Context-aware translation can be achieved by processing a concatenation of consecutive sentences with the standard Transformer architecture. This paper investigates the intuitive idea of providing the model with explicit information about the position of the sentences contained in the concatenation window. We compare various methods to encode sentence positions into token representations, including novel methods. Our results show that the Transformer benefits from certain sentence position encoding methods on English to Russian translation, if trained with a context-discounted loss. However, the same benefits are not observed on English to German. Further empirical efforts are necessary to define the conditions under which the proposed approach is beneficial.

Multi-encoder models are a broad family of context-aware neural machine translation systems that aim to improve translation quality by encoding document-level contextual information alongside the current sentence. The context encoding is undertaken by contextual parameters, trained on document-level data. In this work, we discuss the difficulty of training these parameters effectively, due to the sparsity of the words in need of context (i.e., the training signal), and their relevant context. We propose to pre-train the contextual parameters over split sentence pairs, which makes an efficient use of the available data for two reasons. Firstly, it increases the contextual training signal by breaking intra-sentential syntactic relations, and thus pushing the model to search the context for disambiguating clues more frequently. Secondly, it eases the retrieval of relevant context, since context segments become shorter. We propose four different splitting methods, and evaluate our approach with BLEU and contrastive test sets. Results show that it consistently improves learning of contextual parameters, both in low and high resource settings.

Most low resource language technology development is premised on the need to collect data for training statistical models. When we follow the typical process of recording and transcribing text for small Indigenous languages, we hit up against the so-called “transcription bottleneck.” Therefore it is worth exploring new ways of engaging with speakers which generate data while avoiding the transcription bottleneck. We have deployed a prototype app for speakers to use for confirming system guesses in an approach to transcription based on word spotting. However, in the process of testing the app we encountered many new problems for engagement with speakers. This paper presents a close-up study of the process of deploying data capture technology on the ground in an Australian Aboriginal community. We reflect on our interactions with participants and draw lessons that apply to anyone seeking to develop methods for language data collection in an Indigenous community.

Word and morpheme segmentation are fundamental steps of language documentation as they allow to discover lexical units in a language for which the lexicon is unknown. However, in most language documentation scenarios, linguists do not start from a blank page: they may already have a pre-existing dictionary or have initiated manual segmentation of a small part of their data. This paper studies how such a weak supervision can be taken advantage of in Bayesian non-parametric models of segmentation. Our experiments on two very low resource languages (Mboshi and Japhug), whose documentation is still in progress, show that weak supervision can be beneficial to the segmentation quality. In addition, we investigate an incremental learning scenario where manual segmentations are provided in a sequential manner. This work opens the way for interactive annotation tools for documentary linguists.

Documenting languages helps to prevent the extinction of endangered dialects - many of which are otherwise expected to disappear by the end of the century. When documenting oral languages, unsupervised word segmentation (UWS) from speech is a useful, yet challenging, task. It consists in producing time-stamps for slicing utterances into smaller segments corresponding to words, being performed from phonetic transcriptions, or in the absence of these, from the output of unsupervised speech discretization models. These discretization models are trained using raw speech only, producing discrete speech units that can be applied for downstream (text-based) tasks. In this paper we compare five of these models: three Bayesian and two neural approaches, with regards to the exploitability of the produced units for UWS. For the UWS task, we experiment with two models, using as our target language the Mboshi (Bantu C25), an unwritten language from Congo-Brazzaville. Additionally, we report results for Finnish, Hungarian, Romanian and Russian in equally low-resource settings, using only 4 hours of speech. Our results suggest that neural models for speech discretization are difficult to exploit in our setting, and that it might be necessary to adapt them to limit sequence length. We obtain our best UWS results by using Bayesian models that produce high quality, yet compressed, discrete representations of the input speech signal.

This papers aims at improving spoken language modeling (LM) using very large amount of automatically transcribed speech. We leverage the INA (French National Audiovisual Institute) collection and obtain 19GB of text after applying ASR on 350,000 hours of diverse TV shows. From this, spoken language models are trained either by fine-tuning an existing LM (FlauBERT) or through training a LM from scratch. The new models (FlauBERT-Oral) will be shared with the community and are evaluated not only in terms of word prediction accuracy but also for two downstream tasks : classification of TV shows and syntactic parsing of speech. Experimental results show that FlauBERT-Oral is better than its initial FlauBERT version demonstrating that, despite its inherent noisy nature, ASR-Generated text can be useful to improve spoken language modeling.

Recently, very large pre-trained models achieve state-of-the-art results in various natural language processing (NLP) tasks, but their size makes it more challenging to apply them in resource-constrained environments. Compression techniques allow to drastically reduce the size of the models and therefore their inference time with negligible impact on top-tier metrics. However, the general performance averaged across multiple tasks and/or languages may hide a drastic performance drop on under-represented features, which could result in the amplification of biases encoded by the models. In this work, we assess the impact of compression methods on Multilingual Neural Machine Translation models (MNMT) for various language groups, gender, and semantic biases by extensive analysis of compressed models on different machine translation benchmarks, i.e. FLORES-101, MT-Gender, and DiBiMT. We show that the performance of under-represented languages drops significantly, while the average BLEU metric only slightly decreases. Interestingly, the removal of noisy memorization with compression leads to a significant improvement for some medium-resource languages. Finally, we demonstrate that compression amplifies intrinsic gender and semantic biases, even in high-resource languages.

In recent years, multilingual machine translation models have achieved promising performance on low-resource language pairs by sharing information between similar languages, thus enabling zero-shot translation. To overcome the “curse of multilinguality”, these models often opt for scaling up the number of parameters, which makes their use in resource-constrained environments challenging. We introduce SMaLL-100, a distilled version of the M2M-100(12B) model, a massively multilingual machine translation model covering 100 languages. We train SMaLL-100 with uniform sampling across all language pairs and therefore focus on preserving the performance of low-resource languages. We evaluate SMaLL-100 on different low-resource benchmarks: FLORES-101, Tatoeba, and TICO-19 and demonstrate that it outperforms previous massively multilingual models of comparable sizes (200-600M) while improving inference latency and memory usage. Additionally, our model achieves comparable results to M2M-100 (1.2B), while being 3.6x smaller and 4.3x faster at inference.

An increasing number of papers have been addressing issues related to low-resource languages and the transcription bottleneck paradigm. After several years spent in Northern Australia, where some of the strongest Aboriginal languages are spoken, we could observe a gap between the motivations depicted in research contributions in this space and the Northern Australian context. In this paper, we address this gap in research by exploring the potential of speech recognition in an Aboriginal community. We describe our work from training a spoken term detection system to its implementation in an activity with Aboriginal participants. We report here on one side how speech recognition technologies can find their place in an Aboriginal context and, on the other, methodological paths that allowed us to reach better comprehension and engagement from Aboriginal participants.

