Neural machine translation systems estimate probabilities of target sentences given source sentences, yet these estimates may not align with human preferences. This work introduces QE-fusion, a method that synthesizes translations using a quality estimation metric (QE), which correlates better with human judgments. QE-fusion leverages a pool of candidates sampled from a model, combining spans from different candidates using a QE metric such as CometKiwi. We compare QE-fusion against beam search and recent reranking techniques, such as Minimum Bayes Risk decoding or QE-reranking. Our method consistently improves translation quality in terms of COMET and BLEURT scores when applied to large language models (LLMs) used for translation (PolyLM, XGLM, Llama2, Mistral, ALMA, and Tower) and to multilingual translation models (NLLB), over five language pairs. Notably, QE-fusion exhibits larger improvements for LLMs due to their ability to generate diverse outputs. We demonstrate that our approach generates novel translations in over half of the cases and consistently outperforms other methods across varying numbers of candidates (5–200). Furthermore, we empirically establish that QE-fusion scales linearly with the number of candidates in the pool.

Training neural MT systems for low-resource language pairs or in unsupervised settings (i.e. with no parallel data) often involves a large number of auxiliary systems. These may include parent systems trained on higher-resource pairs and used for initializing the parameters of child systems, multilingual systems for neighboring languages, and several stages of systems trained on pseudo-parallel data obtained through back-translation. We propose here a simplified pipeline, which we compare to the best submissions to the WMT 2021 Shared Task on Unsupervised MT and Very Low Resource Supervised MT. Our pipeline only needs two parents, two children, one round of back-translation for low-resource directions and two for unsupervised ones and obtains better or similar scores when compared to more complex alternatives.

Poem generation with language models requires the modeling of rhyming patterns. We propose a novel solution for learning to rhyme, based on synthetic data generated with a rule-based rhyming algorithm. The algorithm and an evaluation metric use a phonetic dictionary and the definitions of perfect and assonant rhymes. We fine-tune a GPT-2 English model with 124M parameters on 142 MB of natural poems and find that this model generates consecutive rhymes infrequently (11%). We then fine-tune the model on 6 MB of synthetic quatrains with consecutive rhymes (AABB) and obtain nearly 60% of rhyming lines in samples generated by the model. Alternating rhymes (ABAB) are more difficult to model because of longer-range dependencies, but they are still learnable from synthetic data, reaching 45% of rhyming lines in generated samples.

Subword tokenization is the de-facto standard for tokenization in neural language models and machine translation systems. Three advantages are frequently put forward in favor of subwords: shorter encoding of frequent tokens, compositionality of subwords, and ability to deal with unknown words. As their relative importance is not entirely clear yet, we propose a tokenization approach that enables us to separate frequency (the first advantage) from compositionality, thanks to the use of Huffman coding, which tokenizes words using a fixed amount of symbols. Experiments with CS-DE, EN-FR and EN-DE NMT show that frequency alone accounts for approximately 90% of the BLEU scores reached by BPE, hence compositionality has less importance than previously thought.

We explore the roles and interactions of the hyper-parameters governing regularization, and propose a range of values applicable to low-resource neural machine translation. We demonstrate that default or recommended values for high-resource settings are not optimal for low-resource ones, and that more aggressive regularization is needed when resources are scarce, in proportion to their scarcity. We explain our observations by the generalization abilities of sharp vs. flat basins in the loss landscape of a neural network. Results for four regularization factors corroborate our claim: batch size, learning rate, dropout rate, and gradient clipping. Moreover, we show that optimal results are obtained when using several of these factors, and that our findings generalize across datasets of different sizes and languages.

This paper describes a system for interactive poem generation, which combines neural language models (LMs) for poem generation with explicit constraints that can be set by users on form, topic, emotion, and rhyming scheme. LMs cannot learn such constraints from the data, which is scarce with respect to their needs even for a well-resourced language such as French. We propose a method to generate verses and stanzas by combining LMs with rule-based algorithms, and compare several approaches for adjusting the words of a poem to a desired combination of topics or emotions. An approach to automatic rhyme setting using a phonetic dictionary is proposed as well. Our system has been demonstrated at public events, and log analysis shows that users found it engaging.

