as a conference paper at ICLR 2020
Figure 2: BERTSCORE visualization. The cosine similarity of each word matching in PBERT are
color-coded.
The bottom half of Table 7 shows examples where BERTSCORE and human judgments disagree
about the ranking. We observe that BERTSCORE finds it difficult to detect factual errors. For
example, BERTSCORE assigns high similarity to pair 11 when the translation replaces German
language with English and pair 12 where the translation incorrectly outputs Tuesday when it is
supposed to generate Monday. BERTSCORE also fails to identify that 5ft 11in is equivalent with
1.80 metres in pair 18. As a result, BERTSCORE assigns low similarity to the eighth pair in Table 7.
SENTBLEU also suffers from these limitations.
Figure 2 visualizes the BERTSCORE matching of two pairs of candidate and reference sentences.
The figure illustrates how FBERT matches synonymous phrases, such as imported cars and foreign
cars. We also see that FBERT effectively matches words even given a high ordering distortion, for
example the token people in the figure.
16


Published as a conference paper at ICLR 2020
B
REPRESENTATION CHOICE
As suggested by previous works (Peters et al., 2018; Reimers & Gurevych, 2019), selecting a
good layer or a good combination of layers from the BERT model is important. In designing
BERTSCORE, we use WMT16 segment-level human judgment data as a development set to fa-
cilitate our representation choice. For Chinese models, we tune with the WMT17 “en-zh” data
because the language pair “en-zh” is not available in WMT16. In Figure 3, we plot the change of
human correlation of FBERT over different layers of BERT, RoBERTa, XLNet and XLM models.
Based on results from different models, we identify a common trend that FBERT computed with the
intermediate representations tends to work better. We tune the number of layer to use for a range of
publicly available models.8 Table 8 shows the results of our hyperparameter search.
Model
Total Number of Layers
Best Layer
bert-base-uncased
12
9
bert-large-uncased
24
18
bert-base-cased-finetuned-mrpc
12
9
bert-base-multilingual-cased
12
9
bert-base-chinese
12
8
roberta-base
12
10
roberta-large
24
17
roberta-large-mnli
24
19
xlnet-base-cased
12
5
xlnet-large-cased
24
7
xlm-mlm-en-2048
12
7
xlm-mlm-100-1280
16
11
Table 8: Recommended layer of representation to use for BERTSCORE. The layers are chosen
based on a held-out validation set (WMT16).
8https://huggingface.co/pytorch-transformers/pretrained_models.html
17


Published as a conference paper at ICLR 2020
Figure 3: Pearson correlation of FBERT computed with different models, across different layers, with
segment-level human judgments on the WMT16 to-English machine translation task. The WMT17
English-Chinese data is used for the BERT Chinese model. Layer 0 corresponds to using BPE
embeddings. Consistently, correlation drops significantly in the final layers.
18


Published as a conference paper at ICLR 2020
C
ABLATION STUDY OF MOVERSCORE
Word Mover’s Distance (WMD; Kusner et al., 2015) is a semantic similarity metric that relies on
word embeddings and optimal transport. MOVERSCORE (Zhao et al., 2019) combines contextual
embeddings and WMD for text generation evaluation. In contrast, BERTSCORE adopts a greedy
approach to aggregate token-level information. In addition to using WMD for generation evalu-
ation, Zhao et al. (2019) also introduce various other improvements. We do a detailed ablation
study to understand the benefit of each improvement, and to investigate whether it can be applied to
BERTSCORE. We use a 12-layer uncased BERT model on the WMT17 to-English segment-level
data, the same setting as Zhao et al. (2019).
We identify several differences between MOVERSCORE and BERTSCORE by analyzing the released
source code. We isolate each difference, and mark it with a bracketed tag for our ablation study:
1. [MNLI] Use a BERT model fine-tuned on MNLI (Williams et al., 2018).
2. [PMEANS] Apply power means (Rücklé et al., 2018) to aggregate the information of dif-
ferent layers.9
3. [IDF-L] For reference sentences, instead of computing the idf scores on the 560 sen-
tences in the segment-level data ([IDF-S]), compute the idf scores on the 3,005 sentences
in the system-level data.
4. [SEP] For candidate sentences, recompute the idf scores on the candidate sentences. The
weighting of reference tokens are kept the same as in [IDF-S]
5. [RM] Exclude punctuation marks and sub-word tokens except the first sub-word in each
word from the matching.
We follow the setup of Zhao et al. (2019) and use their released fine-tuned BERT model to
conduct the experiments.
Table 9 shows the results of our ablation study.
We report corre-
lations for the two variants of WMD Zhao et al. (2019) study:
unigrams (WMD1) and bi-
grams (WMD2). Our FBERT corresponds to the vanilla setting and the importance weighted vari-
ant corresponds to the [IDF-S] setting.
The complete MOVERSCORE metric corresponds to
[IDF-S]+[SEP]+[PMEANS]+[MNLI]+[RM]. We make several observations. First, for all lan-
guage pairs except fi-en and lv-en, we can replicate the reported performance. For these two lan-
guage pairs, Zhao et al. (2019) did not release their implementations at the time of publication.10
Second, we confirm the effectiveness of [PMEANS] and [MNLI]. In Appendix F, we study more
pre-trained models and further corroborate this conclusion. However, the contribution of other tech-
niques, including [RM] and [SEP], seems less stable. Third, replacing greedy matching with
WMD does not lead to consistent improvement. In fact, oftentimes BERTSCORE is the better met-
ric when given the same setup. In general, for any given language pair, BERTSCORE is always
among the best performing ones. Given the current results, it is not clear tht WMD is better than
greedy matching for text generation evaluation.
9 Zhao et al. (2019) uses the embeddings from the last five layers from BERT and L2-normalizes the embed-
ding vectors at each layer before computing the P-MEANs and L2-normalizing the concatenated P-MEANS.
10A public comment on the project page indicates that some of the techniques are not applied for these two
language pairs (https://github.com/AIPHES/emnlp19-moverscore/issues/1).
19


Published as a conference paper at ICLR 2020
Ablation
Metric
cs-en
de-en
fi-en
lv-en
ru-en
tr-en
zh-en
Vanilla
WMD1
0.628
0.655
0.795
0.692
0.701
0.715
0.699
WMD2
0.638
0.661
0.797
0.695
0.700
0.728
0.714
FBERT
0.659
0.680
0.817
0.702
0.719
0.727
0.717
IDF-S
WMD1
0.636
0.662
0.824
0.709
0.716
0.728
0.713
WMD2
0.643
0.662
0.821
0.708
0.712
0.732
0.715
FBERT
0.657
0.681
0.823
0.713
0.725
0.718
0.711
IDF-L
WMD1
0.633
0.659
0.825
0.708
0.716
0.727
0.715
WMD2
0.641
0.661
0.822
0.708
0.713
0.730
0.716
FBERT
0.655
0.682
0.823
0.713
0.726
0.718
0.712
IDF-L + SEP
WMD1
0.651
0.660
0.819
0.703
0.714
0.724
0.715
WMD2
0.659
0.662
0.816
0.702
0.712
0.729
0.715
FBERT
0.664
0.681
0.818
0.709
0.724
0.716
0.710
IDF-L + SEP
+ RM
WMD1
0.651
0.686
0.803
0.681
0.730
0.730
0.720
WMD2
0.664
0.687
0.797
0.679
0.728
0.735
0.718
FBERT
0.659
0.695
0.800
0.683
0.734
0.722
0.712
IDF-L + SEP
+ PMEANS
WMD1
0.658
0.663
0.820
0.707
0.717
0.725
0.712
WMD2
0.667
0.665
0.817
0.707
0.717
0.727
0.712
FBERT
0.671
0.682
0.819
0.708
0.725
0.715
0.704
IDF-L + SEP
+ MNLI
WMD1
0.659
0.679
0.822
0.732
0.718
0.746
0.725
WMD2
0.664
0.682
0.819
0.731
0.715
0.748
0.722
FBERT
0.668
0.701
0.825
0.737
0.727
0.744
0.725
IDF-L + SEP
+ PMEANS + MNLI
WMD1
0.672
0.686
0.831
0.738
0.725
0.753
0.737
WMD2
0.677
0.690
0.828
0.736
0.722
0.755
0.735
FBERT
0.682
0.707
0.836
0.741
0.732
0.751
0.736
IDF-L + SEP
+ PMEANS + MNLI
+ RM
WMD1
0.670
0.708
0.821
0.717
0.738
0.762
0.744
WMD2
0.679
0.709
0.814
0.716
0.736
0.762
0.738
FBERT
0.676
0.717
0.824
0.719
0.740
0.757
0.738
Table 9: Ablation Study of MOVERSCORE and BERTSCORE using Pearson correlations on the
WMT17 to-English segment-level data. Correlations that are not outperformed by others for that
language pair under Williams Test are bolded. We observe that using WMD does not consistently
improve BERTSCORE.
20


Published as a conference paper at ICLR 2020
Type
Metric
Meaning
Grammar
Combined
BERTSCORE
PBERT
0.36
0.47
0.46
RBERT
0.64
0.29
0.52
FBERT
0.58
0.41
0.56
Common metrics
BLEU
0.46
0.13
0.33
METEOR
0.53
0.11
0.36
ROUGE-L
0.51
0.16
0.38
SARI
0.50
0.15
0.37
Best metrics according to
Toutanova et al. (2016)
SKIP-2+RECALL+MULT-PROB
0.59
N/A
0.51
PARSE-2+RECALL+MULT-MAX
N/A
0.35
0.52
PARSE-2+RECALL+MULT-PROB
0.57
0.35
0.52
Table 10: Pearson correlations with human judgments on the MSR Abstractive Text Compression
Dataset.
D
ADDITIONAL EXPERIMENTS ON ABSTRACTIVE TEXT COMPRESSION
We use the human judgments provided from the MSR Abstractive Text Compression
Dataset (Toutanova et al., 2016) to illustrate the applicability of BERTSCORE to abstractive text
compression evaluation. The data includes three types of human scores: (a) meaning: how well a
compressed text preserve the meaning of the original text; (b) grammar: how grammatically correct
a compressed text is; and (c) combined: the average of the meaning and the grammar scores. We
follow the experimental setup of Toutanova et al. (2016) and report Pearson correlation between
BERTSCORE and the three types of human scores. Table 10 shows that RBERT has the highest cor-
relation with human meaning judgments, and PBERT correlates highly with human grammar judg-
ments. FBERT provides a balance between the two aspects.
21


Published as a conference paper at ICLR 2020
Task
Model
BLEU
ˆ
PBERT
ˆ
RBERT
ˆ
FBERT
PBERT
RBERT
FBERT
WMT14
En-De
ConvS2S (Gehring et al., 2017)
0.266
0.6099 0.6055 0.6075 0.8499 0.8482 0.8488
Transformer-big∗∗(Ott et al., 2018)
0.298
0.6587 0.6528 0.6558 0.8687 0.8664 0.8674
DynamicConv∗∗∗(Wu et al., 2019)
0.297
0.6526 0.6464 0.6495 0.8664 0.8640 0.8650
WMT14
En-Fr
ConvS2S (Gehring et al., 2017)
0.408
0.6998 0.6821 0.6908 0.8876 0.8810 0.8841
Transformer-big (Ott et al., 2018)
0.432
0.7148 0.6978 0.7061 0.8932 0.8869 0.8899
DynamicConv (Wu et al., 2019)
0.432
0.7156 0.6989 0.7071 0.8936 0.8873 0.8902
IWSLT14
De-En
Transformer-iwslt+ (Ott et al., 2019) 0.350
0.6749 0.6590 0.6672 0.9452 0.9425 0.9438
LightConv (Wu et al., 2019)
0.348
0.6737 0.6542 0.6642 0.9450 0.9417 0.9433
DynamicConv (Wu et al., 2019)
0.352
0.6770 0.6586 0.6681 0.9456 0.9425 0.9440
Table 11:
BLEU scores and BERTSCOREs of publicly available pre-trained MT models in
fairseq (Ott et al., 2019). We show both rescaled scores marked with ˆ and raw BERTSCOREs.
∗: trained on unconfirmed WMT data version, ∗∗: trained on WMT16 + ParaCrawl, ∗∗∗: trained on
WMT16, +: trained by us using fairseq.
E
BERTSCORE OF RECENT MT MODELS
Table 11 shows the BLEU scores and the BERTSCOREs of pre-trained machine translation models
on WMT14 English-to-German, WMT14 English-to-French, IWSLT14 German-to-English task.
We used publicly available pre-trained models from fairseq (Ott et al., 2019).11 Because a pre-
trained Transformer model on IWSLT is not released, we trained our own using the fairseq library.
We use multilingual cased BERTbase12 for English-to-German and English-to-French pairs, and
English uncased BERTbase13 for German-to-English pairs. Interestingly, the gap between a Dy-
namicConv (Wu et al., 2019) trained on only WMT16 and a Transformer (Ott et al., 2018) trained
on WMT16 and ParaCrawl14 (about 30× more training data) becomes larger when evaluated with
BERTSCORE rather than BLEU.
11 Code and pre-trained model available at https://github.com/pytorch/fairseq.
12Hash code: bert-base-multilingual-cased_L9_version=0.2.0
13Hash code: roberta-large_L17_version=0.2.0
14http://paracrawl.eu/download.html
22


