La détection d’intention et de concepts sont des tâches essentielles de la compréhension de la parole(SLU). Or il n’existe que peu de données annotées en français permettant d’effectuer ces deux tâches conjointement. Cependant, il existe des ensembles de données annotées en concept, dont le corpus MEDIA. Ce corpus est considéré comme l’un des plus difficiles. Néanmoins, il ne comporte que des annotations en concepts et pas en intentions. Dans cet article, nous proposons une version étendue de MEDIA annotée en intentions pour étendre son utilisation. Cet article présente une méthode semi-automatique pour obtenir cette version étendue. De plus, nous présentons les premiers résultats des expériences menées sur cet ensemble de données en utilisant des modèles joints pour la classification des intentions et la détection de concepts.

Ce travail s’inscrit dans le débat sur l’efficacité des grands modèles de langue par rapport aux petits pour la classification de texte par amorçage (prompting). Nous évaluons ici le potentiel des petits modèles de langue dans la classification de texte sans exemples, remettant en question la prédominance des grands modèles. À travers un ensemble diversifié de jeux de données, notre étude compare les petits et les grands modèles utilisant différentes architectures et données de pré-entraînement. Nos conclusions révèlent que les petits modèles peuvent générer efficacement des étiquettes et, dans certains contextes, rivaliser ou surpasser les performances de leurs homologues plus grands. Ce travail souligne l’idée que le modèle le plus grand n’est pas toujours le meilleur, suggérant que les petits modèles économes en ressources peuvent offrir des solutions viables pour des défis spécifiques de classification de données

Within the current trend of Pretained Language Models (PLM), emerge more and more criticisms about the ethical and ecological impact of such models. In this article, considering these critical remarks, we propose to focus on smaller models, such as compact models like ALBERT, which are more ecologically virtuous than these PLM. However, PLMs enable huge breakthroughs in Natural Language Processing tasks, such as Spoken and Natural Language Understanding, classification, Question–Answering tasks. PLMs also have the advantage of being multilingual, and, as far as we know, a multilingual version of compact ALBERT models does not exist. Considering these facts, we propose the free release of the first version of a multilingual compact ALBERT model, pre-trained using Wikipedia data, which complies with the ethical aspect of such a language model. We also evaluate the model against classical multilingual PLMs in classical NLP tasks. Finally, this paper proposes a rare study on the subword tokenization impact on language performances.

Intent classification and slot-filling are essential tasks of Spoken Language Understanding (SLU). In most SLU systems, those tasks are realized by independent modules, but for about fifteen years, models achieving both of them jointly and exploiting their mutual enhancement have been proposed. A multilingual module using a joint model was envisioned to create a touristic dialogue system for a European project, HumanE-AI-Net. A combination of multiple datasets, including the MEDIA dataset, was suggested for training this joint model. The MEDIA SLU dataset is a French dataset distributed since 2005 by ELRA, mainly used by the French research community and free for academic research since 2020. Unfortunately, it is annotated only in slots but not intents. An enhanced version of MEDIA annotated with intents has been built to extend its use to more tasks and use cases. This paper presents the semi-automatic methodology used to obtain this enhanced version. In addition, we present the first results of SLU experiments on this enhanced dataset using joint models for intent classification and slot-filling.

This study is part of the debate on the efficiency of large versus small language models for text classification by prompting. We assess the performance of small language models in zero-shot text classification, challenging the prevailing dominance of large models. Across 15 datasets, our investigation benchmarks language models from 77M to 40B parameters using different architectures and scoring functions. Our findings reveal that small models can effectively classify texts, getting on par with or surpassing their larger counterparts. We developed and shared a comprehensive open-source repository that encapsulates our methodologies. This research underscores the notion that bigger isn’t always better, suggesting that resource-efficient small models may offer viable solutions for specific data classification challenges.

Le développement de solutions de traitement automatique de la langue pour de nouvelles tâches nécessite des données, dont l’obtention est coûteuses. L’accès aux données peut être limité en raison de la nature sensible des données. La plupart des travaux récents ont exploité de grands modèles pré-entraînés pour initialiser des versions spécialisées de ceux-ci. La spécialisation d’un tel modèle nécessite toujours une quantité élevée de données étiquetées spécifiques à la tâche cible. Nous utilisons l’apprentissage semi-supervisé pour entraîner des modèles dans un contexte où le nombre d’exemples étiquetés est limité et le nombre de données non étiquetées est nul. Nous étudions plusieurs méthodes pour générer le corpus non étiqueté nécessaire à l’utilisation de l’apprentissage semi-supervisé. Nous introduisons les méthodes de génération entre les épisodes d’entraînement et utilisons les modèles entraînés pour filtrer les exemples générés. Nous testons cette génération avec le tri-apprentissage et l’auto-apprentissage sur des corpus Anglais et Français.