State-of-the-art multilingual systems rely on shared vocabularies that sufficiently cover all considered languages. To this end, a simple and frequently used approach makes use of subword vocabularies constructed jointly over several languages. We hypothesize that such vocabularies are suboptimal due to false positives (identical subwords with different meanings across languages) and false negatives (different subwords with similar meanings). To address these issues, we propose Subword Mapping and Anchoring across Languages (SMALA), a method to construct bilingual subword vocabularies. SMALA extracts subword alignments using an unsupervised state-of-the-art mapping technique and uses them to create cross-lingual anchors based on subword similarities. We demonstrate the benefits of SMALA for cross-lingual natural language inference (XNLI), where it improves zero-shot transfer to an unseen language without task-specific data, but only by sharing subword embeddings. Moreover, in neural machine translation, we show that joint subword vocabularies obtained with SMALA lead to higher BLEU scores on sentences that contain many false positives and false negatives.

In this paper, we present the systems submitted by our team from the Institute of ICT (HEIG-VD / HES-SO) to the Unsupervised MT and Very Low Resource Supervised MT task. We first study the improvements brought to a baseline system by techniques such as back-translation and initialization from a parent model. We find that both techniques are beneficial and suffice to reach performance that compares with more sophisticated systems from the 2020 task. We then present the application of this system to the 2021 task for low-resource supervised Upper Sorbian (HSB) to German translation, in both directions. Finally, we present a contrastive system for HSB-DE in both directions, and for unsupervised German to Lower Sorbian (DSB) translation, which uses multi-task training with various training schedules to improve over the baseline.

We study the role of an essential hyper-parameter that governs the training of Transformers for neural machine translation in a low-resource setting: the batch size. Using theoretical insights and experimental evidence, we argue against the widespread belief that batch size should be set as large as allowed by the memory of the GPUs. We show that in a low-resource setting, a smaller batch size leads to higher scores in a shorter training time, and argue that this is due to better regularization of the gradients during training.

We present a voice-based conversational agent which combines the robustness of chatbots and the utility of question answering (QA) systems. Indeed, while data-driven chatbots are typically user-friendly but not goal-oriented, QA systems tend to perform poorly at chitchat. The proposed chatbot relies on a controller which performs dialogue act classification and feeds user input either to a sequence-to-sequence chatbot or to a QA system. The resulting chatbot is a spoken QA application for the Google Home smart speaker. The system is endowed with general-domain knowledge from Wikipedia articles and uses coreference resolution to detect relatedness between questions. We present our choices of data sets for training and testing the components, and present the experimental results that helped us optimize the parameters of the chatbot. In particular, we discuss the appropriateness of using the SQuAD dataset for evaluating end-to-end QA, in the light of our system’s behavior.

In this paper, we present the ForwardQuestions data set, made of human-generated questions related to knowledge triples. This data set results from the conversion and merger of the existing SimpleDBPediaQA and SimpleQuestionsWikidata data sets, including the mapping of predicates from DBPedia to Wikidata, and the selection of ‘forward’ questions as opposed to ‘backward’ ones. The new data set can be used to generate novel questions given an unseen Wikidata triple, by replacing the subjects of existing questions with the new one and then selecting the best candidate questions using semantic and syntactic criteria. Evaluation results indicate that the question generation method using ForwardQuestions improves the quality of questions by about 20% with respect to a baseline not using ranking criteria.