Published as a conference paper at ICLR 2020
F
ADDITIONAL RESULTS
In this section, we present additional experimental results:
1. Segment-level and system-level correlation studies on three years of WMT metric evalua-
tion task (WMT16–18)
2. Model selection study on WMT18 10K hybrid systems
3. System-level correlation study on 2015 COCO captioning challenge
4. Robustness study on PAWS-QQP.
Following BERT (Devlin et al., 2019), a variety of Transformer-based (Vaswani et al., 2017) pre-
trained contextual embeddings have been proposed and released. We conduct additional experiments
with four types of pre-trained embeddings: BERT, XLM (Lample & Conneau, 2019), XLNet (Yang
et al., 2019b), and RoBERTa (Liu et al., 2019b). XLM (Cross-lingual Language Model) is a Trans-
former pre-trained on the translation language modeling of predicting masked tokens from a pair of
sentence in two different languages and masked language modeling tasks using multi-lingual train-
ing data. Yang et al. (2019b) modify the Transformer architecture and pre-train it on a permutation
language modeling task resulting in some improvement on top of the original BERT when fine-tuned
on several downstream tasks. Liu et al. (2019b) introduce RoBERTa (Robustly optimized BERT ap-
proach) and demonstrate that an optimized BERT model is comparable to or sometimes outperforms
an XLNet on downstream tasks.
We perform a comprehensive study with the following pre-trained contextual embedding models:15
• BERT
models:
bert-base-uncased,
bert-large-uncased,
bert-based-chinese,
bert-base-multilingual-cased,
and
bert-base-cased-mrpc
• RoBERTa models: roberta-base, roberta-large, and roberta-large-mnli
• XLNet models: xlnet-base-cased and xlnet-base-large
• XLM models: xlm-mlm-en-2048 and xlm-mlm-100-1280
F.1
WMT CORRELATION STUDY
Experimental setup
Because of missing data in the released WMT16 dataset (Bojar et al., 2016),
we are only able to experiment with to-English segment-level data, which contains the outputs of
50 different systems on 6 language pairs. We use this data as the validation set for hyperparam-
eter tuning (Appendix B). Table 12 shows the Pearson correlations of all participating metrics and
BERTSCOREs computed with different pre-trained models. Significance testing for this dataset does
not include the baseline metrics because the released dataset does not contain the original outputs
from the baseline metrics. We conduct significance testing between BERTSCORE results only.
The WMT17 dataset (Bojar et al., 2017) contains outputs of 152 different translations on 14 lan-
guage pairs. We experiment on the segment-level and system-level data on both to-English and
from-English language pairs. We exclude fi-en data from the segment-level experiment due to an
error in the released data. We compare our results to all participating metrics and perform standard
significance testing as done by Bojar et al. (2017). Tables 13–16 show the results.
The WMT18 dataset (Ma et al., 2018) contains outputs of 159 translation systems on 14 lan-
guage pairs. In addition to the results in Tables 1–4, we complement the study with the correla-
tions of all participating metrics in WMT18 and results from using different contextual models for
BERTSCORE.
Results
Table 12–22 collectively showcase the effectiveness of BERTSCORE in correlating with
human judgments. The improvement of BERTSCORE is more pronounced on the segment-level
than on the system-level. We also see that more optimized or larger BERT models can produce better
contextual representations (e.g., comparing FRoBERTa–Large and FBERT–Large). In contrast, the smaller
XLNet performs better than a large one. Based on the evidence in Figure 8 and Tables 12–22, we
15Denoted by names specified at https://huggingface.co/pytorch-transformers/pretrained_models.html.
23


Published as a conference paper at ICLR 2020
hypothesize that the permutation language task, though leading to a good set of model weights for
fine-tuning on downstream tasks, does not necessarily produce informative pre-trained embeddings
for generation evaluation. We also observe that fine-tuning pre-trained models on a related task,
such as natural language inference (Williams et al., 2018), can lead to better human correlation in
evaluating text generation. Therefore, for evaluating English sentences, we recommend computing
BERTSCORE with a 24-layer RoBERTa model fine-tuned on the MNLI dataset. For evaluating
Non-English sentences, both the multilingual BERT model and the XLM model trained on 100
languages are suitable candidates. We also recommend using domain- or language-specific contex-
tual embeddings when possible, such as using BERT Chinese models for evaluating Chinese tasks.
In general, we advise users to consider the target domain and languages when selecting the exact
configuration to use.
F.2
MODEL SELECTION STUDY
Experimental setup
Similar to Section 4, we use the 10K hybrid systems super-sampled from
WMT18. We randomly select 100 out of 10K hybrid systems, rank them using automatic metrics,
and repeat this process 100K times. We add to the results in the main paper (Table 3) performance of
all participating metrics in WMT18 and results from using different contextual embedding models
for BERTSCORE. We reuse the hybrid configuration and metric outputs released in WMT18. In
addition to the Hits@1 measure, we evaluate the metrics using (a) mean reciprocal rank (MRR) of
the top metric-rated system in human rankings, and (b) the absolute human score difference (Diff)
between the top metric- and human-rated systems. Hits@1 captures a metric’s ability to select the
best system. The other two measures quantify the amount of error a metric makes in the selection
process. Tables 23–28 show the results from these experiments.
Results
The additional results further support our conclusion from Table 3: BERTSCORE demon-
strates better model selection performance. We also observe that the supervised metric RUSE dis-
plays strong model selection ability.
F.3
IMAGE CAPTIONING ON COCO
We follow the experimental setup described in Section 4. Table 29 shows the correlations of several
pre-trained contextual embeddings. We observe that precision-based methods such as BLEU and
PBERT are weakly correlated with human judgments on image captioning tasks. We hypothesize
that this is because human judges prefer captions that capture the main objects in a picture for
image captioning. In general, RBERT has a high correlation, even surpassing the task-specific metric
SPICE Anderson et al. (2016). While the fine-tuned RoBERTa-Large model does not result in the
highest correlation, it is one of the best metrics.
F.4
ROBUSTNESS ANALYSIS ON PAWS-QQP
We present the full results of the robustness study described in Section 6 in Table 30. In general,
we observe that BERTSCORE is more robust than other commonly used metrics. BERTSCORE
computed with the 24-layer RoBERTa model performs the best. Fine-tuning RoBERTa-Large on
MNLI (Williams et al., 2018) can significantly improve the robustness against adversarial sentences.
However, a fine-tuned BERT on MRPC (Microsoft Research Paraphrasing Corpus) (Dolan & Brock-
ett, 2005) performs worse than its counterpart.
24


Published as a conference paper at ICLR 2020
Setting
Metric
cs-en
de-en
fi-en
ro-en
ru-en
tr-en
560
560
560
560
560
560
Unsupervised
DPMFCOMB
0.713
0.584
0.598
0.627
0.615
0.663
METRICS-F
0.696
0.601
0.557
0.662
0.618
0.649
COBALT-F.
0.671
0.591
0.554
0.639
0.618
0.627
UPF-COBA.
0.652
0.550
0.490
0.616
0.556
0.626
MPEDA
0.644
0.538
0.513
0.587
0.545
0.616
CHRF2
0.658
0.457
0.469
0.581
0.534
0.556
CHRF3
0.660
0.455
0.472
0.582
0.535
0.555
CHRF1
0.644
0.454
0.452
0.570
0.522
0.551
UOW-REVAL
0.577
0.528
0.471
0.547
0.528
0.531
WORDF3
0.599
0.447
0.473
0.525
0.504
0.536
WORDF2
0.596
0.445
0.471
0.522
0.503
0.537
WORDF1
0.585
0.435
0.464
0.508
0.497
0.535
SENTBLEU
0.557
0.448
0.484
0.499
0.502
0.532
DTED
0.394
0.254
0.361
0.329
0.375
0.267
Supervised
BEER
0.661
0.462
0.471
0.551
0.533
0.545
Pre-Trained
PBERT–Base
0.729
0.617
0.719
0.651
0.684
0.678
RBERT–Base
0.741
0.639
0.616
0.693
0.660
0.660
FBERT–Base
0.747
0.640
0.661
0.723
0.672
0.688
PBERT–Base (no idf)
0.723
0.638
0.662
0.700
0.633
0.696
RBERT–Base (no idf)
0.745
0.656
0.638
0.697
0.653
0.674
FBERT–Base (no idf)
0.747
0.663
0.666
0.714
0.662
0.703
PBERT–Base–MRPC
0.697
0.618
0.614
0.676
0.62
0.695
RBERT–Base–MRPC
0.723
0.636
0.587
0.667
0.648
0.664
FBERT–Base–MRPC
0.725
0.644
0.617
0.691
0.654
0.702
PBERT–Base–MRPC (idf)
0.713
0.613
0.630
0.693
0.635
0.691
RBERT–Base–MRPC (idf)
0.727
0.631
0.573
0.666
0.642
0.662
FBERT–Base–MRPC (idf)
0.735
0.637
0.620
0.700
0.658
0.697
PBERT–Large
0.756
0.671
0.701
0.723
0.678
0.706
RBERT–Large
0.768
0.684
0.677
0.720
0.686
0.699
FBERT–Large
0.774
0.693
0.705
0.736
0.701
0.717
PBERT–Large (idf)
0.758
0.653
0.704
0.734
0.685
0.705
RBERT–Large (idf)
0.771
0.680
0.661
0.718
0.687
0.692
FBERT–Large (idf)
0.774
0.678
0.700
0.740
0.701
0.711
PRoBERTa–Base
0.738
0.642
0.671
0.712
0.669
0.671
RRoBERTa–Base
0.745
0.669
0.645
0.698
0.682
0.653
FRoBERTa–Base
0.761
0.674
0.686
0.732
0.697
0.689
PRoBERTa–Base (idf)
0.751
0.626
0.678
0.723
0.685
0.668
RRoBERTa–Base (idf)
0.744
0.652
0.638
0.699
0.685
0.657
FRoBERTa–Base (idf)
0.767
0.653
0.688
0.737
0.705
0.685
PRoBERTa–Large
0.757
0.702
0.709
0.735
0.721
0.676
RRoBERTa–Large
0.765
0.713
0.686
0.718
0.714
0.676
FRoBERTa–Large
0.780
0.724
0.728
0.753
0.738
0.709
PRoBERTa–Large (idf)
0.771
0.682
0.705
0.727
0.714
0.681
RRoBERTa–Large (idf)
0.762
0.695
0.683
0.711
0.708
0.678
FRoBERTa–Large (idf)
0.786
0.704
0.727
0.747
0.732
0.711
PRoBERTa–Large–MNLI
0.777
0.718
0.733
0.744
0.729
0.747
RRoBERTa–Large–MNLI
0.790
0.731
0.702
0.741
0.727
0.732
FRoBERTa–Large–MNLI
0.795
0.736
0.733
0.757
0.744
0.756
PRoBERTa–Large–MNLI (idf)
0.794
0.695
0.731
0.752
0.732
0.747
RRoBERTa–Large–MNLI (idf)
0.792
0.706
0.694
0.737
0.724
0.733
FRoBERTa–Large–MNLI (idf)
0.804
0.710
0.729
0.760
0.742
0.754
PXLNet–Base
0.708
0.612
0.639
0.650
0.606
0.690
RXLNet–Base
0.728
0.630
0.617
0.645
0.621
0.675
FXLNet–Base
0.727
0.631
0.640
0.659
0.626
0.695
PXLNet–Base (idf)
0.726
0.618
0.655
0.678
0.629
0.700
RXLNet–Base (idf)
0.734
0.633
0.618
0.66
0.635
0.682
FXLNet–Base (idf)
0.739
0.633
0.649
0.681
0.643
0.702
PXL-NET–LARGE
0.710
0.577
0.643
0.647
0.616
0.684
RXL-NET–LARGE
0.732
0.600
0.610
0.636
0.627
0.668
FXL-NET–LARGE
0.733
0.600
0.643
0.655
0.637
0.691
PXL-NET–LARGE (idf)
0.728
0.574
0.652
0.669
0.633
0.681
RXL-NET–LARGE (idf)
0.735
0.592
0.597
0.642
0.629
0.662
FXL-NET–LARGE (idf)
0.742
0.592
0.643
0.670
0.645
0.685
PXLM–En
0.688
0.569
0.613
0.645
0.583
0.659
RXLM–En
0.715
0.603
0.577
0.645
0.609
0.644
FXLM–En
0.713
0.597
0.610
0.657
0.610
0.668
PXLM–En (idf)
0.728
0.576
0.649
0.681
0.604
0.683
RXLM–En (idf)
0.730
0.597
0.591
0.659
0.622
0.669
FXLM–En (idf)
0.739
0.594
0.636
0.682
0.626
0.691
Table 12: Pearson correlations with segment-level human judgments on WMT16 to-English trans-
lations. Correlations of metrics not significantly outperformed by any other for that language pair
are highlighted in bold. For each language pair, we specify the number of examples.
25


