Widely used learned metrics for machine translation evaluation, such as Comet and Bleurt, estimate the quality of a translation hypothesis by providing a single sentence-level score. As such, they offer little insight into translation errors (e.g., what are the errors and what is their severity). On the other hand, generative large language models (LLMs) are amplifying the adoption of more granular strategies to evaluation, attempting to detail and categorize translation errors. In this work, we introduce xcomet, an open-source learned metric designed to bridge the gap between these approaches. xcomet integrates both sentence-level evaluation and error span detection capabilities, exhibiting state-of-the-art performance across all types of evaluation (sentence-level, system-level, and error span detection). Moreover, it does so while highlighting and categorizing error spans, thus enriching the quality assessment. We also provide a robustness analysis with stress tests, and show that xcomet is largely capable of identifying localized critical errors and hallucinations.

Neural metrics for machine translation evaluation, such as COMET, exhibit significant improvements in their correlation with human judgments, as compared to traditional metrics based on lexical overlap, such as BLEU. Yet, neural metrics are, to a great extent, “black boxes” returning a single sentence-level score without transparency about the decision-making process. In this work, we develop and compare several neural explainability methods and demonstrate their effectiveness for interpreting state-of-the-art fine-tuned neural metrics. Our study reveals that these metrics leverage token-level information that can be directly attributed to translation errors, as assessed through comparison of token-level neural saliency maps with Multidimensional Quality Metrics (MQM) annotations and with synthetically-generated critical translation errors. To ease future research, we release our code at: https://github.com/Unbabel/COMET/tree/explainable-metrics

In recent years, there has been a notable rise in research interest regarding the integration of gender-inclusive and gender-neutral language in natural language processing models. A specific area of focus that has gained practical and academic significant interest is gender-neutral rewriting, which involves converting binary-gendered text to its gender-neutral counterpart. However, current approaches to gender-neutral rewriting for gendered languages tend to rely on large datasets, which may not be an option for languages with fewer resources, such as Portuguese. In this paper, we present a rule-based and a neural-based tool for gender-neutral rewriting for Portuguese, a heavily gendered Romance language whose morphology creates different challenges from the ones tackled by other gender-neutral rewriters. Our neural approach relies on fine-tuning large multilingual machine translation models on examples generated by the rule-based model. We evaluate both models on texts from different sources and contexts. We provide the first Portuguese dataset explicitly containing gender-neutral language and neopronouns, as well as a manually annotated golden collection of 500 sentences that allows for evaluation of future work.

Portuguese Sign Language (LGP) is the official language in deaf education in Portugal. Current approaches in developing a translation system between European Portuguese and LGP rely on hand-crafted rules. In this paper, we present a fully automatic corpora-driven rule-based machine translation system between European Portuguese and LGP glosses, and also two neural machine translation models. We also contribute with the LGP-5-Domain corpus, composed of five different text domains, built with the help of our rule-based system, and used to train the neural models. In addition, we provide a gold collection, annotated by LGP experts, that can be used for future evaluations. Compared with the only similar available translation system, PE2LGP, results are always improved with the new rule-based model, which competes for the highest scores with one of the neural models.

Recent approaches have attempted to personalize dialogue systems by leveraging profile information into models. However, this knowledge is scarce and difficult to obtain, which makes the extraction/generation of profile information from dialogues a fundamental asset. To surpass this limitation, we introduce the Profile Generation Task (PGTask). We contribute with a new dataset for this problem, comprising profile sentences aligned with related utterances, extracted from a corpus of dialogues. Furthermore, using state-of-the-art methods, we provide a benchmark for profile generation on this novel dataset. Our experiments disclose the challenges of profile generation, and we hope that this introduces a new research direction.

Task-oriented dialogue systems need to generate appropriate responses to help fulfill users’ requests. This paper explores different strategies, namely prompting, retrieval, and fine-tuning, for task-oriented dialogue generation. Through a systematic evaluation, we aim to provide valuable insights and guidelines for researchers and practitioners working on developing efficient and effective dialogue systems for real-world applications. Evaluation is performed on the MultiWOZ and Taskmaster-2 datasets, and we test various versions of FLAN-T5, GPT-3.5, and GPT-4 models. Costs associated with running these models are analyzed, and dialogue evaluation is briefly discussed. Our findings suggest that when testing data differs from the training data, fine-tuning may decrease performance, favoring a combination of a more general language model and a prompting mechanism based on retrieved examples.

