While many new automatic metrics for machine translation evaluation have been proposed in recent years, BLEU scores are still used as the primary metric in the vast majority of MT research papers. There are many reasons that researchers may be reluctant to switch to new metrics, from external pressures (reviewers, prior work) to the ease of use of metric toolkits. Another reason is a lack of intuition about the meaning of novel metric scores. In this work, we examine “rules of thumb” about metric score differences and how they do (and do not) correspond to human judgments of statistically significant differences between systems. In particular, we show that common rules of thumb about BLEU score differences do not in fact guarantee that human annotators will find significant differences between systems. We also show ways in which these rules of thumb fail to generalize across translation directions or domains.

We analyze the different data sampling approaches used in selecting data for human evaluation and ranking of machine translation systems at the highly influential Conference on Machine Translation (WMT). By using automatic evaluation metrics, we are able to focus on the impact of the data sampling procedure as separate from questions about human annotator consistency. We provide evidence that the latest data sampling approach used at WMT skews the annotated data toward shorter documents, not necessarily representative of the full test set. Lastly, we examine a new data sampling method that uses the available labour budget to sample data in a more representative manner, with the goals of improving representation of various document lengths in the sample and producing more stable rankings of system translation quality.

This paper presents the results of the WMT23 Metrics Shared Task. Participants submitting automatic MT evaluation metrics were asked to score the outputs of the translation systems competing in the WMT23 News Translation Task. All metrics were evaluated on how well they correlate with human ratings at the system and segment level. Similar to last year, we acquired our own human ratings based on expert-based human evaluation via Multidimensional Quality Metrics (MQM). Following last year’s success, we also included a challenge set subtask, where participants had to create contrastive test suites for evaluating metrics’ ability to capture and penalise specific types of translation errors. Furthermore, we improved our meta-evaluation procedure by considering fewer tasks and calculating a global score by weighted averaging across the various tasks. We present an extensive analysis on how well metrics perform on three language pairs: Chinese-English, Hebrew-English on the sentence-level and English-German on the paragraph-level. The results strongly confirm the results reported last year, that neural-based metrics are significantly better than non-neural metrics in their levels of correlation with human judgments. Further, we investigate the impact of bad reference translations on the correlations of metrics with human judgment. We present a novel approach for generating synthetic reference translations based on the collection of MT system outputs and their corresponding MQM ratings, which has the potential to mitigate bad reference issues we observed this year for some language pairs. Finally, we also study the connections between the magnitude of metric differences and their expected significance in human evaluation, which should help the community to better understand and adopt new metrics.

The Metric Score Landscape Challenge (MSLC23) dataset aims to gain insight into metric scores on a broader/wider landscape of machine translation (MT) quality. It provides a collection of low- to medium-quality MT output on the WMT23 general task test set. Together with the high quality systems submitted to the general task, this will enable better interpretation of metric scores across a range of different levels of translation quality. With this wider range of MT quality, we also visualize and analyze metric characteristics beyond just correlation.

This paper presents the results of the WMT22 Metrics Shared Task. Participants submitting automatic MT evaluation metrics were asked to score the outputs of the translation systems competing in the WMT22 News Translation Task on four different domains: news, social, ecommerce, and chat. All metrics were evaluated on how well they correlate with human ratings at the system and segment level. Similar to last year, we acquired our own human ratings based on expert-based human evaluation via Multidimensional Quality Metrics (MQM). This setup had several advantages, among other things: (i) expert-based evaluation is more reliable, (ii) we extended the pool of translations by 5 additional translations based on MBR decoding or rescoring which are challenging for current metrics. In addition, we initiated a challenge set subtask, where participants had to create contrastive test suites for evaluating metrics’ ability to capture and penalise specific types of translation errors. Finally, we present an extensive analysis on how well metrics perform on three language pairs: English to German, English to Russian and Chinese to English. The results demonstrate the superiority of neural-based learned metrics and demonstrate again that overlap metrics like Bleu, spBleu or chrf correlate poorly with human ratings. The results also reveal that neural-based metrics are remarkably robust across different domains and challenges.