Published as a conference paper at ICLR 2020
Setting
Metric
cs-en
de-en
fi-en
lv-en
ru-en
tr-en
zh-en
560
560
560
560
560
560
560
Unsupervised
CHRF
0.514
0.531
0.671
0.525
0.599
0.607
0.591
CHRF++
0.523
0.534
0.678
0.520
0.588
0.614
0.593
MEANT 2.0
0.578
0.565
0.687
0.586
0.607
0.596
0.639
MEANT 2.0-NOSRL
0.566
0.564
0.682
0.573
0.591
0.582
0.630
SENTBLEU
0.435
0.432
0.571
0.393
0.484
0.538
0.512
TREEAGGREG
0.486
0.526
0.638
0.446
0.555
0.571
0.535
UHH_TSKM
0.507
0.479
0.600
0.394
0.465
0.478
0.477
Supervised
AUTODA
0.499
0.543
0.673
0.533
0.584
0.625
0.583
BEER
0.511
0.530
0.681
0.515
0.577
0.600
0.582
BLEND
0.594
0.571
0.733
0.577
0.622
0.671
0.661
BLEU2VEC
0.439
0.429
0.590
0.386
0.489
0.529
0.526
NGRAM2VEC
0.436
0.435
0.582
0.383
0.490
0.538
0.520
Pre-Trained
PBERT–Base
0.625
0.659
0.808
0.688
0.698
0.713
0.675
RBERT–Base
0.653
0.645
0.782
0.662
0.678
0.716
0.715
FBERT–Base
0.654
0.671
0.811
0.692
0.707
0.731
0.714
PBERT–Base (idf)
0.626
0.668
0.819
0.708
0.719
0.702
0.667
RBERT–Base (idf)
0.652
0.658
0.789
0.678
0.696
0.703
0.712
FBERT–Base (idf)
0.657
0.680
0.823
0.712
0.725
0.718
0.711
PBERT–Base–MRPC
0.599
0.630
0.788
0.657
0.659
0.710
0.681
RBERT–Base–MRPC
0.613
0.620
0.754
0.616
0.650
0.685
0.705
FBERT–Base–MRPC
0.627
0.647
0.792
0.656
0.676
0.717
0.712
PBERT–Base–MRPC (idf)
0.609
0.630
0.801
0.680
0.676
0.712
0.682
RBERT–Base–MRPC (idf)
0.611
0.628
0.759
0.633
0.665
0.687
0.703
FBERT–Base–MRPC (idf)
0.633
0.649
0.803
0.678
0.690
0.719
0.713
PBERT–Large
0.638
0.685
0.816
0.717
0.719
0.746
0.693
RBERT–Large
0.661
0.676
0.782
0.693
0.705
0.744
0.730
FBERT–Large
0.666
0.701
0.814
0.723
0.730
0.760
0.731
PBERT–Large (idf)
0.644
0.692
0.827
0.728
0.729
0.734
0.689
RBERT–Large (idf)
0.665
0.686
0.796
0.712
0.729
0.733
0.730
FBERT–Large (idf)
0.671
0.707
0.829
0.738
0.745
0.746
0.729
PRoBERTa–Base
0.639
0.663
0.801
0.689
0.688
0.700
0.704
RRoBERTa–Base
0.648
0.652
0.768
0.651
0.669
0.684
0.734
FRoBERTa–Base
0.675
0.683
0.818
0.693
0.707
0.718
0.740
PRoBERTa–Base (idf)
0.629
0.655
0.804
0.702
0.711
0.707
0.700
RRoBERTa–Base (idf)
0.652
0.646
0.773
0.667
0.676
0.689
0.734
FRoBERTa–Base (idf)
0.673
0.673
0.823
0.708
0.719
0.721
0.739
PRoBERTa–Large
0.658
0.724
0.811
0.743
0.727
0.720
0.744
RRoBERTa–Large
0.685
0.714
0.778
0.711
0.718
0.713
0.759
FRoBERTa–Large
0.710
0.745
0.833
0.756
0.746
0.751
0.775
PRoBERTa–Large (idf)
0.644
0.721
0.815
0.740
0.734
0.736
0.734
RRoBERTa–Large (idf)
0.683
0.705
0.783
0.718
0.720
0.726
0.751
FRoBERTa–Large (idf)
0.703
0.737
0.838
0.761
0.752
0.764
0.767
PRoBERTa–Large–MNLI
0.694
0.736
0.822
0.764
0.741
0.754
0.737
RRoBERTa–Large–MNLI
0.706
0.725
0.785
0.732
0.741
0.750
0.760
FRoBERTa–Large–MNLI
0.722
0.747
0.822
0.764
0.758
0.767
0.765
PRoBERTa–Large–MNLI (idf)
0.686
0.733
0.836
0.772
0.760
0.767
0.738
RRoBERTa–Large–MNLI (idf)
0.697
0.717
0.796
0.741
0.753
0.757
0.762
FRoBERTa–Large–MNLI (idf)
0.714
0.740
0.835
0.774
0.773
0.776
0.767
PXLNET–Base
0.595
0.579
0.779
0.632
0.626
0.688
0.646
RXLNET–Base
0.603
0.560
0.746
0.617
0.624
0.689
0.677
FXLNET–Base
0.610
0.580
0.775
0.636
0.639
0.700
0.675
PXLNET–Base (idf)
0.616
0.603
0.795
0.665
0.659
0.693
0.649
RXLNET–Base (idf)
0.614
0.583
0.765
0.640
0.648
0.697
0.688
FXLNET–Base (idf)
0.627
0.603
0.795
0.663
0.665
0.707
0.684
PXLNET–Large
0.620
0.622
0.796
0.648
0.648
0.694
0.660
RXLNET–Large
0.622
0.601
0.758
0.628
0.645
0.684
0.701
FXLNET–Large
0.635
0.627
0.794
0.654
0.664
0.705
0.698
PXLNET–Large (idf)
0.635
0.633
0.808
0.673
0.672
0.688
0.649
RXLNET–Large (idf)
0.626
0.611
0.770
0.646
0.661
0.682
0.700
FXLNET–Large (idf)
0.646
0.636
0.809
0.675
0.682
0.700
0.695
PXLM–En
0.565
0.594
0.769
0.631
0.649
0.672
0.643
RXLM–En
0.592
0.586
0.734
0.618
0.647
0.673
0.686
FXLM–En
0.595
0.605
0.768
0.641
0.664
0.686
0.683
PXLM–En (idf)
0.599
0.618
0.795
0.670
0.686
0.690
0.657
RXLM–En (idf)
0.624
0.605
0.768
0.652
0.680
0.684
0.698
FXLM–En (idf)
0.630
0.624
0.798
0.676
0.698
0.698
0.694
Table 13: Absolute Pearson correlations with segment-level human judgments on WMT17 to-
English translations. Correlations of metrics not significantly outperformed by any other for that
language pair are highlighted in bold. For each language pair, we specify the number of examples.
26


Published as a conference paper at ICLR 2020
Setting
Metric
en-cs
en-de
en-fi
en-lv
en-ru
en-tr
en-zh
32K
3K
3K
3K
560
247
560
τ
τ
τ
τ
|r|
τ
|r|
Unsupervised
AUTODA
0.041
0.099
0.204
0.130
0.511
0.409
0.609
AUTODA-TECTO
0.336
-
-
-
-
-
-
CHRF
0.376
0.336
0.503
0.420
0.605
0.466
0.608
CHRF+
0.377
0.325
0.514
0.421
0.609
0.474
-
CHRF++
0.368
0.328
0.484
0.417
0.604
0.466
0.602
MEANT 2.0
-
0.350
-
-
-
-
0.727
MEANT 2.0-NOSRL
0.395
0.324
0.565
0.425
0.636
0.482
0.705
SENTBLEU
0.274
0.269
0.446
0.259
0.468
0.377
0.642
TREEAGGREG
0.361
0.305
0.509
0.383
0.535
0.441
0.566
Supervised
BEER
0.398
0.336
0.557
0.420
0.569
0.490
0.622
BLEND
-
-
-
-
0.578
-
-
BLEU2VEC
0.305
0.313
0.503
0.315
0.472
0.425
-
NGRAM2VEC
-
-
0.486
0.317
-
-
-
Pre-Trained
PBERT–Multi
0.412
0.364
0.561
0.435
0.606
0.579
0.759
RBERT–Multi
0.443
0.430
0.587
0.480
0.663
0.571
0.804
FBERT–Multi
0.440
0.404
0.587
0.466
0.653
0.587
0.806
PBERT–Multi (idf)
0.411
0.328
0.568
0.444
0.616
0.555
0.741
RBERT–Multi (idf)
0.449
0.416
0.591
0.479
0.665
0.579
0.796
FBERT–Multi (idf)
0.447
0.379
0.588
0.470
0.657
0.571
0.793
PXLM–100
0.406
0.383
0.553
0.423
0.562
0.611
0.722
RXLM–100
0.446
0.436
0.587
0.458
0.626
0.652
0.779
FXLM–100
0.444
0.424
0.577
0.456
0.613
0.628
0.778
PXLM–100 (idf)
0.419
0.367
0.557
0.427
0.571
0.595
0.719
RXLM–100 (idf)
0.450
0.424
0.592
0.464
0.632
0.644
0.770
FXLM–100 (idf)
0.448
0.419
0.580
0.459
0.617
0.644
0.771
Table 14: Absolute Pearson correlation (|r|) and Kendall correlation (τ) with segment-level human
judgments on WMT17 from-English translations. Correlations of metrics not significantly outper-
formed by any other for that language pair are highlighted in bold. For each language pair, we
specify the number of examples.
27


Published as a conference paper at ICLR 2020
Setting
Metric
cs-en
de-en
fi-en
lv-en
ru-en
tr-en
zh-en
4
11
6
9
9
10
16
Unsupervised
BLEU
0.971
0.923
0.903
0.979
0.912
0.976
0.864
CDER
0.989
0.930
0.927
0.985
0.922
0.973
0.904
CHARACTER
0.972
0.974
0.946
0.932
0.958
0.949
0.799
CHRF
0.939
0.968
0.938
0.968
0.952
0.944
0.859
CHRF++
0.940
0.965
0.927
0.973
0.945
0.960
0.880
MEANT 2.0
0.926
0.950
0.941
0.970
0.962
0.932
0.838
MEANT 2.0-NOSRL
0.902
0.936
0.933
0.963
0.960
0.896
0.800
NIST
1.000
0.931
0.931
0.960
0.912
0.971
0.849
PER
0.968
0.951
0.896
0.962
0.911
0.932
0.877
TER
0.989
0.906
0.952
0.971
0.912
0.954
0.847
TREEAGGREG
0.983
0.920
0.977
0.986
0.918
0.987
0.861
UHH_TSKM
0.996
0.937
0.921
0.990
0.914
0.987
0.902
WER
0.987
0.896
0.948
0.969
0.907
0.925
0.839
Supervised
AUTODA
0.438
0.959
0.925
0.973
0.907
0.916
0.734
BEER
0.972
0.960
0.955
0.978
0.936
0.972
0.902
BLEND
0.968
0.976
0.958
0.979
0.964
0.984
0.894
BLEU2VEC
0.989
0.936
0.888
0.966
0.907
0.961
0.886
NGRAM2VEC
0.984
0.935
0.890
0.963
0.907
0.955
0.880
Pre-Trained
PBERT–Base
0.975
0.936
0.991
0.993
0.918
0.981
0.892
RBERT–Base
0.995
0.975
0.944
0.978
0.953
0.991
0.975
FBERT–Base
0.987
0.961
0.979
0.991
0.937
0.991
0.953
PBERT–Base (idf)
0.983
0.937
0.998
0.992
0.939
0.985
0.878
RBERT–Base (idf)
0.997
0.981
0.962
0.968
0.977
0.985
0.949
FBERT–Base (idf)
0.992
0.967
0.995
0.992
0.960
0.996
0.951
PBERT–Base–MRPC
0.982
0.926
0.990
0.987
0.916
0.970
0.899
RBERT–Base–MRPC
0.999
0.979
0.950
0.982
0.957
0.977
0.985
FBERT–Base–MRPC
0.994
0.957
0.986
0.994
0.938
0.980
0.960
PBERT–Base–MRPC (idf)
0.989
0.936
0.992
0.979
0.931
0.976
0.892
RBERT–Base–MRPC (idf)
0.999
0.987
0.962
0.980
0.975
0.979
0.973
FBERT–Base–MRPC (idf)
0.997
0.968
0.995
0.997
0.956
0.989
0.963
PBERT–Large
0.981
0.937
0.991
0.996
0.921
0.987
0.905
RBERT–Large
0.996
0.975
0.953
0.985
0.954
0.992
0.977
FBERT–Large
0.990
0.960
0.981
0.995
0.938
0.992
0.957
PBERT–Large (idf)
0.986
0.938
0.998
0.995
0.939
0.994
0.897
RBERT–Large (idf)
0.997
0.982
0.967
0.979
0.974
0.992
0.966
FBERT–Large (idf)
0.994
0.965
0.993
0.995
0.958
0.998
0.959
PRoBERTa–Base
0.987
0.930
0.984
0.966
0.916
0.963
0.955
RRoBERTa–Base
0.999
0.982
0.947
0.979
0.956
0.986
0.984
FRoBERTa–Base
0.996
0.961
0.993
0.993
0.937
0.983
0.982
PRoBERTa–Base (idf)
0.990
0.938
0.980
0.956
0.929
0.967
0.962
RRoBERTa–Base (idf)
0.998
0.987
0.963
0.979
0.971
0.986
0.974
FRoBERTa–Base (idf)
0.996
0.970
0.999
0.994
0.952
0.989
0.982
PRoBERTa–Large
0.989
0.948
0.984
0.949
0.927
0.960
0.967
RRoBERTa–Large
0.998
0.988
0.957
0.983
0.969
0.982
0.984
FRoBERTa–Large
0.996
0.973
0.997
0.991
0.949
0.984
0.987
PRoBERTa–Large (idf)
0.989
0.959
0.975
0.935
0.944
0.968
0.974
RRoBERTa–Large (idf)
0.995
0.991
0.962
0.979
0.981
0.981
0.970
FRoBERTa–Large (idf)
0.996
0.982
0.998
0.991
0.965
0.991
0.984
PRoBERTa–Large–MNLI
0.994
0.963
0.995
0.990
0.944
0.981
0.974
RRoBERTa–Large–MNLI
0.995
0.991
0.962
0.981
0.973
0.985
0.984
FRoBERTa–Large–MNLI
0.999
0.982
0.992
0.996
0.961
0.988
0.989
PRoBERTa–Large–MNLI (idf)
0.995
0.970
0.997
0.985
0.955
0.988
0.979
RRoBERTa–Large–MNLI (idf)
0.994
0.992
0.967
0.977
0.983
0.988
0.972
FRoBERTa–Large–MNLI (idf)
0.999
0.989
0.996
0.997
0.972
0.994
0.987
PXLNET–Base
0.988
0.938
0.993
0.993
0.914
0.974
0.960
RXLNET–Base
0.999
0.978
0.956
0.977
0.946
0.981
0.980
FXLNET–Base
0.996
0.963
0.986
0.991
0.932
0.981
0.978
PXLNET–Base (idf)
0.992
0.951
0.998
0.996
0.930
0.982
0.939
RXLNET–Base (idf)
0.999
0.986
0.968
0.973
0.964
0.987
0.955
FXLNET–Base (idf)
0.998
0.974
0.996
0.994
0.950
0.990
0.970
PXLNET–Large
0.991
0.944
0.996
0.995
0.924
0.982
0.943
RXLNET–Large
0.996
0.981
0.945
0.971
0.961
0.986
0.958
FXLNET–Large
0.999
0.969
0.986
0.992
0.945
0.992
0.961
PXLNET–Large (idf)
0.995
0.955
0.999
0.996
0.941
0.985
0.937
RXLNET–Large (idf)
0.993
0.985
0.951
0.960
0.975
0.974
0.910
FXLNET–Large (idf)
1.000
0.978
0.994
0.993
0.962
0.994
0.954
PXLM–En
0.983
0.933
0.994
0.989
0.918
0.973
0.928
RXLM–En
0.998
0.978
0.949
0.983
0.957
0.985
0.972
FXLM–En
0.994
0.960
0.985
0.995
0.938
0.984
0.964
PXLM–En (idf)
0.986
0.940
0.997
0.992
0.939
0.979
0.916
RXLM–En (idf)
0.999
0.983
0.966
0.980
0.975
0.991
0.952
FXLM–En (idf)
0.995
0.967
0.996
0.998
0.959
0.993
0.958
Table 15: Absolute Pearson correlations with system-level human judgments on WMT17 to-English
translations. Correlations of metrics not significantly outperformed by any other for that language
pair are highlighted in bold. For each language pair, we specify the number of systems.
28