Active learning can play an important role in low-resource settings (i.e., where annotated data is scarce), by selecting which instances may be more worthy to annotate. Most active learning approaches for Machine Translation assume the existence of a pool of sentences in a source language, and rely on human annotators to provide translations or post-edits, which can still be costly. In this article, we apply active learning to a real-world human-in-the-loop scenario in which we assume that: (1) the source sentences may not be readily available, but instead arrive in a stream; (2) the automatic translations receive feedback in the form of a rating, instead of a correct/edited translation, since the human-in-the-loop might be a user looking for a translation, but not be able to provide one. To tackle the challenge of deciding whether each incoming pair source–translations is worthy to query for human feedback, we resort to a number of stream-based active learning query strategies. Moreover, because we do not know in advance which query strategy will be the most adequate for a certain language pair and set of Machine Translation models, we propose to dynamically combine multiple strategies using prediction with expert advice. Our experiments on different language pairs and feedback settings show that using active learning allows us to converge on the best Machine Translation systems with fewer human interactions. Furthermore, combining multiple strategies using prediction with expert advice outperforms several individual active learning strategies with even fewer interactions, particularly in partial feedback settings.

We present the joint contribution of Unbabel and Instituto Superior Técnico to the WMT 2023 Shared Task on Quality Estimation (QE). Our team participated on all tasks: Sentence- and Word-level Quality Prediction and Fine-grained error span detection. For all tasks we build on the CometKiwi model (rei et al. 2022). Our multilingual approaches are ranked first for all tasks, reaching state-of-the-art performance for quality estimation at word-, span- and sentence-level granularity. Compared to the previous state-of-the-art, CometKiwi, we show large improvements in correlation with human judgements (up to 10 Spearman points) and surpassing the second-best multilingual submission with up to 3.8 absolute points.

In recent years, several neural fine-tuned machine translation evaluation metrics such as COMET and BLEURT have been proposed. These metrics achieve much higher correlations with human judgments than lexical overlap metrics at the cost of computational efficiency and simplicity, limiting their applications to scenarios in which one has to score thousands of translation hypothesis (e.g. scoring multiple systems or Minimum Bayes Risk decoding). In this paper, we explore optimization techniques, pruning, and knowledge distillation to create more compact and faster COMET versions. Our results show that just by optimizing the code through the use of caching and length batching we can reduce inference time between 39% and 65% when scoring multiple systems. Also, we show that pruning COMET can lead to a 21% model reduction without affecting the model’s accuracy beyond 0.01 Kendall tau correlation. Furthermore, we present DISTIL-COMET a lightweight distilled version that is 80% smaller and 2.128x faster while attaining a performance close to the original model and above strong baselines such as BERTSCORE and PRISM.

In this paper, we present the joint contribution of Unbabel and IST to the WMT 2022 Metrics Shared Task. Our primary submission – dubbed COMET-22 – is an ensemble between a COMET estimator model trained with Direct Assessments and a newly proposed multitask model trained to predict sentence-level scores along with OK/BAD word-level tags derived from Multidimensional Quality Metrics error annotations. These models are ensembled together using a hyper-parameter search that weights different features extracted from both evaluation models and combines them into a single score. For the reference-free evaluation, we present CometKiwi. Similarly to our primary submission, CometKiwi is an ensemble between two models. A traditional predictor-estimator model inspired by OpenKiwi and our new multitask model trained on Multidimensional Quality Metrics which can also be used without references. Both our submissions show improved correlations compared to state-of-the-art metrics from last year as well as increased robustness to critical errors.

We present the joint contribution of IST and Unbabel to the WMT 2022 Shared Task on Quality Estimation (QE). Our team participated in all three subtasks: (i) Sentence and Word-level Quality Prediction; (ii) Explainable QE; and (iii) Critical Error Detection. For all tasks we build on top of the COMET framework, connecting it with the predictor-estimator architecture of OpenKiwi, and equipping it with a word-level sequence tagger and an explanation extractor. Our results suggest that incorporating references during pretraining improves performance across several language pairs on downstream tasks, and that jointly training with sentence and word-level objectives yields a further boost. Furthermore, combining attention and gradient information proved to be the top strategy for extracting good explanations of sentence-level QE models. Overall, our submissions achieved the best results for all three tasks for almost all language pairs by a considerable margin.

