This paper assesses the ability of large language models (LLMs) to translate texts that include inter-sentential dependencies. We use the English-French DiscEvalMT benchmark (Bawden et al., 2018) with pairs of sentences containing translation challenges for pronominal anaphora and lexical cohesion. We evaluate 12 LLMs from the DeepSeek-R1, GPT, Llama, Mistral and Phi families on two tasks: (1) distinguish a correct translation from a wrong but plausible one; and (2) generate a correct translation. We compare prompts that encourage chain-of-thought reasoning with those that do not. The best models take advantage of reasoning and reach about 90% accuracy on the first task and COMET scores of about 92% on the second task, with GPT-4, GPT-4o and Phi standing out. Moreover, we observe a "wise get wiser" effect: the improvements through reasoning are larger for models that already perform well without reasoning.

Multilingual NMT is a viable solution for translating low-resource languages (LRLs) when data from high-resource languages (HRLs) from the same language family is available. However, the training schedule, i.e. the order of presentation of languages, has an impact on the quality of such systems. Here, in a many-to-one translation setting, we propose to apply two algorithms that use reinforcement learning to optimize the training schedule of NMT: (1) Teacher-Student Curriculum Learning and (2) Deep Q Network. The former uses an exponentially smoothed estimate of the returns of each action based on the loss on monolingual or multilingual development subsets, while the latter estimates rewards using an additional neural network trained from the history of actions selected in different states of the system, together with the rewards received. On a 8-to-1 translation dataset with LRLs and HRLs, our second method improves BLEU and COMET scores with respect to both random selection of monolingual batches and shuffled multilingual batches, by adjusting the number of presentations of LRL vs. HRL batches.

The popularity of automatic speech-to-speech translation for human conversations is growing, but the quality varies significantly depending on the language pair. In a context of community interpreting for low-resource languages, namely Turkish and Pashto to/from French, we collected fine-tuning and testing data, and compared systems using several automatic metrics (BLEU, COMET, and BLASER) and human assessments. The pipelines consist of automatic speech recognition, machine translation, and speech synthesis, with local models and cloud-based commercial ones. Some components have been fine-tuned on our data. We evaluated over 60 pipelines and determined the best one for each direction. We also found that the ranks of components are generally independent of the rest of the pipeline.

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 and combines 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 .Furthermore, we empirically show that QE-fusion scales linearly with the number of candidates in the pool.

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.

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.

Lesly Miculicich Werlen, Nikolaos Pappas, Dhananjay Ram, Andrei Popescu-Belis. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

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 Germanto-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.

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.

Representing documents is a crucial component in many NLP tasks, for instance predicting aspect ratings in reviews. Previous methods for this task treat documents globally, and do not acknowledge that target categories are often assigned by their authors with generally no indication of the specific sentences that motivate them. To address this issue, we adopt a weakly supervised learning model, which jointly learns to focus on relevant parts of a document according to the context along with a classifier for the target categories. Derived from the weighted multiple-instance regression (MIR) framework, the model learns decomposable document vectors for each individual category and thus overcomes the representational bottleneck in previous methods due to a fixed-length document vector. During prediction, the estimated relevance or saliency weights explicitly capture the contribution of each sentence to the predicted rating, thus offering an explanation of the rating. Our model achieves state-of-the-art performance on multi-aspect sentiment analysis, improving over several baselines. Moreover, the predicted saliency weights are close to human estimates obtained by crowdsourcing, and increase the performance of lexical and topical features for review segmentation and summarization.

We propose a method for extractive summarization of audiovisual recordings focusing on topic-level segments. We first build a content similarity graph between all segments across the collection, using word vectors from the transcripts, and then select the most central segments for the summaries. We evaluate the method quantitatively on the AMI Meeting Corpus using gold standard reference summaries and the Rouge metric, and qualitatively on lecture recordings using a novel two-tiered approach with human judges. The results show that our method compares favorably with others in terms of Rouge, and outperforms the baselines for human scores, thus also validating our evaluation protocol.

This paper addresses the problem of keyword extraction from conversations, with the goal of using these keywords to retrieve, for each short conversation fragment, a small number of potentially relevant documents, which can be recommended to participants. However, even a short fragment contains a variety of words, which are potentially related to several topics; moreover, using an automatic speech recognition (ASR) system introduces errors among them. Therefore, it is difficult to infer precisely the information needs of the conversation participants. We first propose an algorithm to extract keywords from the output of an ASR system (or a manual transcript for testing), which makes use of topic modeling techniques and of a submodular reward function which favors diversity in the keyword set, to match the potential diversity of topics and reduce ASR noise. Then, we propose a method to derive multiple topically separated queries from this keyword set, in order to maximize the chances of making at least one relevant recommendation when using these queries to search over the English Wikipedia. The proposed methods are evaluated in terms of relevance with respect to conversation fragments from the Fisher, AMI, and ELEA conversational corpora, rated by several human judges. The scores show that our proposal improves over previous methods that consider only word frequency or topic similarity, and represents a promising solution for a document recommender system to be used in conversations.