Training a tagger for Named Entity Recognition (NER) requires a substantial amount of labeled data in the task domain. Manual labeling is a tedious and complicated task. Semisupervised learning methods can reduce the quantity of labeled data necessary to train a model. However, these methods require large quantities of unlabeled data, which remains an issue in many cases.We address this problem by generating unlabeled data. Large language models have proven to be powerful tools for text generation. We use their generative capacity to produce new sentences and variations of the sentences of our available data. This generation method, combined with a semi-supervised method, is evaluated on CoNLL and I2B2. We prepare both of these corpora to simulate a low resource setting. We obtain significant improvements for semisupervised learning with synthetic data against supervised learning on natural data.

Dialog state tracking (DST) is a core step for task-oriented dialogue systems aiming to track the user’s current goal during a dialogue. Recently a special focus has been put on applying existing DST models to new domains, in other words performing zero-shot cross-domain transfer. While recent state-of-the-art models leverage large pre-trained language models, no work has been made on understanding and improving the results of first developed zero-shot models like SUMBT. In this paper, we thus propose to improve SUMBT zero-shot results on MultiWOZ by using attention modulation during inference. This method improves SUMBT zero-shot results significantly on two domains and does not worsen the initial performance with the great advantage of needing no additional training.

Knowledge transfer between neural language models is a widely used technique that has proven to improve performance in a multitude of natural language tasks, in particular with the recent rise of large pre-trained language models like BERT. Similarly, high cross-lingual transfer has been shown to occur in multilingual language models. Hence, it is of great importance to better understand this phenomenon as well as its limits. While most studies about cross-lingual transfer focus on training on independent and identically distributed (i.e. i.i.d.) samples, in this paper we study cross-lingual transfer in a continual learning setting on two sequence labeling tasks: slot-filling and named entity recognition. We investigate this by training multilingual BERT on sequences of 9 languages, one language at a time, on the MultiATIS++ and MultiCoNER corpora. Our first findings are that forward transfer between languages is retained although forgetting is present. Additional experiments show that lost performance can be recovered with as little as a single training epoch even if forgetting was high, which can be explained by a progressive shift of model parameters towards a better multilingual initialization. We also find that commonly used metrics might be insufficient to assess continual learning performance.

For many tasks, state-of-the-art results have been achieved with Transformer-based architectures, resulting in a paradigmatic shift in practices from the use of task-specific architectures to the fine-tuning of pre-trained language models. The ongoing trend consists in training models with an ever-increasing amount of data and parameters, which requires considerable resources. It leads to a strong search to improve resource efficiency based on algorithmic and hardware improvements evaluated only for English. This raises questions about their usability when applied to small-scale learning problems, for which a limited amount of training data is available, especially for under-resourced languages tasks. The lack of appropriately sized corpora is a hindrance to applying data-driven and transfer learning-based approaches with strong instability cases. In this paper, we establish a state-of-the-art of the efforts dedicated to the usability of Transformer-based models and propose to evaluate these improvements on the question-answering performances of French language which have few resources. We address the instability relating to data scarcity by investigating various training strategies with data augmentation, hyperparameters optimization and cross-lingual transfer. We also introduce a new compact model for French FrALBERT which proves to be competitive in low-resource settings.

Supervised deep learning-based approaches have been applied to task-oriented dialog and have proven to be effective for limited domain and language applications when a sufficient number of training examples are available. In practice, these approaches suffer from the drawbacks of domain-driven design and under-resourced languages. Domain and language models are supposed to grow and change as the problem space evolves. On one hand, research on transfer learning has demonstrated the cross-lingual ability of multilingual Transformers-based models to learn semantically rich representations. On the other, in addition to the above approaches, meta-learning have enabled the development of task and language learning algorithms capable of far generalization. Through this context, this article proposes to investigate the cross-lingual transferability of using synergistically few-shot learning with prototypical neural networks and multilingual Transformers-based models. Experiments in natural language understanding tasks on MultiATIS++ corpus shows that our approach substantially improves the observed transfer learning performances between the low and the high resource languages. More generally our approach confirms that the meaningful latent space learned in a given language can be can be generalized to unseen and under-resourced ones using meta-learning.

In this paper, we present a study on a French Spoken Language Understanding (SLU) task: the MEDIA task. Many works and studies have been proposed for many tasks, but most of them are focused on English language and tasks. The exploration of a richer language like French within the framework of a SLU task implies to recent approaches to handle this difficulty. Since the MEDIA task seems to be one of the most difficult, according several previous studies, we propose to explore Neural Networks approaches focusing of three aspects: firstly, the Neural Network inputs and more specifically the word embeddings; secondly, we compared French version of BERT against the best setup through different ways; Finally, the comparison against State-of-the-Art approaches. Results show that the word embeddings trained on a small corpus need to be updated during SLU model training. Furthermore, the French BERT fine-tuned approaches outperform the classical Neural Network Architectures and achieves state of the art results. However, the contextual embeddings extracted from one of the French BERT approaches achieve comparable results in comparison to word embedding, when integrated into the proposed neural architecture.

