International audience

International audience

Pretrained Language Models (PLMs) are the de facto backbone of most state-of-the-art NLP systems. In this paper, we introduce a family of domain-specific pretrained PLMs for French, focusing on three important domains: transcribed speech, medicine, and law. We use a transformer architecture based on efficient methods (LinFormer) to maximise their utility, since these domains often involve processing long documents. We evaluate and compare our models to state-of-the-art models on a diverse set of tasks and datasets, some of which are introduced in this paper. We gather the datasets into a new French-language evaluation benchmark for these three domains. We also compare various training configurations: continued pretraining, pretraining from scratch, as well as single- and multi-domain pretraining. Extensive domain-specific experiments show that it is possible to attain competitive downstream performance even when pre-training with the approximative LinFormer attention mechanism. For full reproducibility, we release the models and pretraining data, as well as contributed datasets.

This paper describes MAKE-NMTViz, a project designed to help translators visualize neural machine translation outputs using explainable artificial intelligence visualization tools initially developed for computer vision.

This paper describes work in progress on Visualisation tools to foster collaborations between translators and computational scientists. We aim to describe how visualisation features can be used to explain translation and NMT outputs. We tested several visualisation functionalities with three NMT models based on Chinese-English, Spanish-English and French-English language pairs. We created three demos containing different visualisation tools and analysed them within the framework of performance-explainability, focusing on the translator’s perspective.

Self-Supervised Learning (SSL) has proven to be effective in various domains, including speech processing. However, SSL is computationally and memory expensive. This is in part due the quadratic complexity of multi-head self-attention (MHSA). Alternatives for MHSA have been proposed and used in the speech domain, but have yet to be investigated properly in an SSL setting. In this work, we study the effects of replacing MHSA with recent state-of-the-art alternatives that have linear complexity, namely, HyperMixing, Fastformer, SummaryMixing, and Mamba. We evaluate these methods by looking at the speed, the amount of VRAM consumed, and the performance on the SSL MP3S benchmark. Results show that these linear alternatives maintain competitive performance compared to MHSA while, on average, decreasing VRAM consumption by around 20% to 60% and increasing speed from 7% to 65% for input sequences ranging from 20 to 80 seconds.

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.

International audience

In this paper we present the final result of a project focused on Tunisian Arabic encoded in Arabizi, the Latin-based writing system for digital conversations. The project led to the realization of two integrated and independent tools: a linguistic corpus and a neural network architecture created to annotate the former with various levels of linguistic information (code-switching classification, transliteration, tokenization, POS-tagging, lemmatization). We discuss the choices made in terms of computational and linguistic methodology and the strategies adopted to improve our results. We report on the experiments performed in order to outline our research path. Finally, we explain the reasons why we believe in the potential of these tools for both computational and linguistic researches.

Self-Supervised Learning (SSL) using huge unlabeled data has been successfully explored for image and natural language processing. Recent works also investigated SSL from speech. They were notably successful to improve performance on downstream tasks such as automatic speech recognition (ASR). While these works suggest it is possible to reduce dependence on labeled data for building efficient speech systems, their evaluation was mostly made on ASR and using multiple and heterogeneous experimental settings (most of them for English). This questions the objective comparison of SSL approaches and the evaluation of their impact on building speech systems. In this paper, we propose LeBenchmark: a reproducible framework for assessing SSL from speech. It not only includes ASR (high and low resource) tasks but also spoken language understanding, speech translation and emotion recognition. We also focus on speech technologies in a language different than English: French. SSL models of different sizes are trained from carefully sourced and documented datasets. Experiments show that SSL is beneficial for most but not all tasks which confirms the need for exhaustive and reliable benchmarks to evaluate its real impact. LeBenchmark is shared with the scientific community for reproducible research in SSL from speech.

End-to-end architectures have been recently proposed for spoken language understanding (SLU) and semantic parsing. Based on a large amount of data, those models learn jointly acoustic and linguistic-sequential features. Such architectures give very good results in the context of domain, intent and slot detection, their application in a more complex semantic chunking and tagging task is less easy. For that, in many cases, models are combined with an external a language model to enhance their performance.In this paper we introduce a data efficient system which is trained end-to-end, with no additional, pre-trained external module. One key feature of our approach is an incremental training procedure where acoustic, language and semantic models are trained sequentially one after the other. The proposed model has a reasonable size and achieves competitive results with respect to state-of-the-art while using a small training dataset. In particular, we reach 24.02% Concept Error Rate (CER) on MEDIA/test while training on MEDIA/train without any additional data.