We present a collection of expanded human annotations of the WMT20 English-Inuktitut machine translation shared task, covering the Nunavut Hansard portion of the dataset. Additionally, we recompute News rankings to take into account the completed set of human annotations and certain irregularities in the annotation task construction. We show the effect of these changes on the downstream task of the evaluation of automatic metrics. Finally, we demonstrate that character-level metrics correlate well with human judgments for the task of automatically evaluating translation into this polysynthetic language.

This paper presents the results of the WMT21 Metrics Shared Task. Participants were asked to score the outputs of the translation systems competing in the WMT21 News Translation Task with automatic metrics on two different domains: news and TED talks. All metrics were evaluated on how well they correlate at the system- and segment-level with human ratings. Contrary to previous years’ editions, this year we acquired our own human ratings based on expert-based human evaluation via Multidimensional Quality Metrics (MQM). This setup had several advantages: (i) expert-based evaluation has been shown to be more reliable, (ii) we were able to evaluate all metrics on two different domains using translations of the same MT systems, (iii) we added 5 additional translations coming from the same system during system development. In addition, we designed three challenge sets that evaluate the robustness of all automatic metrics. We present an extensive analysis on how well metrics perform on three language pairs: English to German, English to Russian and Chinese to English. We further show the impact of different reference translations on reference-based metrics and compare our expert-based MQM annotation with the DA scores acquired by WMT.

The Inuktitut language, a member of the Inuit-Yupik-Unangan language family, is spoken across Arctic Canada and noted for its morphological complexity. It is an official language of two territories, Nunavut and the Northwest Territories, and has recognition in additional regions. This paper describes a newly released sentence-aligned Inuktitut–English corpus based on the proceedings of the Legislative Assembly of Nunavut, covering sessions from April 1999 to June 2017. With approximately 1.3 million aligned sentence pairs, this is, to our knowledge, the largest parallel corpus of a polysynthetic language or an Indigenous language of the Americas released to date. The paper describes the alignment methodology used, the evaluation of the alignments, and preliminary experiments on statistical and neural machine translation (SMT and NMT) between Inuktitut and English, in both directions.

This paper presents the results of the news translation task and the similar language translation task, both organised alongside the Conference on Machine Translation (WMT) 2020. In the news task, participants were asked to build machine translation systems for any of 11 language pairs, to be evaluated on test sets consisting mainly of news stories. The task was also opened up to additional test suites to probe specific aspects of translation. In the similar language translation task, participants built machine translation systems for translating between closely related pairs of languages.

We present an extended study on using pretrained language models and YiSi-1 for machine translation evaluation. Although the recently proposed contextual embedding based metrics, YiSi-1, significantly outperform BLEU and other metrics in correlating with human judgment on translation quality, we have yet to understand the full strength of using pretrained language models for machine translation evaluation. In this paper, we study YiSi-1’s correlation with human translation quality judgment by varying three major attributes (which architecture; which inter- mediate layer; whether it is monolingual or multilingual) of the pretrained language mod- els. Results of the study show further improvements over YiSi-1 on the WMT 2019 Metrics shared task. We also describe the pretrained language model we trained for evaluating Inuktitut machine translation output.

We present a study on using YiSi-2 with massive multilingual pretrained language models for machine translation (MT) reference-less evaluation. Aiming at finding better semantic representation for semantic MT evaluation, we first test YiSi-2 with contextual embed- dings extracted from different layers of two different pretrained models, multilingual BERT and XLM-RoBERTa. We also experiment with learning bilingual mappings that trans- form the vector subspace of the source language to be closer to that of the target language in the pretrained model to obtain more accurate cross-lingual semantic similarity representations. Our results show that YiSi-2’s correlation with human direct assessment on translation quality is greatly improved by replacing multilingual BERT with XLM-RoBERTa and projecting the source embeddings into the tar- get embedding space using a cross-lingual lin- ear projection (CLP) matrix learnt from a small development set.

The National Research Council of Canada’s team submissions to the parallel corpus filtering task at the Fifth Conference on Machine Translation are based on two key components: (1) iteratively refined statistical sentence alignments for extracting sentence pairs from document pairs and (2) a crosslingual semantic textual similarity metric based on a pretrained multilingual language model, XLM-RoBERTa, with bilingual mappings learnt from a minimal amount of clean parallel data for scoring the parallelism of the extracted sentence pairs. The translation quality of the neural machine translation systems trained and fine-tuned on the parallel data extracted by our submissions improved significantly when compared to the organizers’ LASER-based baseline, a sentence-embedding method that worked well last year. For re-aligning the sentences in the document pairs (component 1), our statistical approach has outperformed the current state-of-the-art neural approach in this low-resource context.