A straightforward approach to context-aware neural machine translation consists in feeding the standard encoder-decoder architecture with a window of consecutive sentences, formed by the current sentence and a number of sentences from its context concatenated to it. In this work, we propose an improved concatenation approach that encourages the model to focus on the translation of the current sentence, discounting the loss generated by target context. We also propose an additional improvement that strengthen the notion of sentence boundaries and of relative sentence distance, facilitating model compliance to the context-discounted objective. We evaluate our approach with both average-translation quality metrics and contrastive test sets for the translation of inter-sentential discourse phenomena, proving its superiority to the vanilla concatenation approach and other sophisticated context-aware systems.

We consider the problem of multilingual unsupervised machine translation, translating to and from languages that only have monolingual data by using auxiliary parallel language pairs. For this problem the standard procedure so far to leverage the monolingual data is _back-translation_, which is computationally costly and hard to tune. In this paper we propose instead to use _denoising adapters_, adapter layers with a denoising objective, on top of pre-trained mBART-50. In addition to the modularity and flexibility of such an approach we show that the resulting translations are on-par with back-translating as measured by BLEU, and furthermore it allows adding unseen languages incrementally.

While End-2-End Text-to-Speech (TTS) has made significant progresses over the past few years, these systems still lack intuitive user controls over prosody. For instance, generating speech with fine-grained prosody control (prosodic prominence, contextually appropriate emotions) is still an open challenge. In this paper, we investigate whether we can control prosody directly from the input text, in order to code information related to contrastive focus which emphasizes a specific word that is contrary to the presuppositions of the interlocutor. We build and share a specific dataset for this purpose and show that it allows to train a TTS system were this fine-grained prosodic feature can be correctly conveyed using control tokens. Our evaluation compares synthetic and natural utterances and shows that prosodic patterns of contrastive focus (variations of Fo, Intensity and Duration) can be learnt accurately. Such a milestone is important to allow, for example, smart speakers to be programmatically controlled in terms of output prosody.

This paper presents an interactive data dashboard that provides users with an overview of the preservation of discourse relations among 28 language pairs. We display a graph network depicting the cross-lingual discourse relations between a pair of languages for multilingual TED talks and provide a search function to look for sentences with specific keywords or relation types, facilitating ease of analysis on the cross-lingual discourse relations.

We investigate the efficiency of two very different spoken term detection approaches for transcription when the available data is insufficient to train a robust speech recognition system. This work is grounded in a very low-resource language documentation scenario where only a few minutes of recording have been transcribed for a given language so far. Experiments on two oral languages show that a pretrained universal phone recognizer, fine-tuned with only a few minutes of target language speech, can be used for spoken term detection through searches in phone confusion networks with a lexicon expressed as a finite state automaton. Experimental results show that a phone recognition based approach provides better overall performances than Dynamic Time Warping when working with clean data, and highlight the benefits of each methods for two types of speech corpus.

Les modèles neuronaux de type seq2seq manifestent d’étonnantes capacités de prédiction quand ils sont entraînés sur des données de taille suffisante. Cependant, ils échouent à généraliser de manière satisfaisante quand la tâche implique d’apprendre et de réutiliser des règles systématiques de composition et non d’apprendre simplement par imitation des exemples d’entraînement. Le jeu de données SCAN, constitué d’un ensemble de commandes en langage naturel associées à des séquences d’action, a été spécifiquement conçu pour évaluer les capacités des réseaux de neurones à apprendre ce type de généralisation compositionnelle. Dans cet article, nous nous proposons d’étudier la contribution d’informations syntaxiques sur les capacités de généralisation compositionnelle des réseaux de neurones seq2seq convolutifs.

In this paper we question the impact of gender representation in training data on the performance of an end-to-end ASR system. We create an experiment based on the Librispeech corpus and build 3 different training corpora varying only the proportion of data produced by each gender category. We observe that if our system is overall robust to the gender balance or imbalance in training data, it is nonetheless dependant of the adequacy between the individuals present in the training and testing sets.

Adapter modules were recently introduced as an efficient alternative to fine-tuning in NLP. Adapter tuning consists in freezing pre-trained parameters of a model and injecting lightweight modules between layers, resulting in the addition of only a small number of task-specific trainable parameters. While adapter tuning was investigated for multilingual neural machine translation, this paper proposes a comprehensive analysis of adapters for multilingual speech translation (ST). Starting from different pre-trained models (a multilingual ST trained on parallel data or a multilingual BART (mBART) trained on non parallel multilingual data), we show that adapters can be used to: (a) efficiently specialize ST to specific language pairs with a low extra cost in terms of parameters, and (b) transfer from an automatic speech recognition (ASR) task and an mBART pre-trained model to a multilingual ST task. Experiments show that adapter tuning offer competitive results to full fine-tuning, while being much more parameter-efficient.

Language domains that require very careful use of terminology are abundant and reflect a significant part of the translation industry. In this work we introduce a benchmark for evaluating the quality and consistency of terminology translation, focusing on the medical (and COVID-19 specifically) domain for five language pairs: English to French, Chinese, Russian, and Korean, as well as Czech to German. We report the descriptions and results of the participating systems, commenting on the need for further research efforts towards both more adequate handling of terminologies as well as towards a proper formulation and evaluation of the task.

Avec l’essor de l’intelligence artificielle (IA) et l’utilisation croissante des architectures d’apprentissage profond, la question de l’éthique et de la transparence des systèmes d’IA est devenue une préoccupation centrale au sein de la communauté de recherche. Dans cet article, nous proposons une étude sur la représentation du genre dans les ressources de parole disponibles sur la plateforme Open Speech and Language Resource. Un tout premier résultat est la difficulté d’accès aux informations sur le genre des locuteurs. Ensuite, nous montrons que l’équilibre entre les catégories de genre dépend de diverses caractéristiques des corpus (discours élicité ou non, tâche adressée). En nous appuyant sur des travaux antérieurs, nous reprenons quelques principes concernant les métadonnées dans l’optique d’assurer une meilleure transparence des systèmes de parole construits à l’aide de ces corpus.