The task of automatic misogyny identification and categorization has not received as much attention as other natural language tasks have, even though it is crucial for identifying hate speech in social Internet interactions. In this work, we address this sentence classification task from a representation learning perspective, using both a bidirectional LSTM and BERT optimized with the following metric learning loss functions: contrastive loss, triplet loss, center loss, congenerous cosine loss and additive angular margin loss. We set new state-of-the-art for the task with our fine-tuned BERT, whose sentence embeddings can be compared with a simple cosine distance, and we release all our code as open source for easy reproducibility. Moreover, we find that almost every loss function performs equally well in this setting, matching the regular cross entropy loss.

This paper describes the participation of LIMSI_UPV team in SemEval-2020 Task 9: Sentiment Analysis for Code-Mixed Social Media Text. The proposed approach competed in SentiMix HindiEnglish subtask, that addresses the problem of predicting the sentiment of a given Hindi-English code-mixed tweet. We propose Recurrent Convolutional Neural Network that combines both the recurrent neural network and the convolutional network to better capture the semantics of the text, for code-mixed sentiment analysis. The proposed system obtained 0.69 (best run) in terms of F1 score on the given test data and achieved the 9th place (Codalab username: somban) in the SentiMix Hindi-English subtask.

Named entity recognition (NER) from speech is usually made through a pipeline process that consists in (i) processing audio using an automatic speech recognition system (ASR) and (ii) applying a NER to the ASR outputs. The latest data available for named entity extraction from speech in French were produced during the ETAPE evaluation campaign in 2012. Since the publication of ETAPE’s campaign results, major improvements were done on NER and ASR systems, especially with the development of neural approaches for both of these components. In addition, recent studies have shown the capability of End-to-End (E2E) approach for NER / SLU tasks. In this paper, we propose a study of the improvements made in speech recognition and named entity recognition for pipeline approaches. For this type of systems, we propose an original 3-pass approach. We also explore the capability of an E2E system to do structured NER. Finally, we compare the performances of ETAPE’s systems (state-of-the-art systems in 2012) with the performances obtained using current technologies. The results show the interest of the E2E approach, which however remains below an updated pipeline approach.

La reconnaissance des entités nommées (REN) à partir de la parole est traditionnellement effectuée par l’intermédiaire d’une chaîne de composants, exploitant un système de reconnaissance de la parole (RAP), puis un système de REN appliqué sur les transcriptions automatiques. Les dernières données disponibles pour la REN structurées à partir de la parole en français proviennent de la campagne d’évaluation ETAPE en 2012. Depuis la publication des résultats, des améliorations majeures ont été réalisées pour les systèmes de REN et de RAP. Notamment avec le développement des systèmes neuronaux. De plus, certains travaux montrent l’intérêt des approches de bout en bout pour la tâche de REN dans la parole. Nous proposons une étude des améliorations en RAP et REN dans le cadre d’une chaîne de composants, ainsi qu’une nouvelle approche en trois étapes. Nous explorons aussi les capacités d’une approche bout en bout pour la REN structurées. Enfin, nous comparons ces deux types d’approches à l’état de l’art de la campagne ETAPE. Nos résultats montrent l’intérêt de l’approche bout en bout, qui reste toutefois en deçà d’une chaîne de composants entièrement mise à jour.

Cet article présente une nouvelle méthode d’étiquetage en parties du discours adaptée aux langues peu dotées : la définition du contexte utilisé pour construire les plongements lexicaux est adaptée à la tâche, et de nouveaux vecteurs sont créés pour les mots inconnus. Les expériences menées sur le picard, le malgache et l’alsacien montrent que cette méthode améliore l’état de l’art pour ces trois langues peu dotées.

In this article, we provide several approaches to the automatic identification of parallel sentences that require sentence-external linguistic context to be correctly translated. Our long-term goal is to automatically construct a test set of context-dependent sentences in order to evaluate machine translation models designed to improve the translation of contextual, discursive phenomena. We provide a discussion and critique that show that current approaches do not allow us to achieve our goal, and suggest that for now evaluating individual phenomena is likely the best solution.

We present the work-in-progress of automating the classification of doctor-patient questions in the context of a simulated consultation with a virtual patient. We classify questions according to the computational strategy (rule-based or other) needed for looking up data in the clinical record. We compare ‘traditional’ machine learning methods (Gaussian and Multinomial Naive Bayes, and Support Vector Machines) and a neural network classifier (FastText). We obtained the best results with the SVM using semantic annotations, whereas the neural classifier achieved promising results without it.

La désambiguïsation d’entités (ou liaison d’entités), qui consiste à relier des mentions d’entités d’un texte à des entités d’une base de connaissance, est un problème qui se pose, entre autre, pour le peuplement automatique de bases de connaissances à partir de textes. Une difficulté de cette tâche est la résolution d’ambiguïtés car les systèmes ont à choisir parmi un nombre important de candidats. Cet article propose une nouvelle approche fondée sur l’apprentissage joint de représentations distribuées des mots et des entités dans le même espace, ce qui permet d’établir un modèle robuste pour la comparaison entre le contexte local de la mention d’entité et les entités candidates.