This article describes the constitution process of the first morpho-syntactically annotated Tunisian Arabish Corpus (TArC). Arabish, also known as Arabizi, is a spontaneous coding of Arabic dialects in Latin characters and “arithmographs” (numbers used as letters). This code-system was developed by Arabic-speaking users of social media in order to facilitate the writing in the Computer-Mediated Communication (CMC) and text messaging informal frameworks. Arabish differs for each Arabic dialect and each Arabish code-system is under-resourced, in the same way as most of the Arabic dialects. In the last few years, the attention of NLP studies on Arabic dialects has considerably increased. Taking this into consideration, TArC will be a useful support for different types of analyses, computational and linguistic, as well as for NLP tools training. In this article we will describe preliminary work on the TArC semi-automatic construction process and some of the first analyses we developed on TArC. In addition, in order to provide a complete overview of the challenges faced during the building process, we will present the main Tunisian dialect characteristics and its encoding in Tunisian Arabish.

Modelling target label dependencies is important for sequence labelling tasks. This may become crucial in the case of Spoken Language Understanding (SLU) applications, especially for the slot-filling task where models have to deal often with a high number of target labels. Conditional Random Fields (CRF) were previously considered as the most efficient algorithm in these conditions. More recently, different architectures of Recurrent Neural Networks (RNNs) have been proposed for the SLU slot-filling task. Most of them, however, have been successfully evaluated on the simple ATIS database, on which it is difficult to draw significant conclusions. In this paper we propose new variants of RNNs able to learn efficiently and effectively label dependencies by integrating label embeddings. We show first that modeling label dependencies is useless on the (simple) ATIS database and unstructured models can produce state-of-the-art results on this benchmark. On ATIS our new variants achieve the same results as state-of-the-art models, while being much simpler. On the other hand, on the MEDIA benchmark, we show that the modification introduced in the proposed RNN outperforms traditional RNNs and CRF models.

Spoken language understanding (SLU) is concerned with the extraction of meaning structures from spoken utterances. Recent computational approaches to SLU, e.g., conditional random fields (CRFs), optimize local models by encoding several features, mainly based on simple n-grams. In contrast, recent works have shown that the accuracy of CRF can be significantly improved by modeling long-distance dependency features. In this paper, we propose novel approaches to encode all possible dependencies between features and most importantly among parts of the meaning structure, e.g., concepts and their combination. We rerank hypotheses generated by local models, e.g., stochastic finite state transducers (SFSTs) or CRF, with a global model. The latter encodes a very large number of dependencies (in the form of trees or sequences) by applying kernel methods to the space of all meaning (sub) structures. We performed comparative experiments between SFST, CRF, support vector machines (SVMs), and our proposed discriminative reranking models (DRMs) on representative conversational speech corpora in three different languages: the ATIS (English), the MEDIA (French), and the LUNA (Italian) corpora. These corpora have been collected within three different domain applications of increasing complexity: informational, transactional, and problem-solving tasks, respectively. The results show that our DRMs consistently outperform the state-of-the-art models based on CRF.

One of the first steps in building a spoken language understanding (SLU) module for dialogue systems is the extraction of flat concepts out of a given word sequence, usually provided by an automatic speech recognition (ASR) system. In this paper, six different modeling approaches are investigated to tackle the task of concept tagging. These methods include classical, well-known generative and discriminative methods like Finite State Transducers (FSTs), Statistical Machine Translation (SMT), Maximum Entropy Markov Models (MEMMs), or Support Vector Machines (SVMs) as well as techniques recently applied to natural language processing such as Conditional Random Fields (CRFs) or Dynamic Bayesian Networks (DBNs). Following a detailed description of the models, experimental and comparative results are presented on three corpora in different languages and with different complexity. The French MEDIA corpus has already been exploited during an evaluation campaign and so a direct comparison with existing benchmarks is possible. Recently collected Italian and Polish corpora are used to test the robustness and portability of the modeling approaches. For all tasks, manual transcriptions as well as ASR inputs are considered. Additionally to single systems, methods for system combination are investigated. The best performing model on all tasks is based on conditional random fields. On the MEDIA evaluation corpus, a concept error rate of 12.6% could be achieved. Here, additionally to attribute names, attribute values have been extracted using a combination of a rule-based and a statistical approach. Applying system combination using weighted ROVER with all six systems, the concept error rate (CER) drops to 12.0%.

We present a call routing application for complex problem solving tasks. Up to date work on call routing has been mainly dealing with call-type classification. In this paper we take call routing further: Initial call classification is done in parallel with a robust statistical Spoken Language Understanding module. This is followed by a dialogue to elicit further task-relevant details from the user before passing on the call. The dialogue capability also allows us to obtain clarifications of the initial classifier guess. Based on an evaluation, we show that conducting a dialogue significantly improves upon call routing based on call classification alone. We present both subjective and objective evaluation results of the system according to standard metrics on real users.