Les modèles de langue pré-entraînés sont désormais indispensables pour obtenir des résultats à l’état-de-l’art dans de nombreuses tâches du TALN. Tirant avantage de l’énorme quantité de textes bruts disponibles, ils permettent d’extraire des représentations continues des mots, contextualisées au niveau de la phrase. L’efficacité de ces représentations pour résoudre plusieurs tâches de TALN a été démontrée récemment pour l’anglais. Dans cet article, nous présentons et partageons FlauBERT, un ensemble de modèles appris sur un corpus français hétérogène et de taille importante. Des modèles de complexité différente sont entraînés à l’aide du nouveau supercalculateur Jean Zay du CNRS. Nous évaluons nos modèles de langue sur diverses tâches en français (classification de textes, paraphrase, inférence en langage naturel, analyse syntaxique, désambiguïsation automatique) et montrons qu’ils surpassent souvent les autres approches sur le référentiel d’évaluation FLUE également présenté ici.

Nous proposons une réflexion sur les pratiques d’évaluation des systèmes de reconnaissance automatique de la parole (ASR). Après avoir défini la notion de discrimination d’un point de vue légal et la notion d’équité dans les systèmes d’intelligence artificielle, nous nous intéressons aux pratiques actuelles lors des grandes campagnes d’évaluation. Nous observons que la variabilité de la parole et plus particulièrement celle de l’individu n’est pas prise en compte dans les protocoles d’évaluation actuels rendant impossible l’étude de biais potentiels dans les systèmes.

Language models have become a key step to achieve state-of-the art results in many different Natural Language Processing (NLP) tasks. Leveraging the huge amount of unlabeled texts nowadays available, they provide an efficient way to pre-train continuous word representations that can be fine-tuned for a downstream task, along with their contextualization at the sentence level. This has been widely demonstrated for English using contextualized representations (Dai and Le, 2015; Peters et al., 2018; Howard and Ruder, 2018; Radford et al., 2018; Devlin et al., 2019; Yang et al., 2019b). In this paper, we introduce and share FlauBERT, a model learned on a very large and heterogeneous French corpus. Models of different sizes are trained using the new CNRS (French National Centre for Scientific Research) Jean Zay supercomputer. We apply our French language models to diverse NLP tasks (text classification, paraphrasing, natural language inference, parsing, word sense disambiguation) and show that most of the time they outperform other pre-training approaches. Different versions of FlauBERT as well as a unified evaluation protocol for the downstream tasks, called FLUE (French Language Understanding Evaluation), are shared to the research community for further reproducible experiments in French NLP.

The CMU Wilderness Multilingual Speech Dataset (Black, 2019) is a newly published multilingual speech dataset based on recorded readings of the New Testament. It provides data to build Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) models for potentially 700 languages. However, the fact that the source content (the Bible) is the same for all the languages is not exploited to date. Therefore, this article proposes to add multilingual links between speech segments in different languages, and shares a large and clean dataset of 8,130 parallel spoken utterances across 8 languages (56 language pairs). We name this corpus MaSS (Multilingual corpus of Sentence-aligned Spoken utterances). The covered languages (Basque, English, Finnish, French, Hungarian, Romanian, Russian and Spanish) allow researches on speech-to-speech alignment as well as on translation for typologically different language pairs. The quality of the final corpus is attested by human evaluation performed on a corpus subset (100 utterances, 8 language pairs). Lastly, we showcase the usefulness of the final product on a bilingual speech retrieval task.

With the rise of artificial intelligence (AI) and the growing use of deep-learning architectures, the question of ethics, transparency and fairness of AI systems has become a central concern within the research community. We address transparency and fairness in spoken language systems by proposing a study about gender representation in speech resources available through the Open Speech and Language Resource platform. We show that finding gender information in open source corpora is not straightforward and that gender balance depends on other corpus characteristics (elicited/non elicited speech, low/high resource language, speech task targeted). The paper ends with recommendations about metadata and gender information for researchers in order to assure better transparency of the speech systems built using such corpora.

We propose a novel transcription workflow which combines spoken term detection and human-in-the-loop, together with a pilot experiment. This work is grounded in an almost zero-resource scenario where only a few terms have so far been identified, involving two endangered languages. We show that in the early stages of transcription, when the available data is insufficient to train a robust ASR system, it is possible to take advantage of the transcription of a small number of isolated words in order to bootstrap the transcription of a speech collection.

We introduce dual-decoder Transformer, a new model architecture that jointly performs automatic speech recognition (ASR) and multilingual speech translation (ST). Our models are based on the original Transformer architecture (Vaswani et al., 2017) but consist of two decoders, each responsible for one task (ASR or ST). Our major contribution lies in how these decoders interact with each other: one decoder can attend to different information sources from the other via a dual-attention mechanism. We propose two variants of these architectures corresponding to two different levels of dependencies between the decoders, called the parallel and cross dual-decoder Transformers, respectively. Extensive experiments on the MuST-C dataset show that our models outperform the previously-reported highest translation performance in the multilingual settings, and outperform as well bilingual one-to-one results. Furthermore, our parallel models demonstrate no trade-off between ASR and ST compared to the vanilla multi-task architecture. Our code and pre-trained models are available at https://github.com/formiel/speech-translation.

We conduct in this work an evaluation study comparing offline and online neural machine translation architectures. Two sequence-to-sequence models: convolutional Pervasive Attention (Elbayad et al. 2018) and attention-based Transformer (Vaswani et al. 2017) are considered. We investigate, for both architectures, the impact of online decoding constraints on the translation quality through a carefully designed human evaluation on English-German and German-English language pairs, the latter being particularly sensitive to latency constraints. The evaluation results allow us to identify the strengths and shortcomings of each model when we shift to the online setup.

We propose a novel adapter layer formalism for adapting multilingual models. They are more parameter-efficient than existing adapter layers while obtaining as good or better performance. The layers are specific to one language (as opposed to bilingual adapters) allowing to compose them and generalize to unseen language-pairs. In this zero-shot setting, they obtain a median improvement of +2.77 BLEU points over a strong 20-language multilingual Transformer baseline trained on TED talks.

The language acquisition literature shows that children do not build their lexicon by segmenting the spoken input into phonemes and then building up words from them, but rather adopt a top-down approach and start by segmenting word-like units and then break them down into smaller units. This suggests that the ideal way of learning a language is by starting from full semantic units. In this paper, we investigate if this is also the case for a neural model of Visually Grounded Speech trained on a speech-image retrieval task. We evaluated how well such a network is able to learn a reliable speech-to-image mapping when provided with phone, syllable, or word boundary information. We present a simple way to introduce such information into an RNN-based model and investigate which type of boundary is the most efficient. We also explore at which level of the network’s architecture such information should be introduced so as to maximise its performances. Finally, we show that using multiple boundary types at once in a hierarchical structure, by which low-level segments are used to recompose high-level segments, is beneficial and yields better results than using low-level or high-level segments in isolation.