Le travail que nous présentons ici s’inscrit dans le domaine de l’évaluation des systèmes de reconnaissance automatique de la parole en vue de leur utilisation dans une tâche aval, ici la reconnaissance des entités nommées. Plus largement, la question que nous nous posons est “que peut apporter une métrique d’évaluation en dehors d’un score ?". Nous nous intéressons particulièrement aux erreurs des systèmes et à leur analyse et éventuellement à l’utilisation de ce que nous connaissons de ces erreurs. Nous étudions dans ce travail les listes ordonnées d’erreurs générées à partir de différentes métriques et analysons ce qui en ressort. Nous avons appliqué la même méthode sur les sorties de différents systèmes de reconnaissance de la parole. Nos expériences mettent en évidence que certaines métriques apportent une information plus pertinente étant donné une tâche et transverse à différents systèmes.

Nous nous intéressons à l’évaluation de la qualité des systèmes de reconnaissance de la parole étant donné une tâche de compréhension. L’objectif de ce travail est de fournir un outil permettant la sélection d’un système de reconnaissance automatique de la parole le plus adapté pour un système de dialogue donné. Nous comparons ici différentes métriques, notamment le WER, NE-WER et ATENE métrique proposée récemment pour l’évaluation des systèmes de reconnaissance de la parole étant donné une tâche de reconnaissance d’entités nommées. Cette dernière métrique montrait une meilleure corrélation avec les résultats de la tâche globale que toutes les autres métriques testées. Nos mesures indiquent une très forte corrélation avec la mesure ATENE et une moins forte avec le WER.

Cet article examine l’utilisation du raisonnement analogique dans le contexte de l’apprentissage incrémental. Le problème d’apprentissage sous-jacent développé est le transfert de requêtes formulées en langue naturelle vers des commandes dans un langage de programmation. Nous y explorons deux questions principales : Comment se comporte le raisonnement par analogie dans le contexte de l’apprentissage incrémental ? De quelle manière la séquence d’apprentissage influence-t-elle la performance globale ? Pour y répondre, nous proposons un protocole expérimental simulant deux utilisateurs et différentes séquences d’apprentissage. Nous montrons que l’ordre dans la séquence d’apprentissage incrémental n’a d’influence notable que sous des conditions spécifiques. Nous constatons également la complémentarité de l’apprentissage incrémental avec l’analogie pour un nombre d’exemples d’apprentissage minimal.

Cette démonstration présente un système de dialogue en domaine ouvert qui utilise une base d’exemples de dialogue automatiquement constituée depuis un corpus de sous-titres afin de gérer un dialogue social de type « chatbot ».

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.

Automatic Speech recognition (ASR) is one of the most widely used components in spoken language processing applications. ASR errors are of varying importance with respect to the application, making error analysis keys to improving speech processing applications. Knowing the most serious errors for the applicative case is critical to build better systems. In the context of Automatic Speech Recognition (ASR) used as a first step towards Named Entity Recognition (NER) in speech, error seriousness is usually determined by their frequency, due to the use of the WER as metric to evaluate the ASR output, despite the emergence of more relevant measures in the literature. We propose to use a different evaluation metric form the literature in order to classify ASR errors according to their seriousness for NER. Our results show that the ASR errors importance is ranked differently depending on the used evaluation metric. A more detailed analysis shows that the estimation of the error impact given by the ATENE metric is more adapted to the NER task than the estimation based only on the most used frequency metric WER.

While measuring the readability of texts has been a long-standing research topic, assessing the technicality of terms has only been addressed more recently and mostly for the English language. In this paper, we train a learning-to-rank model to determine a specialization degree for each term found in a given list. Since no training data for this task exist for French, we train our system with non-lexical features on English data, namely, the Consumer Health Vocabulary, then apply it to French. The features include the likelihood ratio of the term based on specialized and lay language models, and tests for containing morphologically complex words. The evaluation of this approach is conducted on 134 terms from the UMLS Metathesaurus and 868 terms from the Eugloss thesaurus. The Normalized Discounted Cumulative Gain obtained by our system is over 0.8 on both test sets. Besides, thanks to the learning-to-rank approach, adding morphological features to the language model features improves the results on the Eugloss thesaurus.

This paper presents an automatic corpus-based process to author an open-domain conversational strategy usable both in chatterbot systems and as a fallback strategy for out-of-domain human utterances. Our approach is implemented on a corpus of television drama subtitles. This system is used as a chatterbot system to collect a corpus of 41 open-domain textual dialogues with 27 human participants. The general capabilities of the system are studied through objective measures and subjective self-reports in terms of understandability, repetition and coherence of the system responses selected in reaction to human utterances. Subjective evaluations of the collected dialogues are presented with respect to amusement, engagement and enjoyability. The main factors influencing those dimensions in our chatterbot experiment are discussed.