Neural sequence-to-sequence networks with attention have achieved remarkable performance for machine translation. One of the reasons for their effectiveness is their ability to capture relevant source-side contextual information at each time-step prediction through an attention mechanism. However, the target-side context is solely based on the sequence model which, in practice, is prone to a recency bias and lacks the ability to capture effectively non-sequential dependencies among words. To address this limitation, we propose a target-side-attentive residual recurrent network for decoding, where attention over previous words contributes directly to the prediction of the next word. The residual learning facilitates the flow of information from the distant past and is able to emphasize any of the previously translated words, hence it gains access to a wider context. The proposed model outperforms a neural MT baseline as well as a memory and self-attention network on three language pairs. The analysis of the attention learned by the decoder confirms that it emphasizes a wider context, and that it captures syntactic-like structures.

This paper demonstrates that word sense disambiguation (WSD) can improve neural machine translation (NMT) by widening the source context considered when modeling the senses of potentially ambiguous words. We first introduce three adaptive clustering algorithms for WSD, based on k-means, Chinese restaurant processes, and random walks, which are then applied to large word contexts represented in a low-rank space and evaluated on SemEval shared-task data. We then learn word vectors jointly with sense vectors defined by our best WSD method, within a state-of-the-art NMT system. We show that the concatenation of these vectors, and the use of a sense selection mechanism based on the weighted average of sense vectors, outperforms several baselines including sense-aware ones. This is demonstrated by translation on five language pairs. The improvements are more than 1 BLEU point over strong NMT baselines, +4% accuracy over all ambiguous nouns and verbs, or +20% when scored manually over several challenging words.

We propose a method to decide whether two occurrences of the same noun in a source text should be translated consistently, i.e. using the same noun in the target text as well. We train and test classifiers that predict consistent translations based on lexical, syntactic, and semantic features. We first evaluate the accuracy of our classifiers intrinsically, in terms of the accuracy of consistency predictions, over a subset of the UN Corpus. Then, we also evaluate them in combination with phrase-based statistical MT systems for Chinese-to-English and German-to-English. We compare the automatic post-editing of noun translations with the re-ranking of the translation hypotheses based on the classifiers’ output, and also use these methods in combination. This improves over the baseline and closes up to 50% of the gap in BLEU scores between the baseline and an oracle classifier.

We implement a fully probabilistic model to combine the hypotheses of a Spanish anaphora resolution system with those of a Spanish-English machine translation system. The probabilities over antecedents are converted into probabilities for the features of translated pronouns, and are integrated with phrase-based MT using an additional translation model for pronouns. The system improves the translation of several Spanish personal and possessive pronouns into English, by solving translation divergencies such as ‘ella’ vs. ‘she’/‘it’ or ‘su’ vs. ‘his’/‘her’/‘its’/‘their’. On a test set with 2,286 pronouns, a baseline system correctly translates 1,055 of them, while ours improves this by 41. Moreover, with oracle antecedents, possessives are translated with an accuracy of 83%.

We present the first prototype of the SUMMA Platform: an integrated platform for multilingual media monitoring. The platform contains a rich suite of low-level and high-level natural language processing technologies: automatic speech recognition of broadcast media, machine translation, automated tagging and classification of named entities, semantic parsing to detect relationships between entities, and automatic construction / augmentation of factual knowledge bases. Implemented on the Docker platform, it can easily be deployed, customised, and scaled to large volumes of incoming media streams.

In this paper, we present a proof-of-concept implementation of a coreference-aware decoder for document-level machine translation. We consider that better translations should have coreference links that are closer to those in the source text, and implement this criterion in two ways. First, we define a similarity measure between source and target coreference structures, by projecting the target ones onto the source and reusing existing coreference metrics. Based on this similarity measure, we re-rank the translation hypotheses of a baseline system for each sentence. Alternatively, to address the lack of diversity of mentions in the MT hypotheses, we focus on mention pairs and integrate their coreference scores with MT ones, resulting in post-editing decisions for mentions. The experimental results for Spanish to English MT on the AnCora-ES corpus show that the second approach yields a substantial increase in the accuracy of pronoun translation, with BLEU scores remaining constant.