This paper describes the ON-TRAC Consortium translation systems developed for two challenge tracks featured in the Evaluation Campaign of IWSLT 2020, offline speech translation and simultaneous speech translation. ON-TRAC Consortium is composed of researchers from three French academic laboratories: LIA (Avignon Université), LIG (Université Grenoble Alpes), and LIUM (Le Mans Université). Attention-based encoder-decoder models, trained end-to-end, were used for our submissions to the offline speech translation track. Our contributions focused on data augmentation and ensembling of multiple models. In the simultaneous speech translation track, we build on Transformer-based wait-k models for the text-to-text subtask. For speech-to-text simultaneous translation, we attach a wait-k MT system to a hybrid ASR system. We propose an algorithm to control the latency of the ASR+MT cascade and achieve a good latency-quality trade-off on both subtasks.

For endangered languages, data collection campaigns have to accommodate the challenge that many of them are from oral tradition, and producing transcriptions is costly. Therefore, it is fundamental to translate them into a widely spoken language to ensure interpretability of the recordings. In this paper we investigate how the choice of translation language affects the posterior documentation work and potential automatic approaches which will work on top of the produced bilingual corpus. For answering this question, we use the MaSS multilingual speech corpus (Boito et al., 2020) for creating 56 bilingual pairs that we apply to the task of low-resource unsupervised word segmentation and alignment. Our results highlight that the choice of language for translation influences the word segmentation performance, and that different lexicons are learned by using different aligned translations. Lastly, this paper proposes a hybrid approach for bilingual word segmentation, combining boundary clues extracted from a non-parametric Bayesian model (Goldwater et al., 2009a) with the attentional word segmentation neural model from Godard et al. (2018). Our results suggest that incorporating these clues into the neural models’ input representation increases their translation and alignment quality, specially for challenging language pairs.

In this paper, we study how word-like units are represented and activated in a recurrent neural model of visually grounded speech. The model used in our experiments is trained to project an image and its spoken description in a common representation space. We show that a recurrent model trained on spoken sentences implicitly segments its input into word-like units and reliably maps them to their correct visual referents. We introduce a methodology originating from linguistics to analyse the representation learned by neural networks – the gating paradigm – and show that the correct representation of a word is only activated if the network has access to first phoneme of the target word, suggesting that the network does not rely on a global acoustic pattern. Furthermore, we find out that not all speech frames (MFCC vectors in our case) play an equal role in the final encoded representation of a given word, but that some frames have a crucial effect on it. Finally we suggest that word representation could be activated through a process of lexical competition.

Recently, neural models led to significant improvements in both machine translation (MT) and natural language generation tasks (NLG). However, generation of long descriptive summaries conditioned on structured data remains an open challenge. Likewise, MT that goes beyond sentence-level context is still an open issue (e.g., document-level MT or MT with metadata). To address these challenges, we propose to leverage data from both tasks and do transfer learning between MT, NLG, and MT with source-side metadata (MT+NLG). First, we train document-based MT systems with large amounts of parallel data. Then, we adapt these models to pure NLG and MT+NLG tasks by fine-tuning with smaller amounts of domain-specific data. This end-to-end NLG approach, without data selection and planning, outperforms the previous state of the art on the Rotowire NLG task. We participated to the “Document Generation and Translation” task at WNGT 2019, and ranked first in all tracks.

In this paper, we present our submission for the English to Czech Text Translation Task of IWSLT 2019. Our system aims to study how pre-trained language models, used as input embeddings, can improve a specialized machine translation system trained on few data. Therefore, we implemented a Transformer-based encoder-decoder neural system which is able to use the output of a pre-trained language model as input embeddings, and we compared its performance under three configurations: 1) without any pre-trained language model (constrained), 2) using a language model trained on the monolingual parts of the allowed English-Czech data (constrained), and 3) using a language model trained on a large quantity of external monolingual data (unconstrained). We used BERT as external pre-trained language model (configuration 3), and BERT architecture for training our own language model (configuration 2). Regarding the training data, we trained our MT system on a small quantity of parallel text: one set only consists of the provided MuST-C corpus, and the other set consists of the MuST-C corpus and the News Commentary corpus from WMT. We observed that using the external pre-trained BERT improves the scores of our system by +0.8 to +1.5 of BLEU on our development set, and +0.97 to +1.94 of BLEU on the test set. However, using our own language model trained only on the allowed parallel data seems to improve the machine translation performances only when the system is trained on the smallest dataset.

One of the basic tasks of computational language documentation (CLD) is to identify word boundaries in an unsegmented phonemic stream. While several unsupervised monolingual word segmentation algorithms exist in the literature, they are challenged in real-world CLD settings by the small amount of available data. A possible remedy is to take advantage of glosses or translation in a foreign, well- resourced, language, which often exist for such data. In this paper, we explore and compare ways to exploit neural machine translation models to perform unsupervised boundary detection with bilingual information, notably introducing a new loss function for jointly learning alignment and segmentation. We experiment with an actual under-resourced language, Mboshi, and show that these techniques can effectively control the output segmentation length.

Current state-of-the-art machine translation systems are based on encoder-decoder architectures, that first encode the input sequence, and then generate an output sequence based on the input encoding. Both are interfaced with an attention mechanism that recombines a fixed encoding of the source tokens based on the decoder state. We propose an alternative approach which instead relies on a single 2D convolutional neural network across both sequences. Each layer of our network re-codes source tokens on the basis of the output sequence produced so far. Attention-like properties are therefore pervasive throughout the network. Our model yields excellent results, outperforming state-of-the-art encoder-decoder systems, while being conceptually simpler and having fewer parameters.

Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from ’exposure bias’: during training tokens are predicted given ground-truth sequences, while at test time prediction is conditioned on generated output sequences. To overcome these limitations we build upon the recent reward augmented maximum likelihood approach that encourages the model to predict sentences that are close to the ground truth according to a given performance metric. We extend this approach to token-level loss smoothing, and propose improvements to the sequence-level smoothing approach. Our experiments on two different tasks, image captioning and machine translation, show that token-level and sequence-level loss smoothing are complementary, and significantly improve results.