In Machine Translation, assessing the quality of a large amount of automatic translations can be challenging. Automatic metrics are not reliable when it comes to high performing systems. In addition, resorting to human evaluators can be expensive, especially when evaluating multiple systems. To overcome the latter challenge, we propose a novel application of online learning that, given an ensemble of Machine Translation systems, dynamically converges to the best systems, by taking advantage of the human feedback available. Our experiments on WMT’19 datasets show that our online approach quickly converges to the top-3 ranked systems for the language pairs considered, despite the lack of human feedback for many translations.

We present MT-Telescope, a visualization platform designed to facilitate comparative analysis of the output quality of two Machine Translation (MT) systems. While automated MT evaluation metrics are commonly used to evaluate MT systems at a corpus-level, our platform supports fine-grained segment-level analysis and interactive visualisations that expose the fundamental differences in the performance of the compared systems. MT-Telescope also supports dynamic corpus filtering to enable focused analysis on specific phenomena such as; translation of named entities, handling of terminology, and the impact of input segment length on translation quality. Furthermore, the platform provides a bootstrapped t-test for statistical significance as a means of evaluating the rigor of the resulting system ranking. MT-Telescope is open source, written in Python, and is built around a user friendly and dynamic web interface. Complementing other existing tools, our platform is designed to facilitate and promote the broader adoption of more rigorous analysis practices in the evaluation of MT quality.

We present AIA-BDE, a corpus of 380 domain-oriented FAQs in Portuguese and their variations, i.e., paraphrases or entailed questions, created manually, by humans, or automatically, with Google Translate. Its aims to be used as a benchmark for FAQ retrieval and automatic question-answering, but may be useful in other contexts, such as the development of task-oriented dialogue systems, or models for natural language inference in an interrogative context. We also report on two experiments. Matching variations with their original questions was not trivial with a set of unsupervised baselines, especially for manually created variations. Besides high performances obtained with ELMo and BERT embeddings, an Information Retrieval system was surprisingly competitive when considering only the first hit. In the second experiment, text classifiers were trained with the original questions, and tested when assigning each variation to one of three possible sources, or assigning them as out-of-domain. Here, the difference between manual and automatic variations was not so significant.

Sign Languages are visual languages and the main means of communication used by Deaf people. However, the majority of the information available online is presented through written form. Hence, it is not of easy access to the Deaf community. Avatars that can animate sign languages have gained an increase of interest in this area due to their flexibility in the process of generation and edition. Synthetic animation of conversational agents can be achieved through the use of notation systems. HamNoSys is one of these systems, which describes movements of the body through symbols. Its XML-compliant, SiGML, is a machine-readable input of HamNoSys able to animate avatars. Nevertheless, current tools have no freely available open source libraries that allow the conversion from HamNoSys to SiGML. Our goal is to develop a tool of open access, which can perform this conversion independently from other platforms. This system represents a crucial intermediate step in the bigger pipeline of animating signing avatars. Two cases studies are described in order to illustrate different applications of our tool.

Software for the production of sign languages is much less common than for spoken languages. Such software usually relies on 3D humanoid avatars to produce signs which, inevitably, necessitates the use of animation. One barrier to the use of popular animation tools is their complexity and steep learning curve, which can be hard to master for inexperienced users. Here, we present PE2LGP, an authoring system that features a 3D avatar that signs Portuguese Sign Language. Our Animator is designed specifically to craft sign language animations using a key frame method, and is meant to be easy to use and learn to users without animation skills. We conducted a preliminary evaluation of the Animator, where we animated seven Portuguese Sign Language sentences and asked four sign language users to evaluate their quality. This evaluation revealed that the system, in spite of its simplicity, is indeed capable of producing comprehensible messages.

Building large datasets annotated with semantic information, such as FrameNet, is an expensive process. Consequently, such resources are unavailable for many languages and specific domains. This problem can be alleviated by using unsupervised approaches to induce the frames evoked by a collection of documents. That is the objective of the second task of SemEval 2019, which comprises three subtasks: clustering of verbs that evoke the same frame and clustering of arguments into both frame-specific slots and semantic roles. We approach all the subtasks by applying a graph clustering algorithm on contextualized embedding representations of the verbs and arguments. Using such representations is appropriate in the context of this task, since they provide cues for word-sense disambiguation. Thus, they can be used to identify different frames evoked by the same words. Using this approach we were able to outperform all of the baselines reported for the task on the test set in terms of Purity F1, as well as in terms of BCubed F1 in most cases.