Published as a conference paper at ICLR 2020
Setting
Metric
en-cs
en-de
en-lv
en-ru
en-tr
en-zh
14
16
17
9
8
11
Unsupervised
BLEU
0.956
0.804
0.866
0.898
0.924
–
CDER
0.968
0.813
0.930
0.924
0.957
–
CHARACTER
0.981
0.938
0.897
0.939
0.975
0.933
CHRF
0.976
0.863
0.955
0.950
0.991
0.976
CHRF++
0.974
0.852
0.956
0.945
0.986
0.976
MEANT 2.0
–
0.858
–
–
–
0.956
MEANT 2.0-NOSRL
0.976
0.770
0.959
0.957
0.991
0.943
NIST
0.962
0.769
0.935
0.920
0.986
–
PER
0.954
0.687
0.851
0.887
0.963
–
TER
0.955
0.796
0.909
0.933
0.967
–
TREEAGGREG
0.947
0.773
0.927
0.921
0.983
0.938
UHH_TSKM
–
–
–
–
–
–
WER
0.954
0.802
0.906
0.934
0.956
–
Supervised
AUTODA
0.975
0.603
0.729
0.850
0.601
0.976
BEER
0.970
0.842
0.930
0.944
0.980
0.914
BLEND
–
–
–
0.953
–
–
BLEU2VEC
0.963
0.810
0.859
0.903
0.911
–
NGRAM2VEC
–
–
0.862
–
–
–
Pre-Trained
PBERT–Multi
0.959
0.798
0.960
0.946
0.981
0.970
RBERT–Multi
0.982
0.909
0.957
0.980
0.979
0.994
FBERT–Multi
0.976
0.859
0.959
0.966
0.980
0.992
PBERT–Multi (idf)
0.963
0.760
0.960
0.947
0.984
0.971
RBERT–Multi (idf)
0.985
0.907
0.955
0.981
0.984
0.982
FBERT–Multi (idf)
0.979
0.841
0.958
0.968
0.984
0.991
PXLM–100
0.967
0.825
0.965
0.953
0.974
0.977
RXLM–100
0.980
0.902
0.965
0.982
0.977
0.979
FXLM–100
0.979
0.868
0.969
0.971
0.976
0.986
PXLM–100 (idf)
0.968
0.809
0.965
0.955
0.980
0.975
RXLM–100 (idf)
0.981
0.894
0.964
0.984
0.983
0.968
FXLM–100 (idf)
0.979
0.856
0.966
0.973
0.982
0.979
Table 16: Absolute Pearson correlations with system-level human judgments on WMT17 from-
English translations. Correlations of metrics not significantly outperformed by any other for that
language pair are highlighted in bold. For each language pair, we specify the number of systems.
29


Published as a conference paper at ICLR 2020
Setting
Metric
cs-en
de-en
et-en
fi-en
ru-en
tr-en
zh-en
5K
78K
57K
16K
10K
9K
33K
Unsupervised
CHARACTER
0.256
0.450
0.286
0.185
0.244
0.172
0.202
ITER
0.198
0.396
0.235
0.128
0.139
-0.029
0.144
METEOR++
0.270
0.457
0.329
0.207
0.253
0.204
0.179
SENTBLEU
0.233
0.415
0.285
0.154
0.228
0.145
0.178
UHH_TSKM
0.274
0.436
0.300
0.168
0.235
0.154
0.151
YISI-0
0.301
0.474
0.330
0.225
0.294
0.215
0.205
YISI-1
0.319
0.488
0.351
0.231
0.300
0.234
0.211
YISI-1 SRL
0.317
0.483
0.345
0.237
0.306
0.233
0.209
Supervised
BEER
0.295
0.481
0.341
0.232
0.288
0.229
0.214
BLEND
0.322
0.492
0.354
0.226
0.290
0.232
0.217
RUSE
0.347
0.498
0.368
0.273
0.311
0.259
0.218
Pre-Trained
PBERT–Base
0.349
0.522
0.373
0.264
0.325
0.264
0.232
RBERT–Base
0.370
0.528
0.378
0.291
0.333
0.257
0.244
FBERT–Base
0.373
0.531
0.385
0.287
0.341
0.266
0.243
PBERT–Base (idf)
0.352
0.524
0.382
0.27
0.326
0.277
0.235
RBERT–Base (idf)
0.368
0.536
0.388
0.300
0.340
0.284
0.244
FBERT–Base (idf)
0.375
0.535
0.393
0.294
0.339
0.289
0.243
PBERT–Base–MRPC
0.343
0.520
0.365
0.247
0.333
0.25
0.227
RBERT–Base–MRPC
0.370
0.524
0.373
0.277
0.34
0.261
0.244
FBERT–Base–MRPC
0.366
0.529
0.377
0.271
0.342
0.263
0.242
PBERT–Base–MRPC (idf)
0.348
0.522
0.371
0.25
0.318
0.256
0.224
RBERT–Base–MRPC (idf)
0.379
0.531
0.383
0.285
0.339
0.266
0.242
FBERT–Base–MRPC (idf)
0.373
0.534
0.383
0.274
0.342
0.275
0.242
PBERT–LARGE
0.361
0.529
0.380
0.276
0.340
0.266
0.241
RBERT–LARGE
0.386
0.532
0.386
0.297
0.347
0.268
0.247
FBERT–LARGE
0.402
0.537
0.390
0.296
0.344
0.274
0.252
PBERT–LARGE (idf)
0.377
0.532
0.390
0.287
0.342
0.292
0.246
RBERT–LARGE (idf)
0.386
0.544
0.396
0.308
0.356
0.287
0.251
FBERT–LARGE (idf)
0.388
0.545
0.399
0.309
0.358
0.300
0.257
PRoBERTa–Base
0.368
0.53
0.371
0.274
0.318
0.265
0.235
RRoBERTa–Base
0.383
0.536
0.376
0.283
0.336
0.253
0.245
FRoBERTa–Base
0.391
0.540
0.383
0.273
0.339
0.270
0.249
PRoBERTa–Base (idf)
0.379
0.528
0.372
0.261
0.314
0.265
0.232
RRoBERTa–Base (idf)
0.389
0.539
0.384
0.288
0.332
0.267
0.245
FRoBERTa–Base (idf)
0.400
0.540
0.385
0.274
0.337
0.277
0.247
PRoBERTa–LARGE
0.387
0.541
0.389
0.283
0.345
0.280
0.248
RRoBERTa–LARGE
0.388
0.546
0.391
0.304
0.343
0.290
0.255
FRoBERTa–LARGE
0.404
0.550
0.397
0.296
0.353
0.292
0.264
PRoBERTa–LARGE (idf)
0.391
0.540
0.387
0.280
0.334
0.284
0.252
RRoBERTa–LARGE (idf)
0.386
0.548
0.394
0.305
0.338
0.295
0.252
FRoBERTa–LARGE (idf)
0.408
0.550
0.395
0.293
0.346
0.296
0.260
PRoBERTa–Large–MNLI
0.397
0.549
0.396
0.299
0.351
0.295
0.253
RRoBERTa–Large–MNLI
0.404
0.553
0.393
0.313
0.351
0.279
0.253
FRoBERTa–Large–MNLI
0.418
0.557
0.402
0.312
0.362
0.290
0.258
PRoBERTa–Large–MNLI (idf)
0.414
0.552
0.399
0.301
0.349
0.306
0.249
RRoBERTa–Large–MNLI (idf)
0.412
0.555
0.400
0.316
0.357
0.289
0.258
FRoBERTa–Large–MNLI (idf)
0.417
0.559
0.403
0.309
0.357
0.307
0.258
PXLNet–Base
0.335
0.514
0.359
0.243
0.308
0.247
0.232
RXLNet–Base
0.351
0.515
0.362
0.261
0.311
0.227
0.232
FXLNet–Base
0.351
0.517
0.365
0.257
0.315
0.25
0.237
PXLNet–Base (idf)
0.339
0.516
0.366
0.258
0.307
0.261
0.236
RXLNet–Base (idf)
0.364
0.521
0.371
0.268
0.317
0.242
0.238
FXLNet–Base (idf)
0.355
0.524
0.374
0.265
0.320
0.261
0.241
PXL-NET–LARGE
0.344
0.522
0.371
0.252
0.316
0.264
0.233
RXL-NET–LARGE
0.358
0.524
0.374
0.275
0.332
0.249
0.239
FXL-NET–LARGE
0.357
0.530
0.380
0.265
0.334
0.263
0.238
PXL-NET–LARGE (idf)
0.348
0.520
0.373
0.260
0.319
0.265
0.235
RXL-NET–LARGE (idf)
0.366
0.529
0.378
0.278
0.331
0.266
0.241
FXL-NET–LARGE (idf)
0.375
0.530
0.382
0.274
0.332
0.274
0.240
PXLM–En
0.349
0.516
0.366
0.244
0.310
0.259
0.233
RXLM–En
0.358
0.518
0.364
0.264
0.320
0.244
0.237
FXLM–En
0.358
0.525
0.373
0.259
0.322
0.258
0.238
PXLM–En (idf)
0.355
0.527
0.374
0.254
0.311
0.28
0.238
RXLM–En (idf)
0.362
0.528
0.376
0.274
0.333
0.26
0.24
FXLM–En (idf)
0.367
0.531
0.382
0.273
0.330
0.275
0.246
Table 17: Kendall correlations with segment-level human judgments on WMT18 to-English transla-
tions. Correlations of metrics not significantly outperformed by any other for that language pair are
highlighted in bold. For each language pair, we specify the number of examples.
30


Published as a conference paper at ICLR 2020
Setting
Metric
en-cs
en-de
en-et
en-fi
en-ru
en-tr
en-zh
5K
20K
32K
10K
22K
1K
29K
Unsupervised
CHARACTER
0.414
0.604
0.464
0.403
0.352
0.404
0.313
ITER
0.333
0.610
0.392
0.311
0.291
0.236
-
SENTBLEU
0.389
0.620
0.414
0.355
0.330
0.261
0.311
YISI-0
0.471
0.661
0.531
0.464
0.394
0.376
0.318
YISI-1
0.496
0.691
0.546
0.504
0.407
0.418
0.323
YISI-1 SRL
-
0.696
-
-
-
-
0.310
Supervised
BEER
0.518
0.686
0.558
0.511
0.403
0.374
0.302
BLEND
-
-
-
-
0.394
-
-
Pre-Trained
PBERT–Multi
0.541
0.715
0.549
0.486
0.414
0.328
0.337
RBERT–Multi
0.570
0.728
0.594
0.565
0.420
0.411
0.367
FBERT–Multi
0.562
0.728
0.586
0.546
0.423
0.399
0.364
PBERT–Multi (idf)
0.525
0.7
0.54
0.495
0.423
0.352
0.338
RBERT–Multi (idf)
0.569
0.727
0.601
0.561
0.423
0.420
0.374
FBERT–Multi (idf)
0.553
0.721
0.585
0.537
0.425
0.406
0.366
PXLM–100
0.496
0.711
0.561
0.527
0.417
0.364
0.340
RXLM–100
0.564
0.724
0.612
0.584
0.418
0.432
0.363
FXLM–100
0.533
0.727
0.599
0.573
0.421
0.408
0.362
PXLM–100 (idf)
0.520
0.710
0.572
0.546
0.421
0.370
0.328
RXLM–100 (idf)
0.567
0.722
0.609
0.587
0.420
0.439
0.365
FXLM–100 (idf)
0.554
0.724
0.601
0.584
0.422
0.389
0.355
Table 18: Kendall correlations with segment-level human judgments on WMT18 from-English
translations. Correlations of metrics not significantly outperformed by any other for that language
pair are highlighted in bold. For each language pair, we specify the number of examples.
31