This paper addresses a relatively new task: prediction of ASR performance on unseen broadcast programs. In a previous paper, we presented an ASR performance prediction system using CNNs that encode both text (ASR transcript) and speech, in order to predict word error rate. This work is dedicated to the analysis of speech signal embeddings and text embeddings learnt by the CNN while training our prediction model. We try to better understand which information is captured by the deep model and its relation with different conditioning factors. It is shown that hidden layers convey a clear signal about speech style, accent and broadcast type. We then try to leverage these 3 types of information at training time through multi-task learning. Our experiments show that this allows to train slightly more efficient ASR performance prediction systems that - in addition - simultaneously tag the analyzed utterances according to their speech style, accent and broadcast program origin.

Computational Language Documentation attempts to make the most recent research in speech and language technologies available to linguists working on language preservation and documentation. In this paper, we pursue two main goals along these lines. The first is to improve upon a strong baseline for the unsupervised word discovery task on two very low-resource Bantu languages, taking advantage of the expertise of linguists on these particular languages. The second consists in exploring the Adaptor Grammar framework as a decision and prediction tool for linguists studying a new language. We experiment 162 grammar configurations for each language and show that using Adaptor Grammars for word segmentation enables us to test hypotheses about a language. Specializing a generic grammar with language specific knowledge leads to great improvements for the word discovery task, ultimately achieving a leap of about 30% token F-score from the results of a strong baseline.

In spite of the recent success of Dialogue Act (DA) classification, the majority of prior works focus on text-based classification with oracle transcriptions, i.e. human transcriptions, instead of Automatic Speech Recognition (ASR)’s transcriptions. In spoken dialog systems, however, the agent would only have access to noisy ASR transcriptions, which may further suffer performance degradation due to domain shift. In this paper, we explore the effectiveness of using both acoustic and textual signals, either oracle or ASR transcriptions, and investigate speaker domain adaptation for DA classification. Our multimodal model proves to be superior to the unimodal models, particularly when the oracle transcriptions are not available. We also propose an effective method for speaker domain adaptation, which achieves competitive results.

This paper is a deep investigation of cross-language plagiarism detection methods on a new recently introduced open dataset, which contains parallel and comparable collections of documents with multiple characteristics (different genres, languages and sizes of texts). We investigate cross-language plagiarism detection methods for 6 language pairs on 2 granularities of text units in order to draw robust conclusions on the best methods while deeply analyzing correlations across document styles and languages.

This paper describes speech translation from Amharic-to-English, particularly Automatic Speech Recognition (ASR) with post-editing feature and Amharic-English Statistical Machine Translation (SMT). ASR experiment is conducted using morpheme language model (LM) and phoneme acoustic model(AM). Likewise,SMT conducted using word and morpheme as unit. Morpheme based translation shows a 6.29 BLEU score at a 76.4% of recognition accuracy while word based translation shows a 12.83 BLEU score using 77.4% word recognition accuracy. Further, after post-edit on Amharic ASR using corpus based n-gram, the word recognition accuracy increased by 1.42%. Since post-edit approach reduces error propagation, the word based translation accuracy improved by 0.25 (1.95%) BLEU score. We are now working towards further improving propagated errors through different algorithms at each unit of speech translation cascading component.

This paper proposes to use distributed representation of words (word embeddings) in cross-language textual similarity detection. The main contributions of this paper are the following: (a) we introduce new cross-language similarity detection methods based on distributed representation of words; (b) we combine the different methods proposed to verify their complementarity and finally obtain an overall F1 score of 89.15% for English-French similarity detection at chunk level (88.5% at sentence level) on a very challenging corpus.

We present our submitted systems for Semantic Textual Similarity (STS) Track 4 at SemEval-2017. Given a pair of Spanish-English sentences, each system must estimate their semantic similarity by a score between 0 and 5. In our submission, we use syntax-based, dictionary-based, context-based, and MT-based methods. We also combine these methods in unsupervised and supervised way. Our best run ranked 1st on track 4a with a correlation of 83.02% with human annotations.

Cet article présente un système original de traduction de documents numérisés en arabe. Deux modules sont cascadés : un système de reconnaissance optique de caractères (OCR) en arabe et un système de traduction automatique (TA) arabe-français. Le couplage OCR-TA a été peu abordé dans la littérature et l’originalité de cette étude consiste à proposer un couplage étroit entre OCR et TA ainsi qu’un traitement spécifique des mots hors vocabulaire (MHV) engendrés par les erreurs d’OCRisation. Le couplage OCR-TA par treillis et notre traitement des MHV par remplacement selon une mesure composite qui prend en compte forme de surface et contexte du mot, permettent une amélioration significative des performances de traduction. Les expérimentations sont réalisés sur un corpus de journaux numérisés en arabe et permettent d’obtenir des améliorations en score BLEU de 3,73 et 5,5 sur les corpus de développement et de test respectivement.

Nos travaux portent sur la construction rapide d’outils d’analyse linguistique pour des langues peu dotées en ressources. Dans une précédente contribution, nous avons proposé une méthode pour la construction automatique d’un analyseur morpho-syntaxique via une projection interlingue d’annotations linguistiques à partir de corpus parallèles (méthode fondée sur les réseaux de neurones récurrents). Nous présentons, dans cet article, une amélioration de notre modèle neuronal, avec la prise en compte d’informations linguistiques externes pour un annotateur plus complexe. En particulier, nous proposons d’intégrer des annotations morpho-syntaxiques dans notre architecture neuronale pour l’apprentissage non supervisé d’annotateurs sémantiques multilingues à gros grain (annotation en SuperSenses). Nous montrons la validité de notre méthode et sa généricité sur l’italien et le français et étudions aussi l’impact de la qualité du corpus parallèle sur notre approche (généré par traduction manuelle ou automatique). Nos expériences portent sur la projection d’annotations de l’anglais vers le français et l’italien.

Cet article présente une approche associant réseaux lexico-sémantiques et représentations distribuées de mots appliquée à l’évaluation de la traduction automatique. Cette étude est faite à travers l’enrichissement d’une métrique bien connue pour évaluer la traduction automatique (TA) : METEOR. METEOR permet un appariement approché (similarité morphologique ou synonymie) entre une sortie de système automatique et une traduction de référence. Nos expérimentations s’appuient sur la tâche Metrics de la campagne d’évaluation WMT 2014 et montrent que les représentations distribuées restent moins performantes que les ressources lexico-sémantiques pour l’évaluation en TA mais peuvent néammoins apporter un complément d’information intéressant à ces dernières.