This paper shows that the automatic labeling of discourse connectives with the relations they signal, prior to machine translation (MT), can be used by phrase-based statistical MT systems to improve their translations. This improvement is demonstrated here when translating from English to four target languages-French, German, Italian and Arabic-using several test sets from recent MT evaluation campaigns. Using automatically labeled data for training, tuning and testing MT systems is beneficial on condition that labels are sufficiently accurate, typically above 70%. To reach such an accuracy, a large array of features for discourse connective labeling (morpho-syntactic, semantic and discursive) are extracted using state-of-the-art tools and exploited in factored MT models. The translation of connectives is improved significantly, between 0.7% and 10% as measured with the dedicated ACT metric. The improvements depend mainly on the level of ambiguity of the connectives in the test sets.

This paper introduces a model of multiple-instance learning applied to the prediction of aspect ratings or judgments of specific properties of an item from user-contributed texts such as product reviews. Each variable-length text is represented by several independent feature vectors; one word vector per sentence or paragraph. For learning from texts with known aspect ratings, the model performs multiple-instance regression (MIR) and assigns importance weights to each of the sentences or paragraphs of a text, uncovering their contribution to the aspect ratings. Next, the model is used to predict aspect ratings in previously unseen texts, demonstrating interpretability and explanatory power for its predictions. We evaluate the model on seven multi-aspect sentiment analysis data sets, improving over four MIR baselines and two strong bag-of-words linear models, namely SVR and Lasso, by more than 10% relative in terms of MSE.

In this paper, we present an approach for topic-level search and hyperlinking of video snippets, which relies on content-based recommendation and multimodal re-ranking techniques. We identify topic-level segments using transcripts or subtitles and enrich them with other metadata. Segments are indexed in a word vector space. Given a text query or an anchor, the most similar segments are retrieved using cosine similarity scores, which are then combined with visual similarity scores, computed as the distance from the anchor's visual concept vector. This approach has performed well on the MediaEval 2013 Search and Hyperlinking task, evaluated over 1260 hours of BBC TV broadcast, in terms of overall mean average precision. Experiments showed that topic-segments based on transcripts from automatic speech recognition level systems (ASR) led to better performance than the ones based on subtitles for both search and hyperlinking. Moreover, by analyzing the effect of multimodal re-ranking on hyperlinking performance, we emphasize the merits of rich visual information available in the anchors for the hyperlinking task, and the merits of ASR for large-scale search and hyperlinking.

This paper presents the MUST-VIS system for the MediaMixer/VideoLectures.NET Temporal Segmentation and Annotation Grand Challenge. The system allows users to visualize a lecture as a series of segments represented by keyword clouds, with relations to other similar lectures and segments. Segmentation is performed using a multi-factor algorithm which takes advantage of the audio (through automatic speech recognition and word-based segmentation) and video (through the detection of actions such as writing on the blackboard). The similarity across segments and lectures is computed using a content-based recommendation algorithm. Overall, the graph-based representation of segment similarity appears to be a promising and cost-effective approach to navigating lecture databases.

User-generated texts such as reviews, comments or discussions are valuable indicators of users' preferences. Unlike previous works which focus on labeled data from user-contributed reviews, we focus here on user comments which are not accompanied by explicit rating labels. We investigate their utility for a one-class collaborative filtering task such as bookmarking, where only the user actions are given as ground truth. We propose a sentiment-aware nearest neighbor model (SANN) for multimedia recommendations over TED talks, which makes use of user comments. The model outperforms significantly, by more than 25% on unseen data, several competitive baselines.

The Automatic Content Linking Device monitors a conversation and uses automatically recognized words to retrieve documents that are of potential use to the participants. The document set includes project related reports or emails, transcribed snippets of past meetings, and websites. Retrieval results are displayed at regular intervals.

LIDIAP

Information access in meeting recordings can be assisted by meeting browsers, or can be fully automated following a question-answering (QA) approach. An information access task is defined, aiming at discriminating true vs. false parallel statements about facts in meetings. An automatic QA algorithm is applied to this task, using passage retrieval over a meeting transcript. The algorithm scores 59% accuracy for passage retrieval, while random guessing is below 1%, but only scores 60% on combined retrieval and question discrimination, for which humans reach 70%-80% and the baseline is 50%. The algorithm clearly outperforms humans for speed, at less than 1 second per question, vs. 1.5-2 minutes per question for humans. The degradation on ASR compared to manual transcripts still yields lower but acceptable scores, especially for passage identification. Automatic QA thus appears to be a promising enhancement to meeting browsers used by humans, as an assistant for relevant passage identification.

The AMIDA Mobile Meeting Assistant is a system that allows remote participants to attend a meeting through a mobile device. The system improves the engagement in the meeting of the remote participants with respect to voice-only solutions thanks to the use of visual annotations and the capture of slides. The visual focus of attention of meeting participants and other annotations serve to reconstruct a 2D or a 3D representation of the meeting on a mobile device (smart phone). A first version of the system has been implemented, and feedback from a user study and from industrial partners shows that the Mobile Meeting Assistant's functionalities are positively appreciated, and sets priorities for future developments.

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.

In this paper, we describe a system for coreference resolution and emphasize the role of evaluation for its design. The goal of the system is to group referring expressions (identified beforehand in narrative texts) into sets of coreferring expressions that correspond to discourse entities. Several knowledge sources are distinguished, such as referential compatibility between a referring expression and a discourse entity, activation factors for discourse entities, size of working memory, or meta-rules for the creation of discourse entities. For each of them, the theoretical analysis of its relevance is compared to scores obtained through evaluation. After looping through all knowledge sources, an optimal behavior is chosen, then evaluated on test data. The paper also discusses evaluation measures as well as data annotation, and compares the present approach to others in the field.