In this paper, we define and assess a reference-based metric to evaluate the accuracy of pronoun translation (APT). The metric automatically aligns a candidate and a reference translation using GIZA++ augmented with specific heuristics, and then counts the number of identical or different pronouns, with provision for legitimate variations and omitted pronouns. All counts are then combined into one score. The metric is applied to the results of seven systems (including the baseline) that participated in the DiscoMT 2015 shared task on pronoun translation from English to French. The APT metric reaches around 0.993-0.999 Pearson correlation with human judges (depending on the parameters of APT), while other automatic metrics such as BLEU, METEOR, or those specific to pronouns used at DiscoMT 2015 reach only 0.972-0.986 Pearson correlation.

Hierarchical attention networks have recently achieved remarkable performance for document classification in a given language. However, when multilingual document collections are considered, training such models separately for each language entails linear parameter growth and lack of cross-language transfer. Learning a single multilingual model with fewer parameters is therefore a challenging but potentially beneficial objective. To this end, we propose multilingual hierarchical attention networks for learning document structures, with shared encoders and/or shared attention mechanisms across languages, using multi-task learning and an aligned semantic space as input. We evaluate the proposed models on multilingual document classification with disjoint label sets, on a large dataset which we provide, with 600k news documents in 8 languages, and 5k labels. The multilingual models outperform monolingual ones in low-resource as well as full-resource settings, and use fewer parameters, thus confirming their computational efficiency and the utility of cross-language transfer.

This paper presents a solution to evaluate spoken post-editing of imperfect machine translation output by a human translator. We compare two approaches to the combination of machine translation (MT) and automatic speech recognition (ASR): a heuristic algorithm and a machine learning method. To obtain a data set with spoken post-editing information, we use the French version of TED talks as the source texts submitted to MT, and the spoken English counterparts as their corrections, which are submitted to an ASR system. We experiment with various levels of artificial ASR noise and also with a state-of-the-art ASR system. The results show that the combination of MT with ASR improves over both individual outputs of MT and ASR in terms of BLEU scores, especially when ASR performance is low.

This paper presents a method for verb phrase (VP) alignment in an English-French parallel corpus and its use for improving statistical machine translation (SMT) of verb tenses. The method starts from automatic word alignment performed with GIZA++, and relies on a POS tagger and a parser, in combination with several heuristics, in order to identify non-contiguous components of VPs, and to label the aligned VPs with their tense and voice on each side. This procedure is applied to the Europarl corpus, leading to the creation of a smaller, high-precision parallel corpus with about 320,000 pairs of finite VPs, which is made publicly available. This resource is used to train a tense predictor for translation from English into French, based on a large number of surface features. Three MT systems are compared: (1) a baseline phrase-based SMT; (2) a tense-aware SMT system using the above predictions within a factored translation model; and (3) a system using oracle predictions from the aligned VPs. For several tenses, such as the French “imparfait”, the tense-aware SMT system improves significantly over the baseline and is closer to the oracle system.

This paper shows how the disambiguation of discourse connectives can improve their automatic translation, while preserving the overall performance of statistical MT as measured by BLEU. State-of-the-art automatic classifiers for rhetorical relations are used prior to MT to label discourse connectives that signal those relations. These labels are used for MT in two ways: (1) by augmenting factored translation models; and (2) by using the probability distributions of labels in order to train and tune SMT. The improvement of translation quality is demonstrated using a new semi-automated metric for discourse connectives, on the English/French WMT10 data, while BLEU scores remain comparable to non-discourse-aware systems, due to the low frequency of discourse connectives.

Discourse connectives can often signal multiple discourse relations, depending on their context. The automatic identification of the Arabic translations of seven English discourse connectives shows how these connectives are differently translated depending on their actual senses. Automatic labelling of English source connectives can help a machine translation system to translate them more correctly. The corpus-based analysis of Arabic translations also enables the definition of a connective-specific evaluation metric for machine translation, which is here validated by human judges on sample English/Arabic translation data.