We describe the neural machine translation (NMT) system developed at the National Research Council of Canada (NRC) for the Kazakh-English news translation task of the Fourth Conference on Machine Translation (WMT19). Our submission is a multi-source NMT taking both the original Kazakh sentence and its Russian translation as input for translating into English.

We present YiSi, a unified automatic semantic machine translation quality evaluation and estimation metric for languages with different levels of available resources. Underneath the interface with different language resources settings, YiSi uses the same representation for the two sentences in assessment. Besides, we show significant improvement in the correlation of YiSi-1’s scores with human judgment is made by using contextual embeddings in multilingual BERT–Bidirectional Encoder Representations from Transformers to evaluate lexical semantic similarity. YiSi is open source and publicly available.

We describe the National Research Council Canada team’s submissions to the parallel corpus filtering task at the Fourth Conference on Machine Translation.

We present a fully unsupervised crosslingual semantic textual similarity (STS) metric, based on contextual embeddings extracted from BERT – Bidirectional Encoder Representations from Transformers (Devlin et al., 2019). The goal of crosslingual STS is to measure to what degree two segments of text in different languages express the same meaning. Not only is it a key task in crosslingual natural language understanding (XLU), it is also particularly useful for identifying parallel resources for training and evaluating downstream multilingual natural language processing (NLP) applications, such as machine translation. Most previous crosslingual STS methods relied heavily on existing parallel resources, thus leading to a circular dependency problem. With the advent of massively multilingual context representation models such as BERT, which are trained on the concatenation of non-parallel data from each language, we show that the deadlock around parallel resources can be broken. We perform intrinsic evaluations on crosslingual STS data sets and extrinsic evaluations on parallel corpus filtering and human translation equivalence assessment tasks. Our results show that the unsupervised crosslingual STS metric using BERT without fine-tuning achieves performance on par with supervised or weakly supervised approaches.

The WMT18 shared task on parallel corpus filtering (Koehn et al., 2018b) challenged teams to score sentence pairs from a large high-recall, low-precision web-scraped parallel corpus (Koehn et al., 2018a). Participants could use existing sample corpora (e.g. past WMT data) as a supervisory signal to learn what a “clean” corpus looks like. However, in lower-resource situations it often happens that the target corpus of the language is the only sample of parallel text in that language. We therefore made several unsupervised entries, setting ourselves an additional constraint that we not utilize the additional clean parallel corpora. One such entry fairly consistently scored in the top ten systems in the 100M-word conditions, and for one task—translating the European Medicines Agency corpus (Tiedemann, 2009)—scored among the best systems even in the 10M-word conditions.

We present our semantic textual similarity approach in filtering a noisy web crawled parallel corpus using YiSi—a novel semantic machine translation evaluation metric. The systems mainly based on this supervised approach perform well in the WMT18 Parallel Corpus Filtering shared task (4th place in 100-million-word evaluation, 8th place in 10-million-word evaluation, and 6th place overall, out of 48 submissions). In fact, our best performing system—NRC-yisi-bicov is one of the only four submissions ranked top 10 in both evaluations. Our submitted systems also include some initial filtering steps for scaling down the size of the test corpus and a final redundancy removal step for better semantic and token coverage of the filtered corpus. In this paper, we also describe our unsuccessful attempt in automatically synthesizing a noisy parallel development corpus for tuning the weights to combine different parallelism and fluency features.

We discuss various improvements to our MEANT tuned system, previously presented at IWSLT 2013. In our 2014 system, we incorporate this year’s improved version of MEANT, improved Chinese word segmentation, Chinese named entity recognition and dedicated proper name translation, and number expression handling. This results in a significant performance jump compared to last year’s system. We also ran preliminary experiments on tuning to IMEANT, our new ITG based variant of MEANT. The performance of tuning to IMEANT is comparable to tuning on MEANT (differences are statistically insignificant). We are presently investigating if tuning on IMEANT can produce even better results, since IMEANT was actually shown to correlate with human adequacy judgment more closely than MEANT. Finally, we ran experiments applying our new architectural improvements to a contrastive system tuned to BLEU. We observed a slightly higher jump in comparison to last year, possibly due to mismatches of MEANT’s similarity models to our new entity handling.