Published as a conference paper at ICLR 2020
Setting
Metric
cs-en
de-en
et-en
fi-en
ru-en
tr-en
zh-en
5
16
14
9
8
5
14
Unsupervised
BLEU
0.970
0.971
0.986
0.973
0.979
0.657
0.978
CDER
0.972
0.980
0.990
0.984
0.980
0.664
0.982
CHARACTER
0.970
0.993
0.979
0.989
0.991
0.782
0.950
ITER
0.975
0.990
0.975
0.996
0.937
0.861
0.980
METEOR++
0.945
0.991
0.978
0.971
0.995
0.864
0.962
NIST
0.954
0.984
0.983
0.975
0.973
0.970
0.968
PER
0.970
0.985
0.983
0.993
0.967
0.159
0.931
TER
0.950
0.970
0.990
0.968
0.970
0.533
0.975
UHH_TSKM
0.952
0.980
0.989
0.982
0.980
0.547
0.981
WER
0.951
0.961
0.991
0.961
0.968
0.041
0.975
YISI-0
0.956
0.994
0.975
0.978
0.988
0.954
0.957
YISI-1
0.950
0.992
0.979
0.973
0.991
0.958
0.951
YISI-1 SRL
0.965
0.995
0.981
0.977
0.992
0.869
0.962
Supervised
BEER
0.958
0.994
0.985
0.991
0.982
0.870
0.976
BLEND
0.973
0.991
0.985
0.994
0.993
0.801
0.976
RUSE
0.981
0.997
0.990
0.991
0.988
0.853
0.981
Pre-Trained
PBERT–Base
0.965
0.995
0.986
0.973
0.976
0.941
0.974
RBERT–Base
0.994
0.991
0.979
0.992
0.991
0.067
0.988
FBERT–Base
0.982
0.994
0.983
0.986
0.985
0.949
0.984
PBERT–Base (idf)
0.961
0.993
0.987
0.988
0.976
0.984
0.973
RBERT–Base (idf)
0.996
0.994
0.977
0.995
0.995
0.874
0.983
FBERT–Base (idf)
0.981
0.995
0.984
0.995
0.988
0.994
0.981
PBERT–Base–MRPC
0.957
0.994
0.989
0.953
0.976
0.798
0.977
RBERT–Base–MRPC
0.992
0.994
0.983
0.988
0.993
0.707
0.990
FBERT–Base–MRPC
0.975
0.995
0.987
0.975
0.986
0.526
0.986
PBERT–Base–MRPC (idf)
0.957
0.997
0.989
0.967
0.975
0.894
0.980
RBERT–Base–MRPC (idf)
0.991
0.997
0.981
0.994
0.993
0.052
0.987
FBERT–Base–MRPC (idf)
0.975
0.998
0.987
0.985
0.987
0.784
0.987
PBERT–Large
0.978
0.992
0.987
0.971
0.977
0.920
0.978
RBERT–Large
0.997
0.990
0.985
0.990
0.992
0.098
0.990
FBERT–Large
0.989
0.992
0.987
0.983
0.985
0.784
0.986
PBERT–Large (idf)
0.977
0.992
0.988
0.986
0.976
0.980
0.977
RBERT–Large (idf)
0.998
0.993
0.983
0.996
0.995
0.809
0.986
FBERT–Large (idf)
0.989
0.993
0.986
0.993
0.987
0.976
0.984
PRoBERTa–Base
0.970
0.995
0.991
0.998
0.976
0.796
0.980
RRoBERTa–Base
0.996
0.996
0.982
0.998
0.994
0.477
0.991
FRoBERTa–Base
0.984
0.997
0.989
0.999
0.987
0.280
0.989
PRoBERTa–Base (idf)
0.966
0.993
0.991
0.994
0.977
0.880
0.984
RRoBERTa–Base (idf)
0.995
0.998
0.981
0.998
0.995
0.230
0.989
FRoBERTa–Base (idf)
0.981
0.998
0.989
0.997
0.988
0.741
0.990
PRoBERTa–Large
0.980
0.998
0.990
0.995
0.982
0.791
0.981
RRoBERTa–Large
0.998
0.997
0.986
0.997
0.995
0.054
0.990
FRoBERTa–Large
0.990
0.999
0.990
0.998
0.990
0.499
0.988
PRoBERTa–Large (idf)
0.972
0.997
0.993
0.985
0.982
0.920
0.983
RRoBERTa–Large (idf)
0.996
0.997
0.984
0.997
0.995
0.578
0.989
FRoBERTa–Large (idf)
0.985
0.999
0.992
0.992
0.991
0.826
0.989
PRoBERTa–Large–MNLI
0.989
0.998
0.994
0.998
0.985
0.908
0.982
RRoBERTa–Large–MNLI
1.000
0.996
0.988
0.996
0.995
0.097
0.991
FRoBERTa–Large–MNLI
0.996
0.998
0.992
0.998
0.992
0.665
0.989
PRoBERTa–Large–MNLI (idf)
0.986
0.998
0.994
0.993
0.986
0.989
0.985
RRoBERTa–Large–MNLI (idf)
0.999
0.997
0.986
0.997
0.993
0.633
0.990
FRoBERTa–Large–MNLI (idf)
0.995
0.998
0.991
0.996
0.993
0.963
0.990
PXLNET–Base
0.970
0.996
0.986
0.990
0.979
0.739
0.982
RXLNET–Base
0.994
0.997
0.979
0.995
0.994
0.795
0.990
FXLNET–Base
0.983
0.997
0.983
0.993
0.987
0.505
0.988
PXLNET–Base (idf)
0.968
0.998
0.986
0.990
0.978
0.923
0.982
RXLNET–Base (idf)
0.993
0.998
0.978
0.996
0.994
0.439
0.988
FXLNET–Base (idf)
0.981
0.999
0.984
0.995
0.989
0.722
0.988
PXLNET–Large
0.969
0.998
0.986
0.995
0.979
0.880
0.981
RXLNET–Large
0.995
0.997
0.977
0.997
0.995
0.430
0.988
FXLNET–Large
0.983
0.998
0.983
0.997
0.988
0.713
0.988
PXLNET–Large (idf)
0.963
0.996
0.986
0.995
0.978
0.939
0.979
RXLNET–Large (idf)
0.992
0.997
0.975
0.993
0.996
0.531
0.982
FXLNET–Large (idf)
0.978
0.997
0.983
0.996
0.990
0.886
0.984
PXLM–En
0.965
0.996
0.990
0.978
0.980
0.946
0.981
RXLM–En
0.990
0.995
0.984
0.996
0.996
0.286
0.987
FXLM–En
0.978
0.997
0.988
0.990
0.989
0.576
0.987
PXLM–En (idf)
0.960
0.996
0.990
0.987
0.980
0.989
0.981
RXLM–En (idf)
0.991
0.997
0.983
0.996
0.998
0.612
0.985
FXLM–En (idf)
0.976
0.998
0.988
0.994
0.992
0.943
0.985
Table 19: Absolute Pearson correlations with system-level human judgments on WMT18 to-English
translations. Correlations of metrics not significantly outperformed by any other for that language
pair are highlighted in bold. For each language pair, we specify the number of systems.
32


Published as a conference paper at ICLR 2020
Setting
Metric
en-cs
en-de
en-et
en-fi
en-ru
en-tr
en-zh
5
16
14
12
9
8
14
Unsupervised
BLEU
0.995
0.981
0.975
0.962
0.983
0.826
0.947
CDER
0.997
0.986
0.984
0.964
0.984
0.861
0.961
CHARACTER
0.993
0.989
0.956
0.974
0.983
0.833
0.983
ITER
0.915
0.984
0.981
0.973
0.975
0.865
–
METEOR++
–
–
–
–
–
–
–
NIST
0.999
0.986
0.983
0.949
0.990
0.902
0.950
PER
0.991
0.981
0.958
0.906
0.988
0.859
0.964
TER
0.997
0.988
0.981
0.942
0.987
0.867
0.963
UHH_TSKM
–
–
–
–
–
–
–
WER
0.997
0.986
0.981
0.945
0.985
0.853
0.957
YISI-0
0.973
0.985
0.968
0.944
0.990
0.990
0.957
YISI-1
0.987
0.985
0.979
0.940
0.992
0.976
0.963
YISI-1 SRL
–
0.990
–
–
–
–
0.952
Supervised
BEER
0.992
0.991
0.980
0.961
0.988
0.965
0.928
BLEND
–
–
–
–
0.988
–
–
RUSE
–
–
–
–
–
–
–
Pre-Trained
PBERT–Multi
0.994
0.988
0.981
0.957
0.990
0.935
0.954
RBERT–Multi
0.997
0.990
0.980
0.980
0.989
0.879
0.976
FBERT–Multi
0.997
0.989
0.982
0.972
0.990
0.908
0.967
PBERT–Multi (idf)
0.992
0.986
0.974
0.954
0.991
0.969
0.954
RBERT–Multi (idf)
0.997
0.993
0.982
0.982
0.992
0.901
0.984
FBERT–Multi (idf)
0.995
0.990
0.981
0.972
0.991
0.941
0.973
PXLM–100
0.984
0.992
0.993
0.972
0.993
0.962
0.965
RXLM–100
0.991
0.992
0.992
0.989
0.992
0.895
0.983
FXLM–100
0.988
0.993
0.993
0.986
0.993
0.935
0.976
PXLM–100 (idf)
0.982
0.992
0.994
0.975
0.993
0.968
0.964
RXLM–100 (idf)
0.993
0.993
0.991
0.989
0.993
0.911
0.986
FXLM–100 (idf)
0.989
0.993
0.994
0.985
0.993
0.945
0.979
Table 20: Absolute Pearson correlations with system-level human judgments on WMT18 from-
English translations. Correlations of metrics not significantly outperformed by any other for that
language pair are highlighted in bold. For each language pair, we specify the number of systems.
33


Published as a conference paper at ICLR 2020
Setting
Metric
cs-en
de-en
et-en
fi-en
ru-en
tr-en
zh-en
10K
10K
10K
10K
10K
10K
10K
Unsupervised
BLEU
0.956
0.969
0.981
0.962
0.972
0.586
0.968
CDER
0.964
0.980
0.988
0.976
0.974
0.577
0.973
CHARACTER
0.960
0.992
0.975
0.979
0.984
0.680
0.942
ITER
0.966
0.990
0.975
0.989
0.943
0.742
0.978
METEOR++
0.937
0.990
0.975
0.962
0.989
0.787
0.954
NIST
0.942
0.982
0.980
0.965
0.965
0.862
0.959
PER
0.937
0.982
0.978
0.983
0.955
0.043
0.923
TER
0.942
0.970
0.988
0.960
0.963
0.450
0.967
UHH_TSKM
0.943
0.979
0.987
0.974
0.973
0.443
0.972
WER
0.942
0.961
0.989
0.953
0.962
0.072
0.967
YISI-0
0.947
0.992
0.972
0.969
0.982
0.863
0.950
YISI-1
0.942
0.991
0.976
0.964
0.985
0.881
0.943
YISI-1 SRL
0.957
0.994
0.978
0.968
0.986
0.785
0.954
Supervised
BEER
0.950
0.993
0.983
0.982
0.976
0.723
0.968
BLEND
0.965
0.990
0.982
0.985
0.986
0.724
0.969
RUSE
0.974
0.996
0.988
0.983
0.982
0.780
0.973
Pre-Trained
PBERT–Base
0.954
0.992
0.984
0.980
0.970
0.917
0.965
RBERT–Base
0.988
0.994
0.974
0.987
0.988
0.801
0.975
FBERT–Base
0.973
0.994
0.981
0.987
0.982
0.924
0.973
PBERT–Base (idf)
0.957
0.994
0.983
0.966
0.970
0.875
0.966
RBERT–Base (idf)
0.986
0.990
0.976
0.984
0.984
0.019
0.980
FBERT–Base (idf)
0.974
0.993
0.980
0.978
0.978
0.853
0.976
PBERT–Base–MRPC
0.949
0.995
0.986
0.960
0.969
0.832
0.972
RBERT–Base–MRPC
0.983
0.997
0.979
0.986
0.986
0.099
0.980
FBERT–Base–MRPC
0.967
0.997
0.984
0.978
0.981
0.722
0.979
PBERT–Base–MRPC (idf)
0.949
0.994
0.986
0.946
0.969
0.743
0.969
RBERT–Base–MRPC (idf)
0.984
0.994
0.980
0.980
0.986
0.541
0.982
FBERT–Base–MRPC (idf)
0.967
0.995
0.984
0.968
0.979
0.464
0.978
PBERT–Large
0.969
0.991
0.985
0.979
0.970
0.915
0.969
RBERT–Large
0.990
0.993
0.980
0.988
0.988
0.745
0.978
FBERT–Large
0.982
0.993
0.984
0.986
0.981
0.909
0.976
PBERT–Large (idf)
0.970
0.991
0.984
0.963
0.971
0.858
0.970
RBERT–Large (idf)
0.989
0.990
0.982
0.982
0.985
0.047
0.982
FBERT–Large (idf)
0.981
0.991
0.984
0.976
0.978
0.722
0.978
PRoBERTa–Base
0.959
0.992
0.988
0.986
0.971
0.809
0.976
RRoBERTa–Base
0.987
0.997
0.978
0.989
0.988
0.238
0.981
FRoBERTa–Base
0.973
0.997
0.987
0.989
0.982
0.674
0.982
PRoBERTa–Base (idf)
0.963
0.994
0.988
0.989
0.970
0.711
0.972
RRoBERTa–Base (idf)
0.988
0.996
0.979
0.989
0.987
0.353
0.983
FRoBERTa–Base (idf)
0.976
0.997
0.986
0.990
0.980
0.277
0.980
PRoBERTa–Large
0.965
0.995
0.990
0.976
0.976
0.846
0.975
RRoBERTa–Large
0.989
0.997
0.982
0.989
0.988
0.540
0.981
FRoBERTa–Large
0.978
0.998
0.989
0.983
0.985
0.760
0.981
PRoBERTa–Large (idf)
0.972
0.997
0.988
0.986
0.976
0.686
0.973
RRoBERTa–Large (idf)
0.990
0.996
0.983
0.989
0.989
0.096
0.982
FRoBERTa–Large (idf)
0.982
0.998
0.988
0.989
0.983
0.453
0.980
PRoBERTa–Large–MNLI
0.978
0.997
0.991
0.984
0.980
0.914
0.977
RRoBERTa–Large–MNLI
0.991
0.996
0.984
0.989
0.987
0.566
0.982
FRoBERTa–Large–MNLI
0.987
0.998
0.989
0.988
0.986
0.873
0.982
PRoBERTa–Large–MNLI (idf)
0.982
0.998
0.992
0.990
0.978
0.822
0.974
RRoBERTa–Large–MNLI (idf)
0.992
0.996
0.985
0.988
0.988
0.022
0.983
FRoBERTa–Large–MNLI (idf)
0.989
0.998
0.990
0.990
0.985
0.583
0.980
PXLNET–Base
0.960
0.997
0.984
0.982
0.972
0.849
0.974
RXLNET–Base
0.985
0.997
0.975
0.988
0.988
0.303
0.980
FXLNET–Base
0.974
0.998
0.981
0.986
0.982
0.628
0.980
PXLNET–Base (idf)
0.962
0.995
0.983
0.982
0.972
0.657
0.974
RXLNET–Base (idf)
0.986
0.996
0.976
0.987
0.987
0.666
0.982
FXLNET–Base (idf)
0.975
0.996
0.980
0.985
0.981
0.259
0.980
PXLNET–Large
0.955
0.995
0.983
0.986
0.972
0.875
0.970
RXLNET–Large
0.984
0.996
0.972
0.984
0.989
0.491
0.975
FXLNET–Large
0.971
0.996
0.980
0.987
0.984
0.821
0.976
PXLNET–Large (idf)
0.961
0.997
0.983
0.987
0.973
0.816
0.973
RXLNET–Large (idf)
0.987
0.996
0.975
0.989
0.988
0.320
0.981
FXLNET–Large (idf)
0.976
0.997
0.980
0.989
0.982
0.623
0.980
PXLM–En
0.953
0.995
0.988
0.979
0.974
0.918
0.972
RXLM–En
0.983
0.996
0.980
0.988
0.991
0.561
0.977
FXLM–En
0.969
0.997
0.986
0.986
0.985
0.869
0.977
PXLM–En (idf)
0.957
0.996
0.987
0.970
0.974
0.862
0.973
RXLM–En (idf)
0.982
0.995
0.981
0.988
0.989
0.213
0.980
FXLM–En (idf)
0.970
0.996
0.985
0.982
0.982
0.519
0.978
Table 21: Absolute Pearson correlations with human judgments on WMT18 to-English language
pairs for 10K hybrid systems. Correlations of metrics not significantly outperformed by any other
for that language pair are highlighted in bold. For each language pair, we specify the number of
systems.
34