Discriminative reranking has been successfully used for several tasks of Natural Language Processing (NLP). Recently it has been applied also to Spoken Language Understanding, imrpoving state-of-the-art for some applications. However, such proposed models can be further improved by considering: (i) a better selection of the initial n-best hypotheses to be re-ranked and (ii) the use of a strategy that decides when the reranking model should be used, i.e. in some cases only the basic approach should be applied. In this paper, we apply a semantic inconsistency metric to select the n-best hypotheses from a large set generated by an SLU basic system. Then we apply a state-of-the-art re-ranker based on the Partial Tree Kernel (PTK), which encodes SLU hypotheses in Support Vector Machines (SVM) with complex structured features. Finally, we apply a decision model based on confidence values to select between the first hypothesis provided by the basic SLU model and the first hypothesis provided by the re-ranker. We show the effectiveness of our approach presenting comparative results obtained by reranking hypotheses generated by two very different models: a simple Stochastic Language Model encoded in Finite State Machines (FSM) and a Conditional Random Field (CRF) model. We evaluate our approach on the French MEDIA corpus and on an Italian corpus acquired in the European Project LUNA. The results show a significant improvement with respect to the current state-of-the-art and previous re-ranking models.

Current Spoken Language Understanding models rely on either hand-written semantic grammars or flat attribute-value sequence labeling. In most cases, no relations between concepts are modeled, and both concepts and relations are domain-specific, making it difficult to expand or port the domain model. In contrast, we expand our previous work on a domain model based on an ontology where concepts follow the predicate-argument semantics and domain-independent classical relations are defined on such concepts. We conduct a thorough study on a spoken dialog corpus collected within a customer care problem-solving domain, and we evaluate the coverage and impact of the ontology for the interpretation, grounding and re-ranking of spoken language understanding interpretations.

Spoken Language Understanding performs automatic concept labeling and segmentation of speech utterances. For this task, many approaches have been proposed based on both generative and discriminative models. While all these methods have shown remarkable accuracy on manual transcription of spoken utterances, robustness to noisy automatic transcription is still an open issue. In this paper we study algorithms for Spoken Language Understanding combining complementary learning models: Stochastic Finite State Transducers produce a list of hypotheses, which are re-ranked using a discriminative algorithm based on kernel methods. Our experiments on two different spoken dialog corpora, MEDIA and LUNA, show that the combined generative-discriminative model reaches the state-of-the-art such as Conditional Random Fields (CRF) on manual transcriptions, and it is robust to noisy automatic transcriptions, outperforming, in some cases, the state-of-the-art. Copyright © 2009 ISCA.

Current Spoken Language Understanding systems rely either on hand-written semantic grammars or on flat attribute-value se-quence labeling. In both approaches, concepts and their rela-tions (when modeled at all) are domain-specific, thus making it difficult to expand or port the domain model. To address this issue, we introduce: 1) a domain model based on an ontology where concepts are classified into either predicative or argumentative; 2) the modeling of relations be-tween such concept classes in terms of classical relations as defined in lexical semantics. We study and analyze our ap-proach on a corpus of customer care data, where we evaluate the coverage and relevance of the ontology for the interpreta-tion of speech utterances. Index Terms: Spoken Language Understanding, domain mod-eling, ontology design, semantic relations

In this paper, we describe the semantic content, which can be automatically generated, for the design of advanced dialog systems. Since the latter will be based on machine learning approaches, we created training data by annotating a corpus with the needed content. Given a sentence of our transcribed corpus, domain concepts and other linguistic levels ranging from basic ones, i.e. part-of-speech tagging and constituent chunking level, to more advanced ones, i.e. syntactic and predicate argument structure (PAS) levels are annotated. In particular, the proposed PAS and taxonomy of dialog acts appear to be promising for the design of more complex dialog systems. Statistics about our semantic annotation are reported.

Spoken Language Understanding aims at mapping a natural language spoken sentence into a semantic representation. In the last decade two main approaches have been pursued: generative and discriminative models. The former is more robust to overfitting whereas the latter is more robust to many irrelevant features. Additionally, the way in which these approaches encode prior knowledge is very different and their relative performance changes based on the task. In this paper we describe a training framework where both models are used: a generative model produces a list of ranked hypotheses whereas a discriminative model, depending on string kernels and Support Vector Machines, re-ranks such list. We tested such approach on a new corpus produced in the European LUNA project. The results show a large improvement on the state-of-the-art in concept segmentation and labeling.