We present analyses showing that HMEANT is a reliable, accurate and fine-grained semantic frame based human MT evaluation metric with high inter-annotator agreement (IAA) and correlation with human adequacy judgments, despite only requiring a minimal training of about 15 minutes for lay annotators. Previous work shows that the IAA on the semantic role labeling (SRL) subtask within HMEANT is over 70%. In this paper we focus on (1) the IAA on the semantic role alignment task and (2) the overall IAA of HMEANT. Our results show that the IAA on the alignment task of HMEANT is over 90% when humans align SRL output from the same SRL annotator, which shows that the instructions on the alignment task are sufficiently precise, although the overall IAA where humans align SRL output from different SRL annotators falls to only 61% due to the pipeline effect on the disagreement in the two annotation task. We show that instead of manually aligning the semantic roles using an automatic algorithm not only helps maintaining the overall IAA of HMEANT at 70%, but also provides a finer-grained assessment on the phrasal similarity of the semantic role fillers. This suggests that HMEANT equipped with automatic alignment is reliable and accurate for humans to evaluate MT adequacy while achieving higher correlation with human adequacy judgments than HTER.

We present the results of large-scale human semantic MT evaluation with HMEANT on the IWSLT 2013 German-English MT and SLT tracks and show that HMEANT evaluates the performance of the MT systems differently compared to BLEU and TER. Together with the references, all the translations are annotated by annotators who are native English speakers in both semantic role labeling stage and role filler alignment stage of HMEANT. We obtain high inter-annotator agreement and low annotation time costs which indicate that it is feasible to run a large-scale human semantic MT evaluation campaign using HMEANT. Our results also show that HMEANT is a robust and reliable semantic MT evaluation metric for running large-scale evaluation campaigns as it is inexpensive and simple while maintaining the semantic representational transparency to provide a perspective which is different from BLEU and TER in order to understand the performance of the state-of-the-art MT systems.

We present the first ever results showing that Chinese MT output is significantly improved by tuning a MT system against a semantic frame based objective function, MEANT, rather than an n-gram based objective function, BLEU, as measured across commonly used metrics and different test sets. Recent work showed that by preserving the meaning of the translations as captured by semantic frames in the training process, MT systems for translating into English on both formal and informal genres are constrained to produce more adequate translations by making more accurate choices on lexical output and reordering rules. In this paper we describe our experiments in IWSLT 2013 TED talk MT tasks on tuning MT systems against MEANT for translating into Chinese and English respectively. We show that the Chinese translation output benefits more from tuning a MT system against MEANT than the English translation output due to the ambiguous nature of word boundaries in Chinese. Our encouraging results show that using MEANT is a promising alternative to BLEU in both evaluating and tuning MT systems to drive the progress of MT research across different languages.

We present the methodology that underlies mew metrics for semantic machine translation evaluation we are developing. Unlike widely-used lexical and n-gram based MT evaluation metrics, the aim of semantic MT evaluation is to measure the utility of translations. We discuss the design of empirical studies to evaluate the utility of machine translation output by assessing the accuracy for key semantic roles. These roles are from the English 5W templates (who, what, when, where, why) used in recent GALE distillation evaluations. Recent work by Wu and Fung (2009) introduced semantic role labeling into statistical machine translation to enhance the quality of MT output. However, this approach has so far only been evaluated using lexical and n-gram based SMT evaluation metrics like BLEU which are not aimed at evaluating the utility of MT output. Direct data analysis are still needed to understand how semantic models can be leveraged to evaluate the utility of MT output. In this paper, we discuss a new methodology for evaluating the utility of the machine translation output, by assessing the accuracy with which human readers are able to complete the English 5W templates.

This paper describes the HKUST experiments in the IWSLT 2007 evaluation campaign on spoken language translation. Our primary objective was to compare the open-source phrase-based statistical machine translation toolkit Moses against Pharaoh. We focused on Chinese to English translation, but we also report results on the Arabic to English, Italian to English, and Japanese to English tasks.