We propose BeamSeg, a joint model for segmentation and topic identification of documents from the same domain. The model assumes that lexical cohesion can be observed across documents, meaning that segments describing the same topic use a similar lexical distribution over the vocabulary. The model implements lexical cohesion in an unsupervised Bayesian setting by drawing from the same language model segments with the same topic. Contrary to previous approaches, we assume that language models are not independent, since the vocabulary changes in consecutive segments are expected to be smooth and not abrupt. We achieve this by using a dynamic Dirichlet prior that takes into account data contributions from other topics. BeamSeg also models segment length properties of documents based on modality (textbooks, slides, etc.). The evaluation is carried out in three datasets. In two of them, improvements of up to 4.8% and 7.3% are obtained in the segmentation and topic identifications tasks, indicating that both tasks should be jointly modeled.

This paper describes our approach to the SemEval-2017 “Semantic Textual Similarity” and “Multilingual Word Similarity” tasks. In the former, we test our approach in both English and Spanish, and use a linguistically-rich set of features. These move from lexical to semantic features. In particular, we try to take advantage of the recent Abstract Meaning Representation and SMATCH measure. Although without state of the art results, we introduce semantic structures in textual similarity and analyze their impact. Regarding word similarity, we target the English language and combine WordNet information with Word Embeddings. Without matching the best systems, our approach proved to be simple and effective.

In this paper we describe a corpus of automatic translations annotated with both error type and quality. The 300 sentences that we have selected were generated by Google Translate, Systran and two in-house Machine Translation systems that use Moses technology. The errors present on the translations were annotated with an error taxonomy that divides errors in five main linguistic categories (Orthography, Lexis, Grammar, Semantics and Discourse), reflecting the language level where the error is located. After the error annotation process, we accessed the translation quality of each sentence using a four point comprehension scale from 1 to 5. Both tasks of error and quality annotation were performed by two different annotators, achieving good levels of inter-annotator agreement. The creation of this corpus allowed us to use it as training data for a translation quality classifier. We concluded on error severity by observing the outputs of two machine learning classifiers: a decision tree and a regression model.

We present JUST.ASK, a publicly available Question Answering system, which is freely available. Its architecture is composed of the usual Question Processing, Passage Retrieval and Answer Extraction components. Several details on the information generated and manipulated by each of these components are also provided to the user when interacting with the demonstration. Since JUST.ASK also learns to answer new questions based on users’ feedback, (s)he is invited to identify the correct answers. These will then be used to retrieve answers to future questions.

Analysing the translation errors is a task that can help us finding and describing translation problems in greater detail, but can also suggest where the automatic engines should be improved. Having these aims in mind we have created a corpus composed of 150 sentences, 50 from the TAP magazine, 50 from a TED talk and the other 50 from the from the TREC collection of factoid questions. We have automatically translated these sentences from English into Portuguese using Google Translate and Moses. After we have analysed the errors and created the error annotation taxonomy, the corpus was annotated by a linguist native speaker of Portuguese. Although Google’s overall performance was better in the translation task (we have also calculated the BLUE and NIST scores), there are some error types that Moses was better at coping with, specially discourse level errors.

The task of Statistical Machine Translation depends on large amounts of training corpora. Despite the availability of several parallel corpora, these are typically composed of declarative sentences, which may not be appropriate when the goal is to translate other types of sentences, e.g., interrogatives. There have been efforts to create corpora of questions, specially in the context of the evaluation of Question-Answering systems. One of those corpora is the UIUC dataset, composed of nearly 6,000 questions, widely used in the task of Question Classification. In this work, we make available the Portuguese version of the UIUC dataset, which we manually translated, as well as the translation guidelines. We show the impact of this corpus in the performance of a state-of-the-art SMT system when translating questions. Finally, we present a taxonomy of translation errors, according to which we analyze the output of the automatic translation before and after using the corpus as training data.

The WordNet knowledge model is currently implemented in multiple software frameworks providing procedural access to language instances of it. Frameworks tend to be focused on structural/design aspects of the model thus describing low level interfaces for linguistic knowledge retrieval. Typically the only high level feature directly accessible is word lookup while traversal of semantic relations leads to verbose/complex combinations of data structures, pointers and indexes which are irrelevant in an NLP context. Here is described an extension to the JWNL framework that hides technical requirements of access to WordNet features with an essentially word/sense based API applying terminology from the official online interface. This high level API is applied to the original English version of WordNet and to an SQL based Portuguese lexicon, translated into a WordNet based representation usable by JWNL.