We introduce a dialogue task between a virtual patient and a doctor where the dialogue system, playing the patient part in a simulated consultation, must reconcile a specialized level, to understand what the doctor says, and a lay level, to output realistic patient-language utterances. This increases the challenges in the analysis and generation phases of the dialogue. This paper proposes methods to manage linguistic and terminological variation in that situation and illustrates how they help produce realistic dialogues. Our system makes use of lexical resources for processing synonyms, inflectional and derivational variants, or pronoun/verb agreement. In addition, specialized knowledge is used for processing medical roots and affixes, ontological relations and concept mapping, and for generating lay variants of terms according to the patient’s non-expert discourse. We also report the results of a first evaluation carried out by 11 users interacting with the system. We evaluated the non-contextual analysis module, which supports the Spoken Language Understanding step. The annotation of task domain entities obtained 91.8% of Precision, 82.5% of Recall, 86.9% of F-measure, 19.0% of Slot Error Rate, and 32.9% of Sentence Error Rate.

Recognition of real-world entities is crucial for most NLP applications. Since its introduction some twenty years ago, named entity processing has undergone a significant evolution with, among others, the definition of new tasks (e.g. entity linking) and the emergence of new types of data (e.g. speech transcriptions, micro-blogging). These pose certainly new challenges which affect not only methods and algorithms but especially linguistic resources. Where do we stand with respect to named entity resources? This paper aims at providing a systematic overview of named entity resources, accounting for qualities such as multilingualism, dynamicity and interoperability, and to identify shortfalls in order to guide future developments.

Dans cet article, nous présentons les méthodes que nous avons développées pour analyser des comptes- rendus hospitaliers rédigés en anglais. L’objectif de cette étude consiste à identifier les facteurs de risque de décès pour des patients diabétiques et à positionner les événements médicaux décrits par rapport à la date de création de chaque document. Notre approche repose sur (i) HeidelTime pour identifier les expressions temporelles, (ii) des CRF complétés par des règles de post-traitement pour identifier les traitements, les maladies et facteurs de risque, et (iii) des règles pour positionner temporellement chaque événement médical. Sur un corpus de 514 documents, nous obtenons une F-mesure globale de 0,8451. Nous observons que l’identification des informations directement mentionnées dans les documents se révèle plus performante que l’inférence d’informations à partir de résultats de laboratoire.

Le démonstrateur que nous décrivons ici est un prototype de système de dialogue dont l’objectif est de simuler un patient. Nous décrivons son fonctionnement général en insistant sur les aspects concernant la langue et surtout le rapport entre langue médicale de spécialité et langue générale.

The study provides an original standpoint of the speech transcription errors by focusing on the morpho-syntactic features of the erroneous chunks and of the surrounding left and right context. The typology concerns the forms, the lemmas and the POS involved in erroneous chunks, and in the surrounding contexts. Comparison with error free contexts are also provided. The study is conducted on French. Morpho-syntactic analysis underlines that three main classes are particularly represented in the erroneous chunks: (i) grammatical words (to, of, the), (ii) auxiliary verbs (has, is), and (iii) modal verbs (should, must). Such items are widely encountered in the ASR outputs as frequent candidates to transcription errors. The analysis of the context points out that some left 3-grams contexts (e.g., repetitions, that is disfluencies, bracketing formulas such as “c’est”, etc.) may be better predictors than others. Finally, the surface analysis conducted through a Levensthein distance analysis, highlighted that the most common distance is of 2 characters and mainly involves differences between inflected forms of a unique item.

This paper is concerned with human assessments of the severity of errors in ASR outputs. We did not design any guidelines so that each annotator involved in the study could consider the “seriousness” of an ASR error using their own scientific background. Eight human annotators were involved in an annotation task on three distinct corpora, one of the corpora being annotated twice, hiding this annotation in duplicate to the annotators. None of the computed results (inter-annotator agreement, edit distance, majority annotation) allow any strong correlation between the considered criteria and the level of seriousness to be shown, which underlines the difficulty for a human to determine whether a ASR error is serious or not.

This paper addresses the question of hierarchical named entity evaluation. In particular, we focus on metrics to deal with complex named entity structures as those introduced within the QUAERO project. The intended goal is to propose a smart way of evaluating partially correctly detected complex entities, beyond the scope of traditional metrics. None of the existing metrics are fully adequate to evaluate the proposed QUAERO task involving entity detection, classification and decomposition. We are discussing the strong and weak points of the existing metrics. We then introduce a new metric, the Entity Tree Error Rate (ETER), to evaluate hierarchical and structured named entity detection, classification and decomposition. The ETER metric builds upon the commonly accepted SER metric, but it takes the complex entity structure into account by measuring errors not only at the slot (or complex entity) level but also at a basic (atomic) entity level. We are comparing our new metric to the standard one using first some examples and then a set of real data selected from the ETAPE evaluation results.