Published as a conference paper at ICLR 2020
Setting
Metric
en-cs
en-de
en-et
en-fi
en-ru
en-tr
en-zh
10K
10K
10K
10K
10K
10K
10K
Unsupervised
BLEU
0.993
0.977
0.971
0.958
0.977
0.796
0.941
CDER
0.995
0.984
0.981
0.961
0.982
0.832
0.956
CHARACTER
0.990
0.986
0.950
0.963
0.981
0.775
0.978
ITER
0.865
0.978
0.982
0.966
0.965
0.872
–
METEOR++
–
–
–
–
–
–
–
NIST
0.997
0.984
0.980
0.944
0.988
0.870
0.944
PER
0.987
0.979
0.954
0.904
0.986
0.829
0.950
TER
0.995
0.986
0.977
0.939
0.985
0.837
0.959
UHH_TSKM
–
–
–
–
–
–
–
WER
0.994
0.984
0.977
0.942
0.983
0.824
0.954
YISI-0
0.971
0.983
0.965
0.942
0.988
0.953
0.951
YISI-1
0.985
0.983
0.976
0.938
0.989
0.942
0.957
YISI-1 SRL
–
0.988
–
–
–
–
0.948
Supervised
BEER
0.990
0.989
0.978
0.959
0.986
0.933
0.925
BLEND
–
–
–
–
0.986
–
–
RUSE
–
–
–
–
–
–
–
Pre-Trained
PBERT–Multi
0.989
0.983
0.970
0.951
0.988
0.936
0.950
RBERT–Multi
0.995
0.991
0.979
0.977
0.989
0.872
0.980
FBERT–Multi
0.993
0.988
0.978
0.969
0.989
0.910
0.969
PBERT–Multi (idf)
0.992
0.986
0.978
0.954
0.988
0.903
0.950
RBERT–Multi (idf)
0.995
0.988
0.977
0.976
0.987
0.850
0.972
FBERT–Multi (idf)
0.995
0.988
0.979
0.969
0.987
0.877
0.963
PXLM–100
0.980
0.990
0.991
0.972
0.991
0.936
0.959
RXLM–100
0.991
0.990
0.989
0.985
0.991
0.882
0.981
FXLM–100
0.987
0.990
0.991
0.981
0.991
0.915
0.974
PXLM–100 (idf)
0.982
0.990
0.990
0.968
0.991
0.931
0.960
RXLM–100 (idf)
0.989
0.990
0.990
0.985
0.990
0.867
0.978
FXLM–100 (idf)
0.986
0.991
0.991
0.982
0.991
0.905
0.972
Table 22: Absolute Pearson correlations with human judgments on WMT18 from-English language
pairs for 10K hybrid systems. Correlations of metrics not significantly outperformed by any other
for that language pair are highlighted in bold. For each language pair, we specify the number of
systems.
35


Published as a conference paper at ICLR 2020
Setting
Metric
cs-en
de-en
et-en
fi-en
ru-en
tr-en
zh-en
Unsupervised
BLEU
0.135
0.804
0.757
0.460
0.230
0.096
0.661
CDER
0.162
0.795
0.764
0.493
0.234
0.087
0.660
CHARACTER
0.146
0.737
0.696
0.496
0.201
0.082
0.584
ITER
0.152
0.814
0.746
0.474
0.234
0.100
0.673
METEOR++
0.172
0.804
0.646
0.456
0.253
0.052
0.597
NIST
0.136
0.802
0.739
0.469
0.228
0.135
0.665
PER
0.121
0.764
0.602
0.455
0.218
0.000
0.602
TER
0.139
0.789
0.768
0.470
0.232
0.001
0.652
UHH_TSKM
0.191
0.803
0.768
0.469
0.240
0.002
0.642
WER
0.149
0.776
0.760
0.471
0.227
0.000
0.654
YISI-0
0.148
0.780
0.703
0.483
0.229
0.106
0.629
YISI-1
0.157
0.808
0.752
0.466
0.250
0.110
0.613
YISI-1 SRL
0.159
0.814
0.763
0.484
0.243
0.008
0.620
Supervised
BEER
0.165
0.811
0.765
0.485
0.237
0.030
0.675
BLEND
0.184
0.820
0.779
0.484
0.254
0.003
0.611
RUSE
0.213
0.823
0.788
0.487
0.250
0.109
0.672
Pre-Trained
PBERT–Base
0.190
0.815
0.778
0.468
0.261
0.130
0.655
RBERT–Base
0.189
0.813
0.775
0.481
0.266
0.014
0.663
FBERT–Base
0.194
0.819
0.778
0.474
0.265
0.144
0.670
PBERT–Base (idf)
0.189
0.817
0.775
0.477
0.255
0.131
0.650
RBERT–Base (idf)
0.192
0.808
0.771
0.484
0.248
0.005
0.674
FBERT–Base (idf)
0.193
0.817
0.774
0.483
0.262
0.081
0.669
PBERT–Base–MRPC
0.190
0.701
0.766
0.487
0.254
0.126
0.653
RBERT–Base–MRPC
0.199
0.826
0.765
0.493
0.258
0.000
0.671
FBERT–Base–MRPC
0.197
0.824
0.767
0.491
0.260
0.147
0.668
PBERT–Base–MRPC (idf)
0.186
0.806
0.765
0.492
0.247
0.125
0.661
RBERT–Base–MRPC (idf)
0.200
0.823
0.760
0.495
0.258
0.000
0.680
FBERT–Base–MRPC (idf)
0.196
0.821
0.763
0.497
0.254
0.031
0.676
PBERT–Large
0.200
0.815
0.778
0.474
0.261
0.137
0.661
RBERT–Large
0.194
0.809
0.779
0.493
0.270
0.006
0.672
FBERT–Large
0.199
0.810
0.782
0.484
0.266
0.142
0.672
PBERT–Large (idf)
0.200
0.813
0.772
0.485
0.256
0.136
0.657
RBERT–Large (idf)
0.197
0.806
0.769
0.495
0.262
0.005
0.675
FBERT–Large (idf)
0.199
0.811
0.772
0.494
0.262
0.006
0.673
PRoBERTa–Base
0.173
0.675
0.757
0.502
0.258
0.126
0.654
RRoBERTa–Base
0.165
0.816
0.764
0.483
0.266
0.000
0.674
FRoBERTa–Base
0.173
0.820
0.764
0.498
0.262
0.090
0.669
PRoBERTa–Base (idf)
0.172
0.691
0.755
0.503
0.252
0.123
0.661
RRoBERTa–Base (idf)
0.172
0.809
0.758
0.490
0.268
0.000
0.678
FRoBERTa–Base (idf)
0.178
0.820
0.758
0.501
0.260
0.001
0.674
PRoBERTa–Large
0.174
0.704
0.765
0.497
0.255
0.140
0.663
RRoBERTa–Large
0.163
0.805
0.770
0.491
0.263
0.005
0.679
FRoBERTa–Large
0.175
0.825
0.770
0.499
0.262
0.143
0.675
PRoBERTa–Large (idf)
0.181
0.821
0.758
0.500
0.256
0.089
0.669
RRoBERTa–Large (idf)
0.165
0.787
0.763
0.495
0.270
0.000
0.684
FRoBERTa–Large (idf)
0.179
0.824
0.761
0.502
0.265
0.004
0.679
PRoBERTa–Large–MNLI
0.185
0.828
0.780
0.504
0.263
0.133
0.654
RRoBERTa–Large–MNLI
0.179
0.779
0.775
0.494
0.266
0.004
0.670
FRoBERTa–Large–MNLI
0.186
0.827
0.778
0.502
0.267
0.113
0.669
PRoBERTa–Large–MNLI (idf)
0.190
0.820
0.771
0.504
0.261
0.102
0.661
RRoBERTa–Large–MNLI (idf)
0.181
0.769
0.766
0.494
0.266
0.004
0.674
FRoBERTa–Large–MNLI (idf)
0.188
0.822
0.768
0.501
0.265
0.004
0.671
PXLNET–Base
0.186
0.771
0.762
0.496
0.247
0.153
0.658
RXLNET–Base
0.182
0.823
0.764
0.496
0.256
0.000
0.671
FXLNET–Base
0.186
0.824
0.765
0.499
0.253
0.049
0.673
PXLNET–Base (idf)
0.178
0.819
0.756
0.506
0.241
0.130
0.656
RXLNET–Base (idf)
0.183
0.817
0.754
0.501
0.256
0.000
0.673
FXLNET–Base (idf)
0.182
0.821
0.755
0.505
0.250
0.000
0.670
PXLNET–Large
0.195
0.721
0.767
0.493
0.152
0.144
0.661
RXLNET–Large
0.192
0.821
0.766
0.494
0.260
0.001
0.659
FXLNET–Large
0.196
0.824
0.773
0.496
0.261
0.155
0.675
PXLNET–Large (idf)
0.191
0.811
0.765
0.500
0.167
0.144
0.657
RXLNET–Large (idf)
0.196
0.815
0.762
0.495
0.259
0.000
0.673
FXLNET–Large (idf)
0.195
0.822
0.764
0.499
0.256
0.046
0.674
PXLM–En
0.192
0.796
0.779
0.486
0.255
0.131
0.665
RXLM–En
0.202
0.818
0.772
0.495
0.261
0.005
0.662
FXLM–En
0.199
0.827
0.778
0.491
0.262
0.086
0.674
PXLM–En (idf)
0.189
0.818
0.770
0.485
0.259
0.116
0.662
RXLM–En (idf)
0.202
0.812
0.761
0.490
0.250
0.003
0.668
FXLM–En (idf)
0.196
0.821
0.766
0.490
0.263
0.003
0.672
Table 23: Model selection accuracies (Hits@1) on to-English WMT18 hybrid systems. We report
the average of 100K samples and the 0.95 confidence intervals are below 10−3. We bold the highest
numbers for each language pair and direction.
36