This paper describes methods and results for the annotation of two discourse-level phenomena, connectives and pronouns, over a multilingual parallel corpus. Excerpts from Europarl in English and French have been annotated with disambiguation information for connectives and pronouns, for about 3600 tokens. This data is then used in several ways: for cross-linguistic studies, for training automatic disambiguation software, and ultimately for training and testing discourse-aware statistical machine translation systems. The paper presents the annotation procedures and their results in detail, and overviews the first systems trained on the annotated resources and their use for machine translation.

This paper summarizes the latest, final version of ISO standard 24617-2 ``Semantic annotation framework, Part 2: Dialogue acts"""". Compared to the preliminary version ISO DIS 24617-2:2010, described in Bunt et al. (2010), the final version additionally includes concepts for annotating rhetorical relations between dialogue units, defines a full-blown compositional semantics for the Dialogue Act Markup Language DiAML (resulting, as a side-effect, in a different treatment of functional dependence relations among dialogue acts and feedback dependence relations); and specifies an optimally transparent XML-based reference format for the representation of DiAML annotations, based on the systematic application of the notion of `ideal concrete syntax'. We describe these differences and briefly discuss the design and implementation of an incremental method for dialogue act recognition, which proves the usability of the ISO standard for automatic dialogue annotation.

This paper describes an ISO project which aims at developing a standard for annotating spoken and multimodal dialogue with semantic information concerning the communicative functions of utterances, the kind of semantic content they address, and their relations with what was said and done earlier in the dialogue. The project, ISO 24617-2 ""Semantic annotation framework, Part 2: Dialogue acts"", is currently at DIS stage. The proposed annotation schema distinguishes 9 orthogonal dimensions, allowing each functional segment in dialogue to have a function in each of these dimensions, thus accounting for the multifunctionality that utterances in dialogue often have. A number of core communicative functions is defined in the form of ISO data categories, available at http://semantic-annotation.uvt.nl/dialogue-acts/iso-datcats.pdf; they are divided into ""dimension-specific"" functions, which can be used only in a particular dimension, such as Turn Accept in the Turn Management dimension, and ""general-purpose"" functions, which can be used in any dimension, such as Inform and Request. An XML-based annotation language, ""DiAML"" is defined, with an abstract syntax, a semantics, and a concrete syntax.

This paper presents recent results of the application of the task-based Browser Evaluation Test (BET) to meeting browsers, that is, interfaces to multimodal databases of meeting recordings. The tasks were defined by browser-neutral BET observers. Two groups of human subjects used the Transcript-based Query and Browsing interface (TQB), and attempted to solve as many BET tasks - pairs of true/false statements to disambiguate - as possible in a fixed amount of time. Their performance was measured in terms of precision and speed. Results indicate that the browser’s annotation-based search functionality is frequently used, in particular the keyword search. A more detailed analysis of each test question for each participant confirms that despite considerable variation across strategies, the use of queries is correlated to successful performance.

The Framework for the Evaluation for Machine Translation (FEMTI) contains guidelines for building a quality model that is used to evaluate MT systems in relation to the purpose and intended context of use of the systems. Contextual quality models can thus be constructed, but entering into FEMTI the knowledge required for this operation is a complex task. An experiment has been set up in order to transfer knowledge from MT evaluation experts into the FEMTI guidelines, by polling experts about the evaluation methods they would use in a particular context, then inferring from the results generic relations between characteristics of the context of use and quality characteristics. The results of this hands-on exercise, carried out as part of a conference tutorial, have served to refine FEMTI’s “generic contextual quality model” and to obtain feedback on the FEMTI guidelines in general.

In this paper, we propose a formal framework that takes into account the influence of the intended context of use of an NLP system on the procedure and the metrics used to evaluate the system. We introduce in particular the notion of a context-dependent quality model and explain how it can be adapted to a given context of use. More specifically, we define vector-space representations of contexts of use and of quality models, which are connected by a generic contextual quality model (GCQM). For each domain, experts in evaluation are needed to build a GCQM based on analytic knowledge and on previous evaluations, using the mechanism proposed here. The main inspiration source for this work is the FEMTI framework for the evaluation of machine translation, which implements partly the present model, and which is described briefly along with insights from other domains.