Within the framework of the Quaero project, we proposed a new definition of named entities, based upon an extension of the coverage of named entities as well as the structure of those named entities. In this new definition, the extended named entities we proposed are both hierarchical and compositional. In this paper, we focused on the annotation of a corpus composed of press archives, OCRed from French newspapers of December 1890. We present the methodology we used to produce the corpus and the characteristics of the corpus in terms of named entities annotation. This annotated corpus has been used in an evaluation campaign. We present this evaluation, the metrics we used and the results obtained by the participants.

Text and speech corpora for training a tale telling robot have been designed, recorded and annotated. The aim of these corpora is to study expressive storytelling behaviour, and to help in designing expressive prosodic and co-verbal variations for the artificial storyteller). A set of 89 children tales in French serves as a basis for this work. The tales annotation principles and scheme are described, together with the corpus description in terms of coverage and inter-annotator agreement. Automatic analysis of a new tale with the help of this corpus and machine learning is discussed. Metrics for evaluation of automatic annotation methods are discussed. A speech corpus of about 1 hour, with 12 tales has been recorded and aligned and annotated. This corpus is used for predicting expressive prosody in children tales, above the level of the sentence.

In this paper we deal with named entity detection on data acquired via OCR process on documents dating from 1890. The resulting corpus is very noisy. We perform an analysis to find possible strategies to overcome errors introduced by the OCR process. We propose a preprocessing procedure in three steps to clean data and correct, at least in part, OCR mistakes. The task is made even harder by the complex tree-structure of named entities annotated on data, we solve this problem however by adopting an effective named entity detection system we proposed in previous work. We evaluate our procedure for preprocessing OCR-ized data in two ways: in terms of perplexity and OOV rate of a language model on development and evaluation data, and in terms of the performance of the named entity detection system on the preprocessed data. The preprocessing procedure results to be effective, allowing to improve by a large margin the system we proposed for the official evaluation campaign on Old Press, and allowing to outperform also the best performing system of the evaluation campaign.

Pourtant essentiel pour appréhender rapidement et globalement l’état de santé des patients, l’accès aux informations médicales liées aux prescriptions médicamenteuses et aux concepts médicaux par les outils informatiques se révèle particulièrement difficile. Ces informations sont en effet généralement rédigées en texte libre dans les comptes rendus hospitaliers et nécessitent le développement de techniques dédiées. Cet article présente les stratégies mises en oeuvre pour extraire les prescriptions médicales et les concepts médicaux dans des comptes rendus hospitaliers rédigés en anglais. Nos systèmes, fondés sur des approches à base de règles et d’apprentissage automatique, obtiennent une F1-mesure globale de 0,773 dans l’extraction des prescriptions médicales et dans le repérage et le typage des concepts médicaux.

La variabilité des corpus constitue un problème majeur pour les systèmes de reconnaissance d’entités nommées. L’une des pistes possibles pour y remédier est l’utilisation d’approches linguistiques pour les adapter à de nouveaux contextes : la construction de patrons sémantiques peut permettre de désambiguïser les entités nommées en structurant leur environnement syntaxico-sémantique. Cet article présente une première réalisation sur un corpus de presse d’un système de correction. Après une étape de segmentation sur des critères discursifs de surface, le système extrait et pondère les patrons liés à une classe d’entité nommée fournie par un analyseur. Malgré des modèles encore relativement élémentaires, les résultats obtenus sont encourageants et montrent la nécessité d’un traitement plus approfondi de la classe Organisation.

Cet article présente l’étude d’un corpus de réponses formulées par des humains à des questions factuelles. Des observations qualitatives et quantitatives sur la reprise d’éléments de la question dans les réponses sont exposées. La notion d’information-réponse est introduite et une étude de la présence de cet élément dans le corpus est proposée. Enfin, les formulations des réponses sont étudiées.

The Quaero project organized a set of evaluations of Named Entity recognition systems in 2009. One of the sub-tasks consists in extracting citations from patents, i.e. references to other documents, either other patents or general literature from English-language patents. We present in this paper the participation of LIMSI in this evaluation, with a complete system description and the evaluation results. The corpus shown that patent and non-patent citations have a very different nature. We then separated references to other patents and to general literature papers and we created a hybrid system. For patent citations, the system used rule-based expert knowledge on the form of regular expressions. The system for detecting non-patent citations, on the other hand, is purely stochastic (machine learning with CRF++). Then we mixed both approaches to provide a single output. 4 teams participated to this task and our system obtained the best results of this evaluation campaign, even if the difference between the first two systems is poorly significant.

In the QA and information retrieval domains progress has been assessed via evaluation campaigns(Clef, Ntcir, Equer, Trec).In these evaluations, the systems handle independent questions and should provide one answer to each question, extracted from textual data, for both open domain and restricted domain. Quæro is a program promoting research and industrial innovation on technologies for automatic analysis and classification of multimedia and multilingual documents. Among the many research areas concerned by Quæro. The Quaero project organized a series of evaluations of Question Answering on Web Data systems in 2008 and 2009. For each language, English and French the full corpus has a size of around 20Gb for 2.5M documents. We describe the task and corpora, and especially the methodologies used in 2008 to construct the test of question and a new one in the 2009 campaign. Six types of questions were addressed, factual, Non-factual(How, Why, What), List, Boolean. A description of the participating systems and the obtained results is provided. We show the difficulty for a question-answering system to work with complex data and questions.