This paper aims to unravel the automatic quality assessment for spoken language translation (SLT). More precisely, we propose several effective estimators based on our estimation of transcription (ASR) quality, translation (MT) quality, or both (combined and joint features using ASR and MT information). Our experiments provide an important opportunity to advance the understanding of the prediction quality of words in a SLT output that were revealed by MT and ASR features. These results could be applied to interactive speech translation or computer-assisted translation of speeches and lectures. For reproducible experiments, the code allowing to call our WCE-LIG application and the corpora used are made available to the research community.

In this paper, we describe the organization and the implementation of the CAMOMILE collaborative annotation framework for multimodal, multimedia, multilingual (3M) data. Given the versatile nature of the analysis which can be performed on 3M data, the structure of the server was kept intentionally simple in order to preserve its genericity, relying on standard Web technologies. Layers of annotations, defined as data associated to a media fragment from the corpus, are stored in a database and can be managed through standard interfaces with authentication. Interfaces tailored specifically to the needed task can then be developed in an agile way, relying on simple but reliable services for the management of the centralized annotations. We then present our implementation of an active learning scenario for person annotation in video, relying on the CAMOMILE server; during a dry run experiment, the manual annotation of 716 speech segments was thus propagated to 3504 labeled tracks. The code of the CAMOMILE framework is distributed in open source.

This article presents the data collected and ASR systems developped for 4 sub-saharan african languages (Swahili, Hausa, Amharic and Wolof). To illustrate our methodology, the focus is made on Wolof (a very under-resourced language) for which we designed the first ASR system ever built in this language. All data and scripts are available online on our github repository.

In this paper we describe our effort to create a dataset for the evaluation of cross-language textual similarity detection. We present preexisting corpora and their limits and we explain the various gathered resources to overcome these limits and build our enriched dataset. The proposed dataset is multilingual, includes cross-language alignment for different granularities (from chunk to document), is based on both parallel and comparable corpora and contains human and machine translated texts. Moreover, it includes texts written by multiple types of authors (from average to professionals). With the obtained dataset, we conduct a systematic and rigorous evaluation of several state-of-the-art cross-language textual similarity detection methods. The evaluation results are reviewed and discussed. Finally, dataset and scripts are made publicly available on GitHub: http://github.com/FerreroJeremy/Cross-Language-Dataset.

We present MultiVec, a new toolkit for computing continuous representations for text at different granularity levels (word-level or sequences of words). MultiVec includes word2vec’s features, paragraph vector (batch and online) and bivec for bilingual distributed representations. MultiVec also includes different distance measures between words and sequences of words. The toolkit is written in C++ and is aimed at being fast (in the same order of magnitude as word2vec), easy to use, and easy to extend. It has been evaluated on several NLP tasks: the analogical reasoning task, sentiment analysis, and crosslingual document classification.

This work focuses on the development of linguistic analysis tools for resource-poor languages. We use a parallel corpus to produce a multilingual word representation based only on sentence level alignment. This representation is combined with the annotated source side (resource-rich language) of the parallel corpus to train text analysis tools for resource-poor languages. Our approach is based on Recurrent Neural Networks (RNN) and has the following advantages: (a) it does not use word alignment information, (b) it does not assume any knowledge about foreign languages, which makes it applicable to a wide range of resource-poor languages, (c) it provides truly multilingual taggers. In a previous study, we proposed a method based on Simple RNN to automatically induce a Part-Of-Speech (POS) tagger. In this paper, we propose an improvement of our neural model. We investigate the Bidirectional RNN and the inclusion of external information (for instance low level information from Part-Of-Speech tags) in the RNN to train a more complex tagger (for instance, a multilingual super sense tagger). We demonstrate the validity and genericity of our method by using parallel corpora (obtained by manual or automatic translation). Our experiments are conducted to induce cross-lingual POS and super sense taggers.

This paper presents an approach combining lexico-semantic resources and distributed representations of words applied to the evaluation in machine translation (MT). This study is made through the enrichment of a well-known MT evaluation metric: METEOR. METEOR enables an approximate match (synonymy or morphological similarity) between an automatic and a reference translation. Our experiments are made in the framework of the Metrics task of WMT 2014. We show that distributed representations are a good alternative to lexico-semanticresources for MT evaluation and they can even bring interesting additional information. The augmented versions of METEOR, using vector representations, are made available on our Github page.

Les mesures de confiance au niveau mot (Word Confidence Estimation - WCE) pour la traduction auto- matique (TA) ou pour la reconnaissance automatique de la parole (RAP) attribuent un score de confiance à chaque mot dans une hypothèse de transcription ou de traduction. Dans le passé, l’estimation de ces mesures a le plus souvent été traitée séparément dans des contextes RAP ou TA. Nous proposons ici une estimation conjointe de la confiance associée à un mot dans une hypothèse de traduction automatique de la parole (TAP). Cette estimation fait appel à des paramètres issus aussi bien des systèmes de transcription de la parole (RAP) que des systèmes de traduction automatique (TA). En plus de la construction de ces estimateurs de confiance robustes pour la TAP, nous utilisons les informations de confiance pour re-décoder nos graphes d’hypothèses de traduction. Les expérimentations réalisées montrent que l’utilisation de ces mesures de confiance au cours d’une seconde passe de décodage permettent d’obtenir une amélioration significative des performances de traduction (évaluées avec la métrique BLEU - gains de deux points par rapport à notre système de traduc- tion de parole de référence). Ces expériences sont faites pour une tâche de TAP (français-anglais) pour laquelle un corpus a été spécialement conçu (ce corpus, mis à la disposition de la communauté TALN, est aussi décrit en détail dans l’article).

La construction d’outils d’analyse linguistique pour les langues faiblement dotées est limitée, entre autres, par le manque de corpus annotés. Dans cet article, nous proposons une méthode pour construire automatiquement des outils d’analyse via une projection interlingue d’annotations linguistiques en utilisant des corpus parallèles. Notre approche n’utilise pas d’autres sources d’information, ce qui la rend applicable à un large éventail de langues peu dotées. Nous proposons d’utiliser les réseaux de neurones récurrents pour projeter les annotations d’une langue à une autre (sans utiliser d’information d’alignement des mots). Dans un premier temps, nous explorons la tâche d’annotation morpho-syntaxique. Notre méthode combinée avec une méthode de projection d’annotation basique (utilisant l’alignement mot à mot), donne des résultats comparables à ceux de l’état de l’art sur une tâche similaire.