This article describes an interface for searching and browsing multimodal recordings of group meetings. We provide first an overall perspective of meeting processing and retrieval applications, and distinguish between the media/modalities that are recorded and the ones that are used for browsing. We then proceed to describe the data and the annotations that are stored in a meeting database. Two scenarios of use for the transcript-based query and browsing interface (TQB) are then outlined: search and browse vs. overview and browse. The main functionalities of TQB, namely the database backend and the multimedia rendering solutions are described. An outline of evaluation perspectives is finally provided, with a description of the user interaction features that will be monitored.

This article outlines the evaluation protocol and provides the main results of the French Evaluation Campaign for Machine Translation Systems, CESTA. Following the initial objectives and evaluation plans, the evaluation metrics are briefly described: along with fluency and adequacy assessed by human judges, a number of recently proposed automated metrics are used. Two evaluation campaigns were organized, the first one in the general domain, and the second one in the medical domain. Up to six systems translating from English into French, and two systems translating from Arabic into French, took part in the campaign. The numerical results illustrate the differences between classes of systems, and provide interesting indications about the reliability of the automated metrics for French as a target language, both by comparison to human judges and using correlations between metrics. The corpora that were produced, as well as the information about the reliability of metrics, constitute reusable resources for MT evaluation.

Cet article étudie la résolution des références à des entités lorsqu’une représentation informatique de ces entités est disponible. Nous nous intéressons à un corpus de dialogues entre humains, portant sur les grands titres de la presse francophone du jour, et proposons une méthode pour détecter et résoudre les références faites par les locuteurs aux articles des journaux. La détection des expressions nominales qui réfèrent à ces documents est réalisée grâce à une grammaire, alors que le problème de la détection des pronoms qui réfèrent aux documents est abordé par des moyens statistiques. La résolution de ces expressions, à savoir l’attribution des référents, fait quant à elle l’objet d’un algorithme inspiré de la résolution des coréférences. Ces propositions sont évaluées par le biais de mesures quantitatives spécifiques.

In this paper, we report on the results of a full-size evaluation campaign of various MT systems. This campaign is novel compared to the classical DARPA/NIST MT evaluation campaigns in the sense that French is the target language, and that it includes an experiment of meta-evaluation of various metrics claiming to better predict different attributes of translation quality. We first describe the campaign, its context, its protocol and the data we used. Then we summarise the results obtained by the participating systems and discuss the meta-evaluation of the metrics used.

This paper presents FEMTI, a web-based Framework for the Evaluation of Machine Translation in ISLE. FEMTI offers structured descriptions of potential user needs, linked to an overview of technical characteristics of MT systems. The description of possible systems is mainly articulated around the quality characteristics for software product set out in ISO/IEC standard 9126. Following the philosophy set out there and in the related 14598 series of standards, each quality characteristic bottoms out in metrics which may be applied to a particular instance of a system in order to judge how satisfactory the system is with respect to that characteristic. An evaluator can use the description of user needs to help identify the specific needs of his evaluation and the relations between them. He can then follow the pointers to system description to determine what metrics should be applied and how. In the current state of the framework, emphasis is on being exhaustive, including as much as possible of the information available in the literature on machine translation evaluation. Future work will aim at being more analytic, looking at characteristics and metrics to see how they relate to one another, validating metrics and investigating the correlation between particular metrics and human judgement.

This paper reports results from an experiment that was aimed at comparing evaluation metrics for machine translation. Implemented as a workshop at a major conference in 2002, the experiment defined an evaluation task, description of the metrics, as well as test data consisting of human and machine translations of two texts. Several metrics, either applicable by human judges or automated, were used, and the overall results were analyzed. It appeared that most human metrics and automated metrics provided in general consistent rankings of the various candidate translations; the ranking of the human translations matched the one provided by translation professionals; and human translations were distinguished from machine translations.