Published as a conference paper at ICLR 2020
Setting
Metric
cs-en
de-en
et-en
fi-en
ru-en
tr-en
zh-en
Unsupervised
BLEU
0.338
0.894
0.866
0.666
0.447
0.265
0.799
CDER
0.362
0.890
0.870
0.689
0.451
0.256
0.799
CHARACTER
0.349
0.854
0.814
0.690
0.429
0.254
0.739
ITER
0.356
0.901
0.856
0.676
0.454
0.278
0.811
METEOR++
0.369
0.895
0.798
0.662
0.470
0.174
0.757
NIST
0.338
0.894
0.857
0.672
0.446
0.323
0.803
PER
0.325
0.866
0.771
0.663
0.435
0.021
0.754
TER
0.342
0.885
0.873
0.673
0.447
0.063
0.792
UHH_TSKM
0.387
0.894
0.873
0.671
0.460
0.063
0.788
WER
0.353
0.876
0.868
0.674
0.443
0.034
0.790
YISI-0
0.344
0.881
0.834
0.681
0.452
0.275
0.776
YISI-1
0.352
0.896
0.864
0.671
0.470
0.285
0.765
YISI-1 SRL
0.351
0.901
0.871
0.682
0.464
0.086
0.770
Supervised
BEER
0.364
0.899
0.871
0.684
0.460
0.125
0.811
BLEND
0.382
0.904
0.880
0.681
0.473
0.077
0.767
RUSE
0.417
0.906
0.885
0.686
0.468
0.273
0.809
Pre-Trained
PBERT–Base
0.386
0.901
0.880
0.674
0.481
0.318
0.793
RBERT–Base
0.383
0.899
0.877
0.683
0.486
0.100
0.804
FBERT–Base
0.388
0.903
0.879
0.678
0.484
0.331
0.808
PBERT–Base (idf)
0.390
0.902
0.877
0.681
0.475
0.318
0.786
RBERT–Base (idf)
0.390
0.896
0.874
0.686
0.475
0.077
0.811
FBERT–Base (idf)
0.393
0.902
0.876
0.685
0.483
0.225
0.806
PBERT–Base–MRPC
0.392
0.832
0.872
0.686
0.475
0.319
0.791
RBERT–Base–MRPC
0.397
0.908
0.870
0.691
0.478
0.025
0.811
FBERT–Base–MRPC
0.398
0.907
0.872
0.690
0.481
0.335
0.806
PBERT–Base–MRPC (idf)
0.392
0.896
0.870
0.689
0.467
0.316
0.797
RBERT–Base–MRPC (idf)
0.400
0.906
0.867
0.691
0.479
0.018
0.817
FBERT–Base–MRPC (idf)
0.400
0.905
0.869
0.693
0.475
0.097
0.812
PBERT–Large
0.398
0.901
0.880
0.678
0.481
0.327
0.799
RBERT–Large
0.391
0.897
0.879
0.690
0.490
0.085
0.810
FBERT–Large
0.397
0.898
0.882
0.684
0.486
0.328
0.810
PBERT–Large (idf)
0.398
0.900
0.875
0.685
0.475
0.323
0.794
RBERT–Large (idf)
0.395
0.895
0.873
0.692
0.488
0.080
0.813
FBERT–Large (idf)
0.398
0.899
0.875
0.691
0.482
0.086
0.810
PRoBERTa–Base
0.372
0.814
0.866
0.697
0.475
0.313
0.795
RRoBERTa–Base
0.366
0.902
0.870
0.683
0.483
0.026
0.813
FRoBERTa–Base
0.374
0.904
0.870
0.694
0.480
0.224
0.808
PRoBERTa–Base (idf)
0.373
0.825
0.865
0.697
0.470
0.303
0.802
RRoBERTa–Base (idf)
0.374
0.898
0.866
0.688
0.486
0.028
0.816
FRoBERTa–Base (idf)
0.380
0.904
0.866
0.696
0.479
0.037
0.812
PRoBERTa–Large
0.375
0.833
0.871
0.693
0.474
0.327
0.800
RRoBERTa–Large
0.366
0.895
0.874
0.689
0.480
0.039
0.816
FRoBERTa–Large
0.378
0.907
0.874
0.694
0.480
0.324
0.811
PRoBERTa–Large (idf)
0.384
0.905
0.866
0.694
0.475
0.220
0.806
RRoBERTa–Large (idf)
0.368
0.885
0.869
0.692
0.487
0.030
0.819
FRoBERTa–Large (idf)
0.382
0.907
0.868
0.696
0.484
0.048
0.815
PRoBERTa–Large–MNLI
0.383
0.909
0.880
0.698
0.480
0.323
0.795
RRoBERTa–Large–MNLI
0.378
0.880
0.877
0.692
0.481
0.078
0.811
FRoBERTa–Large–MNLI
0.385
0.909
0.879
0.697
0.484
0.286
0.809
PRoBERTa–Large–MNLI (idf)
0.389
0.905
0.874
0.698
0.478
0.268
0.803
RRoBERTa–Large–MNLI (idf)
0.380
0.874
0.870
0.691
0.483
0.079
0.814
FRoBERTa–Large–MNLI (idf)
0.387
0.906
0.872
0.696
0.482
0.082
0.811
PXLNET–Base
0.385
0.875
0.869
0.692
0.469
0.342
0.796
RXLNET–Base
0.381
0.907
0.869
0.693
0.477
0.026
0.809
FXLNET–Base
0.385
0.907
0.871
0.694
0.476
0.128
0.810
PXLNET–Base (idf)
0.381
0.904
0.864
0.699
0.464
0.289
0.794
RXLNET–Base (idf)
0.384
0.903
0.863
0.696
0.479
0.013
0.812
FXLNET–Base (idf)
0.384
0.905
0.864
0.699
0.472
0.032
0.809
PXLNET–Large
0.392
0.844
0.873
0.689
0.367
0.338
0.799
RXLNET–Large
0.389
0.905
0.871
0.690
0.482
0.031
0.800
FXLNET–Large
0.393
0.907
0.876
0.691
0.483
0.348
0.812
PXLNET–Large (idf)
0.393
0.899
0.870
0.694
0.387
0.333
0.794
RXLNET–Large (idf)
0.395
0.901
0.868
0.690
0.483
0.023
0.810
FXLNET–Large (idf)
0.396
0.906
0.870
0.693
0.478
0.128
0.811
PXLM–En
0.394
0.891
0.880
0.685
0.476
0.322
0.802
RXLM–En
0.401
0.903
0.875
0.692
0.483
0.082
0.803
FXLM–En
0.400
0.909
0.878
0.689
0.483
0.234
0.811
PXLM–En (idf)
0.391
0.903
0.874
0.684
0.480
0.293
0.797
RXLM–En (idf)
0.402
0.900
0.868
0.688
0.477
0.068
0.806
FXLM–En (idf)
0.398
0.905
0.871
0.688
0.487
0.079
0.809
Table 24: Mean Reciprocal Rank (MRR) of the top metric-rated system on to-English WMT18
hybrid systems. We report the average of 100K samples and the 0.95 confidence intervals are below
10−3. We bold the highest numbers for each language pair and direction.
37


Published as a conference paper at ICLR 2020
Setting
Metric
cs-en
de-en
et-en
fi-en
ru-en
tr-en
zh-en
Unsupervised
BLEU
3.85
0.45
1.01
2.17
2.34
4.48
3.19
CDER
3.88
0.43
0.87
1.33
2.30
4.58
3.43
CHARACTER
3.77
0.49
0.94
2.07
2.25
4.07
3.37
ITER
3.55
0.46
1.25
1.43
4.65
3.11
2.92
METEOR++
3.70
0.41
0.69
1.13
2.28
1.40
3.50
NIST
3.93
0.49
1.10
1.19
2.36
1.42
3.92
PER
2.02
0.46
1.71
1.49
2.25
4.22
3.20
TER
3.86
0.43
1.14
1.14
4.34
5.18
3.82
UHH_TSKM
3.98
0.40
1.27
1.10
2.23
4.26
3.47
WER
3.85
0.44
1.48
1.18
4.87
5.96
3.72
YISI-0
3.81
0.48
0.72
1.20
1.75
1.40
3.44
YISI-1
3.88
0.44
0.65
1.13
2.17
1.32
3.40
YISI-1 SRL
3.67
0.41
0.64
1.20
2.15
1.31
3.55
Supervised
BEER
3.82
0.41
0.79
1.08
1.92
1.96
3.43
BLEND
3.77
0.41
0.66
1.09
2.21
1.28
3.46
RUSE
3.13
0.32
0.64
1.03
1.51
1.94
3.15
Pre-Trained
PBERT–Base
3.97
0.36
0.72
1.16
2.20
1.25
3.26
RBERT–Base
1.51
0.43
0.60
1.65
1.33
1.34
3.50
FBERT–Base
3.70
0.36
0.59
1.08
1.92
1.27
3.38
PBERT–Base (idf)
3.94
0.36
0.64
1.18
2.06
2.55
3.54
RBERT–Base (idf)
1.54
0.43
0.63
1.87
1.12
5.96
3.38
FBERT–Base (idf)
2.75
0.39
0.60
1.10
1.38
1.26
3.51
PBERT–Base–MRPC
4.02
0.35
0.74
1.15
1.09
3.33
3.06
RBERT–Base–MRPC
2.66
0.43
0.62
1.75
1.10
5.64
3.34
FBERT–Base–MRPC
3.89
0.36
0.60
1.09
1.08
3.82
3.23
PBERT–Base–MRPC (idf)
4.02
0.35
0.67
1.18
1.48
3.30
3.49
RBERT–Base–MRPC (idf)
1.63
0.43
0.65
1.93
1.13
7.26
3.13
FBERT–Base–MRPC (idf)
3.86
0.38
0.61
1.11
1.14
4.24
3.28
PBERT–Large
3.82
0.34
0.66
1.12
2.10
1.31
3.60
RBERT–Large
1.49
0.40
0.59
1.56
1.17
1.35
3.61
FBERT–Large
1.71
0.35
0.58
1.08
1.65
1.29
3.60
PBERT–Large (idf)
3.74
0.35
0.65
1.12
1.90
1.98
3.77
RBERT–Large (idf)
1.51
0.42
0.62
1.86
1.10
5.84
3.21
FBERT–Large (idf)
1.49
0.38
0.60
1.17
1.24
1.96
3.53
PRoBERTa–Base
3.89
0.37
0.75
1.18
1.07
3.45
2.62
RRoBERTa–Base
1.92
0.39
0.64
1.57
1.11
5.75
3.13
FRoBERTa–Base
3.56
0.37
0.59
1.10
1.08
3.79
2.90
PRoBERTa–Base (idf)
3.89
0.38
0.67
1.20
1.30
3.27
3.47
RRoBERTa–Base (idf)
1.61
0.42
0.67
1.65
1.14
6.55
2.95
FRoBERTa–Base (idf)
3.18
0.38
0.60
1.11
1.13
6.54
3.11
PRoBERTa–Large
3.64
0.36
0.71
1.10
1.03
2.69
2.57
RRoBERTa–Large
1.60
0.37
0.64
1.51
1.09
3.91
3.27
FRoBERTa–Large
2.38
0.35
0.58
1.06
1.05
3.57
2.95
PRoBERTa–Large (idf)
2.70
0.36
0.69
1.13
1.08
3.18
2.89
RRoBERTa–Large (idf)
1.55
0.39
0.66
1.59
1.10
6.66
3.18
FRoBERTa–Large (idf)
1.68
0.37
0.59
1.08
1.08
5.58
2.91
PRoBERTa–Large–MNLI
2.14
0.35
0.61
1.07
1.09
1.21
3.35
RRoBERTa–Large–MNLI
1.45
0.37
0.64
1.49
1.10
4.42
3.55
FRoBERTa–Large–MNLI
1.42
0.35
0.59
1.07
1.07
1.27
3.41
PRoBERTa–Large–MNLI (idf)
1.55
0.35
0.60
1.08
1.12
1.54
3.87
RRoBERTa–Large–MNLI (idf)
1.45
0.39
0.64
1.65
1.09
5.89
3.32
FRoBERTa–Large–MNLI (idf)
1.42
0.36
0.60
1.10
1.08
3.80
3.45
PXLNET–Base
3.90
0.37
0.68
1.07
1.16
2.47
2.91
RXLNET–Base
1.71
0.45
0.72
1.58
1.07
6.29
3.36
FXLNET–Base
3.78
0.39
0.62
1.05
1.07
3.60
3.20
PXLNET–Base (idf)
3.90
0.46
0.65
1.08
2.93
3.30
3.39
RXLNET–Base (idf)
1.51
0.45
0.82
1.78
1.12
10.77
3.13
FXLNET–Base (idf)
3.67
0.42
0.66
1.11
1.22
7.13
3.23
PXLNET–Large
3.94
0.37
0.71
1.10
21.10
1.85
2.90
RXLNET–Large
2.23
0.41
0.69
1.34
1.07
4.46
3.40
FXLNET–Large
3.84
0.36
0.60
1.03
1.07
3.38
3.22
PXLNET–Large (idf)
3.92
0.41
0.64
1.12
21.10
3.24
3.37
RXLNET–Large (idf)
1.60
0.43
0.78
1.70
1.09
6.13
3.20
FXLNET–Large (idf)
3.80
0.38
0.63
1.06
1.09
3.72
3.25
PXLM–En
3.88
0.33
0.75
1.16
2.16
1.28
3.29
RXLM–En
1.98
0.41
0.60
1.41
1.21
3.30
3.47
FXLM–En
3.78
0.36
0.61
1.09
1.71
1.30
3.40
PXLM–En (idf)
3.84
0.36
0.69
1.17
1.86
1.33
3.47
RXLM–En (idf)
1.70
0.42
0.63
1.55
1.11
5.87
3.36
FXLM–En (idf)
3.72
0.40
0.62
1.14
1.32
4.15
3.43
Table 25: Absolute Difference (×100) of the top metric-rated and the top human-rated system on to-
English WMT18 hybrid systems. Smaller difference signify higher agreement with human scores.
We report the average of 100K samples and the 0.95 confidence intervals are below 10−3. We bold
the lowest numbers for each language pair and direction.
38


Published as a conference paper at ICLR 2020
Setting
Metric
en-cs
en-de
en-et
en-fi
en-ru
en-tr
en-zh
Unsupervised
BLEU
0.151
0.611
0.617
0.087
0.519
0.029
0.515
CDER
0.163
0.663
0.731
0.081
0.541
0.032
0.552
CHARACTER
0.135
0.737
0.639
0.492
0.543
0.027
0.667
ITER
0.000
0.691
0.734
0.112
0.534
0.031
–
METEOR++
–
–
–
–
–
–
–
NIST
0.182
0.662
0.549
0.083
0.537
0.033
0.553
PER
0.179
0.555
0.454
0.062
0.535
0.032
0.539
TER
0.175
0.657
0.550
0.065
0.545
0.029
0.551
UHH_TSKM
–
–
–
–
–
–
–
WER
0.155
0.643
0.552
0.067
0.538
0.029
0.546
YISI-0
0.154
0.674
0.622
0.356
0.523
0.383
0.600
YISI-1
0.178
0.670
0.674
0.230
0.548
0.396
0.595
YISI-1 SRL
–
0.708
–
–
–
–
0.537
Supervised
BEER
0.174
0.670
0.662
0.113
0.555
0.296
0.531
BLEND
–
–
–
–
0.559
–
–
RUSE
–
–
–
–
–
–
–
Pre-Trained
PBERT–Multi
0.181
0.665
0.771
0.077
0.550
0.373
0.550
RBERT–Multi
0.184
0.728
0.722
0.146
0.544
0.031
0.657
FBERT–Multi
0.185
0.703
0.764
0.081
0.548
0.032
0.629
PBERT–Multi (idf)
0.175
0.713
0.769
0.080
0.542
0.031
0.549
RBERT–Multi (idf)
0.177
0.725
0.752
0.178
0.538
0.031
0.628
FBERT–Multi (idf)
0.178
0.721
0.766
0.081
0.543
0.030
0.594
PXLM–100
0.175
0.669
0.748
0.079
0.550
0.314
0.582
RXLM–100
0.195
0.671
0.770
0.222
0.555
0.034
0.658
FXLM–100
0.187
0.670
0.775
0.099
0.552
0.034
0.615
PXLM–100 (idf)
0.163
0.664
0.750
0.091
0.550
0.288
0.578
RXLM–100 (idf)
0.191
0.681
0.770
0.231
0.548
0.033
0.645
FXLM–100 (idf)
0.180
0.672
0.774
0.127
0.550
0.033
0.616
Table 26: Model selection accuracies (Hits@1) on to-English WMT18 hybrid systems. We report
the average of 100K samples and the 0.95 confidence intervals are below 10−3. We bold the highest
numbers for each language pair and direction.
39