Data selection is a common technique for adapting statistical translation models for a specific domain, which has been shown to both improve translation quality and to reduce model size. Selection relies on some in-domain data, of the same domain of the texts expected to be translated. Selecting the sentence-pairs that are most similar to the in-domain data from a pool of parallel texts has been shown to be effective; yet, this approach holds the risk of resulting in a limited coverage, when necessary n-grams that do appear in the pool are less similar to in-domain data that is available in advance. Some methods select additional data based on the actual text that needs to be translated. While useful, this is not always a practical scenario. In this work we describe an extensive exploration of data selection techniques over Arabic to French datasets, and propose methods to address both similarity and coverage considerations while maintaining a limited model size.

Corpus-based approaches to machine translation (MT) rely on the availability of parallel corpora. To produce user-acceptable translation outputs, such systems need high quality data to be efficiency trained, optimized and evaluated. However, building high quality dataset is a relatively expensive task. In this paper, we describe the data collection and analysis of a large database of 10.881 SMT translation output hypotheses manually corrected. These post-editions were collected using Amazon's Mechanical Turk, following some ethical guidelines. A complete analysis of the collected data pointed out a high quality of the corrections with more than 87 % of the collected post-editions that improve hypotheses and more than 94 % of the crowdsourced post-editions which are at least of professional quality. We also post-edited 1,500 gold-standard reference translations (of bilingual parallel corpora generated by professional) and noticed that 72 % of these translations needed to be corrected during post-edition. We computed a proximity measure between the differents kind of translations and pointed out that reference translations are as far from the hypotheses than from the corrected hypotheses (i.e. the post-editions). In light of these last findings, we discuss the adequation of text-based generated reference translations to train setence-to-sentence based SMT systems.

The PORTMEDIA project is intended to develop new corpora for the evaluation of spoken language understanding systems. The newly collected data are in the field of human-machine dialogue systems for tourist information in French in line with the MEDIA corpus. Transcriptions and semantic annotations, obtained by low-cost procedures, are provided to allow a thorough evaluation of the systems' capabilities in terms of robustness and portability across languages and domains. A new test set with some adaptation data is prepared for each case: in Italian as an example of a new language, for ticket reservation as an example of a new domain. Finally the work is complemented by the proposition of a new high level semantic annotation scheme well-suited to dialogue data.

This paper presents the LIG participation to the E-F MT task of IWSLT 2012. The primary system proposed made a large improvement (more than 3 point of BLEU on tst2010 set) compared to our last year participation. Part of this improvment was due to the use of an extraction from the Gigaword corpus. We also propose a preliminary adaptation of the driven decoding concept for machine translation. This method allows an efficient combination of machine translation systems, by rescoring the log-linear model at the N-best list level according to auxiliary systems: the basis technique is essentially guiding the search using one or previous system outputs. The results show that the approach allows a significant improvement in BLEU score using Google translate to guide our own SMT system. We also try to use a confidence measure as an additional log-linear feature but we could not get any improvment with this technique.

We describe several experiments to better understand the usefulness of statistical post-edition (SPE) to improve phrase-based statistical MT (PBMT) systems raw outputs. Whatever the size of the training corpus, we show that SPE systems trained on general domain data offers no breakthrough to our baseline general domain PBMT system. However, using manually post-edited system outputs to train the SPE led to a slight improvement in the translations quality compared with the use of professional reference translations. We also show that SPE is far more effective for domain adaptation, mainly because it recovers a lot of specific terms unknown to our general PBMT system. Finally, we compare two domain adaptation techniques, post-editing a general domain PBMT system vs building a new domain-adapted PBMT system with two different techniques, and show that the latter outperforms the first one. Yet, when the PBMT is a “black box”, SPE trained with post-edited system outputs remains an interesting option for domain adaptation.

This paper describes the system developed by the LIG laboratory for the 2011 IWSLT evaluation. We participated to the English-French MT and SLT tasks. The development of a reference translation system (MT task), as well as an ASR output translation system (SLT task) are presented. We focus this year on the SLT task and on the use of multiple 1-best ASR outputs to improve overall translation quality. The main experiment presented here compares the performance of a SLT system where multiple ASR 1-best are combined before translation (source combination), with a SLT system where multiple ASR 1-best are translated, the system combination being conducted afterwards on the target side (target combination). The experimental results show that the second approach (target combination) overpasses the first one, when the performance is measured with BLEU.

Dans cet article, nous proposons plusieurs approches pour la portabilité du module de compréhension de la parole (SLU) d’un système de dialogue d’une langue vers une autre. On montre que l’utilisation des traductions automatiques statistiques (SMT) aide à réduire le temps et le cout de la portabilité d’un tel système d’une langue source vers une langue cible. Pour la tache d’étiquetage sémantique on propose d’utiliser soit les champs aléatoires conditionnels (CRF), soit l’approche à base de séquences (PH-SMT). Les résultats expérimentaux montrent l’efficacité des méthodes proposées pour une portabilité rapide du SLU vers une nouvelle langue. On propose aussi deux méthodes pour accroître la robustesse du SLU aux erreurs de traduction. Enfin on montre que la combinaison de ces approches réduit les erreurs du système. Ces travaux sont motivés par la disponibilité du corpus MEDIA français et de la traduction manuelle vers l’italien d’une sous partie de ce corpus.

Cet article présente une méthode non-supervisée pour extraire des paires de phrases parallèles à partir d’un corpus comparable. Un système de traduction automatique est utilisé pour exploiter le corpus comparable et détecter les paires de phrases parallèles. Un processus itératif est exécuté non seulement pour augmenter le nombre de paires de phrases parallèles extraites, mais aussi pour améliorer la qualité globale du système de traduction. Une comparaison avec une méthode semi-supervisée est présentée également. Les expériences montrent que la méthode non-supervisée peut être réellement appliquée dans le cas où on manque de données parallèles. Bien que les expériences préliminaires soient menées sur la traduction français-anglais, cette méthode non-supervisée est également appliquée avec succès à un couple de langues peu doté : vietnamien-français.

General purpose, high quality and fully automatic MT is believed to be impossible. We are interested in scriptural translation problems, which are weak sub-problems of the general problem of translation. We introduce the characteristics of the weak problems of translation and of the scriptural translation problems, describe different computational approaches (finite-state, statistical and hybrid) to solve these problems, and report our results on several combinations of Indo-Pak languages and writing systems.

In this work, automatic recognition of Arabic dialects is proposed. An acoustic survey of the proportion of vocalic intervals and the standard deviation of consonantal intervals in nine dialects (Tunisia, Morocco, Algeria, Egypt, Syria, Lebanon, Yemen, Golf’s Countries and Iraq) is performed using the platform Alize and Gaussian Mixture Models (GMM). The results show the complexity of the automatic identification of Arabic dialects since. No clear border can be found between the dialects, but a gradual transition between them. They can even vary slightly from one city to another. The existence of this gradual change is easy to understand: it corresponds to a human and social reality, to the contact, friendships forged and affinity in the environment more or less immediate of the individual. This document also raises questions about the classes or macro classes of Arabic dialects noticed from the confusion matrix and the design of the hierarchical tree obtained.