In Statistical Machine Translation, words that were not seen during training are unknown words, that is, words that the system will not know how to translate. In this paper we contribute to this research problem by profiting from orthographic cues given by words. Thus, we report a study of the impact of word distance metrics in cognates' detection and, in addition, on the possibility of obtaining possible translations of unknown words through Logical Analogy. Our approach is tested in the translation of corpora from Portuguese to English (and vice-versa).

This work describes a process to extract Named Entity (NE) translations from the text available in web links (anchor texts). It translates a NE by retrieving a list of web documents in the target language, extracting the anchor texts from the links to those documents and finding the best translation from the anchor texts, using a combination of features, some of which, are specific to anchor texts. Experiments performed on a manually built corpora, suggest that over 70% of the NEs, ranging from unpopular to popular entities, can be translated correctly using sorely anchor texts. Tests on a Machine Translation task indicate that the system can be used to improve the quality of the translations of state-of-the-art statistical machine translation systems.

In this paper we describe the Instituto de Engenharia de Sistemas e Computadores Investigac ̧a ̃o e Desenvolvimento (INESC-ID) system that participated in the IWSLT 2010 evaluation campaign. Our main goal for this evaluation was to employ several state-of-the-art methods applied to phrase-based machine translation in order to improve the translation quality. Aside from the IBM M4 alignment model, two constrained alignment models were tested, which produced better overall results. These results were further improved by using weighted alignment matrixes during phrase extraction, rather than the single best alignment. Finally, we tested several filters that ruled out phrase pairs based on puntuation. Our system was evaluated on the BTEC and DIALOG tasks, having achieved a better overall ranking in the DIALOG task.

Phrase-based systems deeply depend on the quality of their phrase tables and therefore, the process of phrase extraction is always a fundamental step. In this paper we present a general and extensible phrase extraction algorithm, where we have highlighted several control points. The instantiation of these control points allows the simulation of previous approaches, as in each one of these points different strategies/heuristics can be tested. We show how previous approaches fit in this algorithm, compare several of them and, in addition, we propose alternative heuristics, showing their impact on the final translation results. Considering two different test scenarios from the IWSLT 2010 competition (BTEC, Fr-En and DIALOG, Cn-En), we have obtained an improvement in the results of 2.4 and 2.8 BLEU points, respectively.

Named Entity Recognition (NER) plays a relevant role in several Natural Language Processing tasks. Question-Answering (QA) is an example of such, since answers are frequently named entities in agreement with the semantic category expected by a given question. In this context, the recognition of named entities is usually applied in free text data. NER in natural language questions can also aid QA and, thus, should not be disregarded. Nevertheless, it has not yet been given the necessary importance. In this paper, we approach the identification and classification of named entities in natural language questions. We hypothesize that NER results can benefit with the inclusion of previously labeled questions in the training corpus. We present a broad study addressing that hypothesis, focusing on the balance to be achieved between the amount of free text and questions in order to build a suitable training corpus. This work also contributes by providing a set of nearly 5,500 annotated questions with their named entities, freely available for research purposes.

This paper reports an experience on producing manual word alignments over six different language pairs (all combinations between Portuguese, English, French and Spanish) (Graça et al., 2008). Word alignment of each language pair is made over the first 100 sentences of the common test set from the Europarl corpora (Koehn, 2005), corresponding to 600 new annotated sentences. This collection is publicly available at http://www.l2f.inesc- id.pt/resources/translation/. It contains, to our knowledge, the first word alignment gold set for the Portuguese language, with three other languages. Besides, it is to our knowledge, the first multi-language manual word aligned parallel corpus, where the same sentences are annotated for each language pair. We started by using the guidelines presented at (Mariño, 2005) and performed several refinements: some due to under-specifications on the original guidelines, others because of disagreement on some choices. This lead to the development of an extensive new set of guidelines for multi-lingual word alignment annotation that, we believe, makes the alignment process less ambiguous. We evaluate the inter-annotator agreement obtaining an average of 91.6% agreement between the different language pairs.

We present the machine translation system used by L2F from INESC-ID in the evaluation campaign of the International Workshop on Spoken Language Translation (2007), in the task of translating spontaneous conversations in the travel domain from Italian to English.