The Quæro program that promotes research and industrial innovation on technologies for automatic analysis and classification of multimedia and multilingual documents. Within its context a set of evaluations of Named Entity recognition systems was held in 2009. Four tasks were defined. The first two concerned traditional named entities in French broadcast news for one (a rerun of ESTER 2) and of OCR-ed old newspapers for the other. The third was a gene and protein name extraction in medical abstracts. The last one was the detection of references in patents. Four different partners participated, giving a total of 16 systems. We provide a synthetic descriptions of all of them classifying them by the main approaches chosen (resource-based, rules-based or statistical), without forgetting the fact that any modern system is at some point hybrid. The metric (the relatively standard Slot Error Rate) and the results are also presented and discussed. Finally, a process is ongoing with preliminary acceptance of the partners to ensure the availability for the community of all the corpora used with the exception of the non-Quæro produced ESTER 2 one.

This paper presents a corpus of human answers in natural language collected in order to build a base of examples useful when generating natural language answers. We present the corpus and the way we acquired it. Answers correspond to questions with fixed linguistic form, focus, and topic. Answers to a given question exist for two modalities of interaction: oral and written. The whole corpus of answers was annotated manually and automatically on different levels including words from the questions being reused in the answer, the precise element answering the question (or information-answer), and completions. A detailed description of the annotations is presented. Two examples of corpus analyses are described. The first analysis shows some differences between oral and written modality especially in terms of length of the answers. The second analysis concerns the reuse of the question focus in the answers.

Question Answering (QA) technology aims at providing relevant answers to natural language questions. Most Question Answering research has focused on mining document collections containing written texts to answer written questions. In addition to written sources, a large (and growing) amount of potentially interesting information appears in spoken documents, such as broadcast news, speeches, seminars, meetings or telephone conversations. The QAST track (Question-Answering on Speech Transcripts) was introduced in CLEF to investigate the problem of question answering in such audio documents. This paper describes in detail the evaluation protocol and tools designed and developed for the CLEF-QAST evaluation campaigns that have taken place between 2007 and 2009. We first remind the data, question sets, and submission procedures that were produced or set up during these three campaigns. As for the evaluation procedure, the interface that was developed to ease the assessors’ work is described. In addition, this paper introduces a methodology for a semi-automatic evaluation of QAST systems based on time slot comparisons. Finally, the QAST Evaluation Package 2007-2009 resulting from these evaluation campaigns is also introduced.

This paper presents a preliminary analysis of the role of some discourse markers and the vocalic hesitation ""euh"" in a corpus of spoken human utterances collected with the Ritel system, an open domain and spoken dialog system. The frequency and contextual combinatory of classical discourse markers and of the vocalic hesitation have been studied. This analysis pointed out some specificity in terms of combinatory of the analyzed items. The classical discourse markers seem to help initiating larger discursive blocks both at initial and medial positions of the on-going turns. The vocalic hesitation stand also for marking the user's embarrassments and wish to close the dialog.

The performance of question answering system is evaluated through successive evaluations campaigns. A set of questions are given to the participating systems which are to find the correct answer in a collection of documents. The creation process of the questions may change from one evaluation to the next. This may entail an uncontroled question difficulty shift. For the QAst 2009 evaluation campaign, a new procedure was adopted to build the questions. Comparing results of QAst 2008 and QAst 2009 evaluations, a strong performance loss could be measured in 2009 for French and English, while the Spanish systems globally made progress. The measured loss might be related to this new way of elaborating questions. The general purpose of this paper is to propose a measure to calibrate the difficulty of a question set. In particular, a reasonable measure should output higher values for 2009 than for 2008. The proposed measure relies on a distance measure between the critical elements of a question and those of the associated correct answer. An increase of the proposed distance measure for French and English 2009 evaluations as compared to 2008 could be established. This increase correlates with the previously observed degraded performances. We conclude on the potential of this evaluation criterion: the importance of such a measure for the elaboration of new question corpora for questions answering systems and a tool to control the level of difficulty for successive evaluation campaigns.

Notre travail se situe dans le cadre des systèmes de réponse a une question et à pour but de fournir une réponse en langue naturelle aux questions posées en langue naturelle. Cet article présente une expérience permettant d’analyser les réponses de locuteurs du français à des questions que nous leur posons. L’expérience se déroule à l’écrit comme à l’oral et propose à des locuteurs français des questions relevant de différents types sémantiques et syntaxiques. Nous mettons en valeur une large variabilité dans les formes de réponses possibles en langue française. D’autre part nous établissons un certain nombre de liens entre formulation de question et formulation de réponse. Nous proposons d’autre part une comparaison des réponses selon la modalité oral / écrit. Ces résultats peuvent être intégrés à des systèmes existants pour produire une réponse en langue naturelle de façon dynamique.