This paper describes the LIG experiments in the context of IWSLT09 evaluation (Arabic to English Statistical Machine Translation task). Arabic is a morphologically rich language, and recent experimentations in our laboratory have shown that the performance of Arabic to English SMT systems varies greatly according to the Arabic morphological segmenters applied. Based on this observation, we propose to use simultaneously multiple segmentations for machine translation of Arabic. The core idea is to keep the ambiguity of the Arabic segmentation in the system input (using confusion networks or lattices). Then, we hope that the best segmentation will be chosen during MT decoding. The mathematics of this multiple segmentation approach are given. Practical implementations in the case of verbatim text translation as well as speech translation (outside of the scope of IWSLT09 this year) are proposed. Experiments conducted in the framework of IWSLT evaluation campaign show the potential of the multiple segmentation approach. The last part of this paper explains in detail the different systems submitted by LIG at IWSLT09 and the results obtained.

Cet article présente nos premiers travaux en vue de la construction d’un système de traduction probabiliste pour le couple de langue vietnamien-français. La langue vietnamienne étant considérée comme une langue peu dotée, une des difficultés réside dans la constitution des corpus parallèles, indispensable à l’apprentissage des modèles. Nous nous concentrons sur la constitution d’un grand corpus parallèle vietnamien-français. La méthode d’identification automatique des paires de documents parallèles fondée sur la date de publication, les mots spéciaux et les scores d’alignements des phrases est appliquée. Cet article présente également la construction d’un premier système de traduction automatique probabiliste vietnamienfrançais et français-vietnamien à partir de ce corpus et discute l’opportunité d’utiliser des unités lexicales ou sous-lexicales pour le vietnamien (syllabes, mots, ou leurs combinaisons). Les performances du système sont encourageantes et se comparent avantageusement à celles du système de Google.

Dans cet article, nous traitons du problème de la modélisation statistique du langage pour les langues peu dotées et sans segmentation entre les mots. Tandis que le manque de données textuelles a un impact sur la performance des modèles, les erreurs introduites par la segmentation automatique peuvent rendre ces données encore moins exploitables. Pour exploiter au mieux les données textuelles, nous proposons une méthode qui effectue des segmentations multiples sur le corpus d’apprentissage au lieu d’une segmentation unique. Cette méthode basée sur les automates d’état finis permet de retrouver les n-grammes non trouvés par la segmentation unique et de générer des nouveaux n-grammes pour l’apprentissage de modèle du langage. L’application de cette approche pour l’apprentissage des modèles de langage pour les systèmes de reconnaissance automatique de la parole en langue khmère et vietnamienne s’est montrée plus performante que la méthode par segmentation unique, à base de règles.

In this paper we present an overview on the development of a large vocabulary continuous speech recognition (LVCSR) system for Khmer, the official language of Cambodia, spoken by more than 15 million people. As an under-resourced language, develop a LVCSR system for Khmer is a challenging task. We describe our methodologies for quick language data collection and processing for language modeling and acoustic modeling. For language modeling, we investigate the use of word and sub-word as basic modeling unit in order to see the potential of sub-word units in the case of unsegmented language like Khmer. Grapheme-based acoustic modeling is used to quickly build our Khmer language acoustic model. Furthermore, the approaches and tools used for the development of our system are documented and made publicly available on the web. We hope this will contribute to accelerate the development of LVCSR system for a new language, especially for under-resource languages of developing countries where resources and expertise are limited.

This paper is a description of the system presented by the LIG laboratory to the IWSLT07 speech translation evaluation. The LIG participated, for the first time this year, in the Arabic to English speech translation task. For translation, we used a conventional statistical phrase-based system developed using the moses open source decoder. Our baseline MT system is described and we discuss particularly the use of an additional bilingual dictionary which seems useful when few training data is available. The main contribution of this paper concerns the proposal of a lattice decomposition algorithm that allows transforming a word lattice into a sub word lattice compatible with our MT model that uses word segmentation on the Arabic part. The lattice is then transformed into a confusion network which can be directly decoded into moses. The results show that this method outperforms the conventional 1-best translation which consists in translating only the most probable ASR hypothesis. The best BLEU score, from ASR output obtained on IWSLT06 evaluation data is 0.2253. The results confirm the interest of full CN decoding for speech translation, compared to traditional ASR 1-best approach. Our primary system was ranked 7/14 for IWSLT07 AE ASR task with a BLEU score of 0.3804.

Automatic speech recognition (ASR) technology has achieved a level of maturity, where it is already practical to be used by novice users. However, most non-native speakers are still not comfortable with services including ASR systems, because of the accuracy on non-native speakers. This paper describes our approach in constructing a non-native corpus particularly in French for testing and adapting non-native speaker for automatic speech recognition. Finally, we also propose in this paper a method for detecting pronunciation variants and possible pronunciation mistakes by non-native speakers.

Dans cet article, nous détaillons les résultats de la seconde évaluation du projet européen NESPOLE! auquel nous avons pris part pour le français. Dans ce projet, ainsi que dans ceux qui l’ont précédé, des techniques d’évaluation subjectives — réalisées par des évaluateurs humains — ont été mises en oeuvre. Nous présentons aussi les nouvelles techniques objectives — automatiques — proposées en traduction de l’écrit et mises en oeuvre dans le projet C-STAR III. Nous conclurons en proposant quelques idées et perspectives pour le domaine.

Le travail présenté dans cet article a été réalisé dans le cadre d’un projet global de traduction automatique de la parole. L’approche de traduction est fondée sur un langage pivot ou Interchange Format (IF), qui représente le sens de la phrase indépendamment de la langue. Nous proposons une méthode qui intègre des informations sémantiques dans le modèle statistique de langage du système de Reconnaissance Automatique de Parole. Le principe consiste a utiliser certaines classes définies dans l’IF comme des classes sémantiques dans le modèle de langage. Ceci permet au système de reconnaissance de la parole d’analyser partiellement en IF les tours de parole. Les expérimentations realisées montrent qu’avec cette approche, le système de reconnaissance peut analyser directement en IF une partie des données de dialogues de notre application, sans faire appel au système de traduction (35% des mots ; 58% des tours de parole), tout en maintenant le même niveau de performance du système global.