Published as a conference paper at ICLR 2020
Setting
Metric
en-cs
en-de
en-et
en-fi
en-ru
en-tr
en-zh
Unsupervised
BLEU
0.363
0.764
0.766
0.323
0.714
0.205
0.666
CDER
0.371
0.803
0.851
0.319
0.729
0.210
0.700
CHARACTER
0.346
0.853
0.781
0.667
0.732
0.205
0.809
ITER
0.044
0.825
0.853
0.365
0.717
0.210
–
METEOR++
–
–
–
–
–
–
–
NIST
0.393
0.803
0.710
0.326
0.726
0.211
0.698
PER
0.387
0.719
0.624
0.301
0.725
0.211
0.678
TER
0.384
0.798
0.708
0.305
0.733
0.209
0.695
UHH_TSKM
–
–
–
–
–
–
–
WER
0.367
0.787
0.710
0.308
0.728
0.209
0.696
YISI-0
0.370
0.811
0.775
0.553
0.715
0.602
0.753
YISI-1
0.390
0.808
0.811
0.439
0.735
0.612
0.750
YISI-1 SRL
–
0.835
–
–
–
–
0.691
Supervised
BEER
0.388
0.808
0.804
0.353
0.739
0.507
0.683
BLEND
–
–
–
–
0.742
–
–
RUSE
–
–
–
–
–
–
–
Pre-Trained
PBERT–Multi
0.395
0.805
0.876
0.314
0.736
0.586
0.694
RBERT–Multi
0.401
0.849
0.844
0.368
0.732
0.212
0.802
FBERT–Multi
0.400
0.832
0.872
0.317
0.735
0.214
0.775
PBERT–Multi (idf)
0.390
0.839
0.875
0.320
0.730
0.213
0.691
RBERT–Multi (idf)
0.395
0.847
0.864
0.398
0.727
0.212
0.776
FBERT–Multi (idf)
0.395
0.844
0.873
0.319
0.730
0.212
0.739
PXLM–100
0.391
0.808
0.862
0.316
0.735
0.522
0.733
RXLM–100
0.413
0.809
0.876
0.435
0.738
0.216
0.803
FXLM–100
0.404
0.809
0.878
0.333
0.737
0.216
0.767
PXLM–100 (idf)
0.377
0.805
0.863
0.326
0.735
0.497
0.729
RXLM–100 (idf)
0.409
0.816
0.876
0.444
0.733
0.214
0.793
FXLM–100 (idf)
0.396
0.810
0.878
0.355
0.735
0.214
0.767
Table 27: Mean Reciprocal Rank (MRR) of the top metric-rated system on to-English WMT18
hybrid systems. We report the average of 100K samples and the 0.95 confidence intervals are below
10−3. We bold the highest numbers for each language pair and direction.
40


Published as a conference paper at ICLR 2020
Setting
Metric
en-cs
en-de
en-et
en-fi
en-ru
en-tr
en-zh
Unsupervised
BLEU
1.26
6.36
2.59
0.92
0.76
9.40
3.01
CDER
1.25
6.70
1.90
1.41
0.87
9.37
1.75
CHARACTER
1.23
6.90
2.19
4.35
0.93
5.22
1.64
ITER
1.25
9.14
2.52
1.52
1.35
7.33
–
METEOR++
–
–
–
–
–
–
–
NIST
1.24
5.28
2.55
1.02
0.75
8.82
3.34
PER
1.25
6.62
4.92
7.43
0.68
9.76
2.31
TER
1.21
6.02
4.34
2.17
0.73
8.80
1.43
UHH_TSKM
–
–
–
–
–
–
–
WER
1.22
6.15
4.19
2.43
0.72
9.28
1.49
YISI-0
1.25
6.62
1.53
1.46
0.75
3.47
2.87
YISI-1
1.22
6.27
1.21
1.13
0.71
3.51
3.33
YISI-1 SRL
–
6.57
–
–
–
–
3.71
Supervised
BEER
1.21
5.96
1.84
0.77
0.74
3.36
1.96
BLEND
–
–
–
–
0.71
–
–
RUSE
–
–
–
–
–
–
–
Pre-Trained
PBERT–Multi
1.17
3.27
1.38
1.24
0.75
4.14
2.08
RBERT–Multi
1.16
6.68
0.77
0.94
0.68
3.22
1.31
FBERT–Multi
1.15
5.17
0.90
0.98
0.71
3.26
1.62
PBERT–Multi (idf)
1.14
3.82
1.66
1.27
0.76
4.57
2.04
RBERT–Multi (idf)
1.15
6.97
0.83
3.65
0.68
3.32
1.37
FBERT–Multi (idf)
1.14
5.63
1.13
1.19
0.71
3.38
1.58
PXLM–100
1.22
6.30
1.14
0.79
0.74
3.73
2.21
RXLM–100
1.18
6.89
0.76
0.77
0.66
3.26
1.68
FXLM–100
1.19
6.44
0.82
0.76
0.69
3.21
1.57
PXLM–100 (idf)
1.21
6.61
1.07
0.78
0.72
5.59
2.02
RXLM–100 (idf)
1.19
7.07
0.77
0.77
0.66
3.33
1.60
FXLM–100 (idf)
1.20
6.57
0.86
0.76
0.68
3.28
1.56
Table 28: Absolute Difference (×100) of the top metric-rated and the top human-rated system on to-
English WMT18 hybrid systems. Smaller difference indicate higher agreement with human scores.
We report the average of 100K samples and the 0.95 confidence intervals are below 10−3. We bold
the lowest numbers for each language pair and direction.
41


Published as a conference paper at ICLR 2020
Metric
M1
M2
BLEU-1
0.124∗
0.135∗
BLEU-2
0.037∗
0.048∗
BLEU-3
0.004∗
0.016∗
BLEU-4
-0.019∗
-0.005∗
METEOR
0.606∗
0.594∗
ROUGE-L
0.090∗
0.096∗
CIDER
0.438∗
0.440∗
SPICE
0.759∗
0.750∗
LEIC
0.939∗
0.949∗
BEER
0.491
0.562
EED
0.545
0.599
CHRF++
0.702
0.729
CHARACTER
0.800
0.801
PBERT–Base
0.313
0.344
RBERT–Base
0.679
0.622
FBERT–Base
0.531
0.519
PBERT–Base (idf)
0.243
0.286
RBERT–Base (idf)
0.834
0.783
FBERT–Base (idf)
0.579
0.581
PBERT–Base–MRPC
0.252
0.331
RBERT–Base–MRPC
0.644
0.641
FBERT–Base–MRPC
0.470
0.512
PBERT–Base–MRPC (idf)
0.264
0.300
RBERT–Base–MRPC (idf)
0.794
0.767
FBERT–Base–MRPC (idf)
0.575
0.583
PBERT–Large
0.454
0.486
RBERT–Large
0.756
0.697
FBERT–Large
0.649
0.634
PBERT–Large (idf)
0.327
0.372
RBERT–Large (idf)
0.873
0.821
FBERT–Large (idf)
0.645
0.647
PRoBERTa–Base
-0.223
-0.179
RRoBERTa–Base
0.827
0.800
FRoBERTa–Base
0.176
0.191
PRoBERTa–Base (idf)
-0.256
-0.267
RRoBERTa–Base (idf)
0.901
0.869
FRoBERTa–Base (idf)
0.188
0.157
PRoBERTa–Large
-0.105
-0.041
RRoBERTa–Large
0.888
0.863
FRoBERTa–Large
0.322
0.350
PRoBERTa–Large (idf)
0.063
-0.011
RRoBERTa–Large (idf)
0.917
0.889
FRoBERTa–Large (idf)
0.519
0.453
PRoBERTa–Large–MNLI
0.129
0.208
RRoBERTa–Large–MNLI
0.820
0.823
FRoBERTa–Large–MNLI
0.546
0.592
PRoBERTa–Large–MNLI (idf)
0.081
0.099
RRoBERTa–Large–MNLI (idf)
0.906
0.875
FRoBERTa–Large–MNLI (idf)
0.605
0.596
PXLNet–Base
-0.046
0.080
RXLNet–Base
0.409
0.506
FXLNet–Base
0.146
0.265
PXLNet–Base (idf)
0.006
0.145
RXLNet–Base (idf)
0.655
0.720
FXLNet–Base (idf)
0.270
0.391
PXLNet–Large
-0.188
-0.115
RXLNet–Large
0.178
0.195
FXLNet–Large
-0.014
0.036
PXLNet–Large (idf)
-0.186
-0.072
RXLNet–Large (idf)
0.554
0.555
FXLNet–Large (idf)
0.151
0.234
PXLM–En
0.230
0.220
RXLM–En
0.333
0.263
FXLM–En
0.297
0.243
PXLM–En (idf)
0.266
0.275
RXLM–En (idf)
0.700
0.640
FXLM–En (idf)
0.499
0.470
Table 29: Pearson correlation on the 2015 COCO Captioning Challenge. The M1 and M2 measures
are described in Section 4. We bold the best correlating task-specific and task-agnostic metrics in
each setting LEIC uses images as additional inputs. Numbers with ∗are cited from Cui et al. (2018).
42


Published as a conference paper at ICLR 2020
Type
Method
QQP
PAWSQQP
Trained on QQP (supervised)
DecAtt
0.939*
0.263
DIIN
0.952*
0.324
BERT
0.963*
0.351
Trained on QQP + PAWSQQP (supervised)
DecAtt
-
0.511
DIIN
-
0.778
BERT
-
0.831
Metric (Not trained on QQP or PAWSQQP)
BLEU-1
0.737
0.402
BLEU-2
0.720
0.548
BLEU-3
0.712
0.527
BLEU-4
0.707
0.527
METEOR
0.755
0.532
ROUGE-L
0.740
0.536
CHRF++
0.577
0.608
BEER
0.741
0.564
EED
0.743
0.611
CHARACTER
0.698
0.650
PBERT–Base
0.750
0.654
RBERT–Base
0.739
0.655
FBERT–Base
0.755
0.654
PBERT–Base (idf)
0.766
0.665
RBERT–Base (idf)
0.752
0.665
FBERT–Base (idf)
0.770
0.664
PBERT–Base–MRPC
0.742
0.615
RBERT–Base–MRPC
0.729
0.617
FBERT–Base–MRPC
0.746
0.614
PBERT–Base–MRPC (idf)
0.752
0.618
RBERT–Base–MRPC (idf)
0.737
0.619
FBERT–Base–MRPC (idf)
0.756
0.617
PBERT–Large
0.752
0.706
RBERT–Large
0.740
0.710
FBERT–Large
0.756
0.707
PBERT–Large (idf)
0.766
0.713
RBERT–Large (idf)
0.751
0.718
FBERT–Large (idf)
0.769
0.714
PRoBERTa–Base
0.746
0.657
RRoBERTa–Base
0.736
0.656
FRoBERTa–Base
0.751
0.654
PRoBERTa–Base (idf)
0.760
0.666
RRoBERTa–Base (idf)
0.745
0.666
FRoBERTa–Base (idf)
0.765
0.664
PRoBERTa–Large
0.757
0.687
RRoBERTa–Large
0.744
0.685
FRoBERTa–Large
0.761
0.685
PRoBERTa–Large (idf)
0.773
0.691
RRoBERTa–Large (idf)
0.757
0.697
FRoBERTa–Large (idf)
0.777
0.693
PRoBERTa–Large–MNLI
0.763
0.767
RRoBERTa–Large–MNLI
0.750
0.772
FRoBERTa–Large–MNLI
0.766
0.770
PRoBERTa–Large–MNLI (idf)
0.783
0.756
RRoBERTa–Large–MNLI (idf)
0.767
0.764
FRoBERTa–Large–MNLI (idf)
0.784
0.759
PXLNet–Base
0.737
0.603
RXLNet–Base
0.731
0.607
FXLNet–Base
0.739
0.605
PXLNet–Base (idf)
0.751
0.625
RXLNet–Base (idf)
0.743
0.630
FXLNet–Base (idf)
0.751
0.626
PXLNet–Large
0.742
0.593
RXLNet–Large
0.734
0.598
FXLNet–Large
0.744
0.596
PXLNet–Large (idf)
0.759
0.604
RXLNet–Large (idf)
0.749
0.610
FXLNet–Large (idf)
0.760
0.606
PXLM–En
0.734
0.600
RXLM–En
0.725
0.604
FXLM–En
0.737
0.602
PXLM–En (idf)
0.757
0.596
RXLM–En (idf)
0.745
0.603
FXLM–En (idf)
0.759
0.600
Table 30: Area under ROC curve (AUC) on QQP and PAWSQQP datasets. The scores of trained
DecATT (Parikh et al., 2016), DIIN (Gong et al., 2018), and fine-tuned BERT are reported by Zhang
et al. (2019). We bold the best task-specific and task-agnostic metrics. Numbers with ∗are scores
on the held-out test set of QQP.
43