This paper reports on the QAST track of CLEF aiming to evaluate Question Answering on Speech Transcriptions. Accessing information in spoken documents provides additional challenges to those of text-based QA, needing to address the characteristics of spoken language, as well as errors in the case of automatic transcriptions of spontaneous speech. The framework and results of the pilot QAst evaluation held as part of CLEF 2007 is described, illustrating some of the additional challenges posed by QA in spoken documents relative to written ones. The current plans for future multiple-language and multiple-task QAst evaluations are described.

The RITEL project aims to integrate a spoken language dialogue system and an open-domain information retrieval system in order to enable human users to ask a general question and to refine their search for information interactively. This type of system is often referred to as an Interactive Question Answering (IQA) system. In this paper, we present an evaluation of how the performance of the RITEL system differs when users interact with it using spoken versus textual input and output. Our results indicate that while users do not perceive the two versions to perform significantly differently, many more questions are asked in a typical text-based dialogue.

In this paper we present a real (as opposed to Wizard-of-Oz) Human-Computer QA-oriented spoken dialog corpus collected with our Ritel platform. This corpus has been orthographically transcribed and annotated in terms of Specific Entities and Topics. Twelve main topics have been chosen. They are refined into 22 sub-topics. The Specific Entities are from five categories and cover Named Entities, linguistic entities, topic-defining entities, general entities and extended entities. The corpus contains 582 dialogs for 6 hours of user speech.

The aim of the Media-Evalda project is to evaluate the understanding capabilities of dialog systems. This paper presents the Media protocol for speech understanding evaluation and describes the results of the June 2005 literal evaluation campaign. Five systems, both symbolic or corpus-based, participated to the evaluation which is based on a common semantic representation. Different scorings have been performed on the system results. The understanding error rate, for the Full scoring is, depending on the systems, from 29% to 41.3%. A diagnosis analysis of these results is proposed.

Ces dernières années, il y a eu de nombreux travaux portant sur l’utilisation d’actes de dialogue pour caractériser les dialogues homme-homme ou homme-machine. Cet article fait état de nos travaux sur la détection automatique d’actes de dialogue dans des corpus réels de dialogue homme-homme. Notre travail est fondé essentiellement sur deux hypothèses . (i) la position des mots et la classe sémantique du mot sont plus importants que les mots eux-mêmes pour identifier l’acte de dialogue et (ii) il y a une forte prédictivité dans la succession des actes de dialogues portés sur un même segment dialogique. Une approche de type Memory Based Learning a été utilisée pour la détection automatique des actes de dialogue. Le premier modèle n’utilise pas d’autres informations que celles contenus dans le tour de parole. Dans lex expériences suivantes, des historiques dialogiques de taille variables sont utilisés. Le taux d’erreur de détection d’actes de dialogue est d’environ 16% avec le premier modèle est descend avec une utilisation plus large de l’historique du dialogue à environ 14%.

L’objectif du projet RITEL est de réaliser un système de dialogue homme-machine permettant à un utilisateur de poser oralement des questions, et de dialoguer avec un système de recherche d’information généraliste (par exemple, chercher sur l’Internet “Qui est le Président du Sénat ?”) et d’en étudier les potentialités. Actuellement, la plateforme RITEL permet de collecter des corpus de dialogue homme-machine. Les utilisateurs peuvent parfois obtenir une réponse, de type factuel (Q : qui est le président de la France ; R : Jacques Chirac.). Cet article présente brièvement la plateforme développée, le corpus collecté ainsi que les questions que soulèvent un tel système et quelques unes des premières solutions envisagées.

The aim of the MEDIA project is to design and test a methodology for the evaluat ion of context-dependent and independent spoken dialogue systems. We propose an evaluation paradigm based on the use of test suites from real-world corpora and a common semantic representation and common metrics. This paradigm should allow us to diagnose the context-sensitive understanding capability of dialogue system s. This paradigm will be used within an evaluation campaign involving several si tes all of which will carry out the task of querying information from a database .

Dans cet article, nous présentons un gestionnaire de dialogue pour un système de demande d’informations à reconnaissance vocale. Le gestionnaire de dialogue dispose de différentes sources de connaissance, des connaissances statiques et des connaissances dynamiques. Ces connaissances sont gérées et utilisées par le gestionnaire de dialogue via des stratégies. Elles sont mises en oeuvre et organisées en fonction des objectifs concernant le système de dialogue et en fonction des choix ergonomiques que nous avons retenus. Le gestionnaire de dialogue utilise un modèle de dialogue fondé sur la détermination de phases et un modèle de la tâche dynamique. Il augmente les possibilités d’adaptation de la stratégie en fonction des historiques et de l’état du dialogue. Ce gestionnaire de dialogue, implémenté et évalué lors de la dernière campagne d’évaluation du projet LE-3 ARISE, a permi une amélioration du taux de succès de dialogue (de 53% à 85%).