International audience

In recent years, Pre-trained Language Models(PLMs) have been widely used to analyze various documents, playing a crucial role in Natural Language Processing (NLP). However, administrative texts have rarely been used in information extraction tasks, even though this resource is available as open data in many countries. Most of these texts contain many specific domain terms. Moreover, especially in France, they are unstructured because many administrations produce them without a standardized framework. Due to this fact, current language models do not process these documents correctly. In this paper, we propose AdminBERT, the first French pre-trained language models for the administrative domain. Since interesting information in such texts corresponds to named entities and the relations between them, we compare this PLM with general domain language models, fine-tuned on the Named Entity Recognition (NER) task applied to administrative texts, as well as to a Large Language Model (LLM) and to a language model with an architecture different from the BERT one. We show that taking advantage of a PLM for French administrative data increases the performance in the administrative and general domains, on these texts. We also release AdminBERT as well as AdminSet, the pre-training corpus of administrative texts in French and the subset AdminSet-NER, the first NER dataset consisting exclusively of administrative texts in French.

The biomedical domain has sparked a significant interest in the field of Natural Language Processing (NLP), which has seen substantial advancements with pre-trained language models (PLMs). However, comparing these models has proven challenging due to variations in evaluation protocols across different models. A fair solution is to aggregate diverse downstream tasks into a benchmark, allowing for the assessment of intrinsic PLMs qualities from various perspectives. Although still limited to few languages, this initiative has been undertaken in the biomedical field, notably English and Chinese. This limitation hampers the evaluation of the latest French biomedical models, as they are either assessed on a minimal number of tasks with non-standardized protocols or evaluated using general downstream tasks. To bridge this research gap and account for the unique sensitivities of French, we present the first-ever publicly available French biomedical language understanding benchmark called DrBenchmark. It encompasses 20 diversified tasks, including named-entity recognition, part-of-speech tagging, question-answering, semantic textual similarity, or classification. We evaluate 8 state-of-the-art pre-trained masked language models (MLMs) on general and biomedical-specific data, as well as English specific MLMs to assess their cross-lingual capabilities. Our experiments reveal that no single model excels across all tasks, while generalist models are sometimes still competitive.

Yanis Labrak, Adrien Bazoge, Richard Dufour, Mickael Rouvier, Emmanuel Morin, Béatrice Daille, Pierre-Antoine Gourraud. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Abstract Significant advances have been achieved in bilingual word-level alignment, yet the challenge remains for phrase-level alignment. Moreover, the need for parallel data is a critical drawback for the alignment task. This work proposes a system that alleviates these two problems: a unified phrase representation model using cross-lingual word embeddings as input and an unsupervised training algorithm inspired by recent works on neural machine translation. The system consists of a sequence-to-sequence architecture where a short sequence encoder constructs cross-lingual representations of phrases of any length, then an LSTM network decodes them w.r.t their contexts. After training with comparable corpora and existing key phrase extraction, our encoder provides cross-lingual phrase representations that can be compared without further transformation. Experiments on five data sets show that our method obtains state-of-the-art results on the bilingual phrase alignment task and improves the results of different length phrase alignment by a mean of 8.8 points in MAP.

Recent work has demonstrated the importance of dealing with Multi-Word Terms (MWTs) in several Natural Language Processing applications. In particular, MWTs pose serious challenges for alignment and machine translation systems because of their syntactic and semantic properties. Thus, developing algorithms that handle MWTs is becoming essential for many NLP tasks. However, the availability of bilingual and more generally multi-lingual resources is limited, especially for low-resourced languages and in specialized domains. In this paper, we propose an approach for building comparable corpora and bilingual term dictionaries that help evaluate bilingual term alignment in comparable corpora. To that aim, we exploit parallel corpora to perform automatic bilingual MWT extraction and comparable corpus construction. Parallel information helps to align bilingual MWTs and makes it easier to build comparable specialized sub-corpora. Experimental validation on an existing dataset and on manually annotated data shows the interest of the proposed methodology.

Lexical substitution task requires to substitute a target word by candidates in a given context. Candidates must keep meaning and grammatically of the sentence. The task, introduced in the SemEval 2007, has two objectives. The first objective is to find a list of substitutes for a target word. This list of substitutes can be obtained with lexical resources like WordNet or generated with a pre-trained language model. The second objective is to rank these substitutes using the context of the sentence. Most of the methods use vector space models or more recently embeddings to rank substitutes. Embedding methods use high contextualized representation. This representation can be over contextualized and in this way overlook good substitute candidates which are more similar on non-contextualized layers. SemDis 2014 introduced the lexical substitution task in French. We propose an application of the state-of-the-art method based on BERT in French and a novel method using contextualized and non-contextualized layers to increase the suggestion of words having a lower probability in a given context but that are more semantically similar. Experiments show our method increases the BERT based system on the OOT measure but decreases on the BEST measure in the SemDis 2014 benchmark.

Narrow specialized comparable corpora are often small in size.This particularity makes it difficult to build efficient models to acquire translation equivalents, especially for less frequent and rare words.One way to overcome this issue is to enrich the specialized corpora with out-ofdomain resources.Although some recent studies have shown improvements using data augmentation, the enrichment method was roughly conducted by adding out-of-domain data with no particular attention given to how to enrich words and how to do it optimally.In this paper, we contrast several data selection techniques to improve bilingual lexicon induction from specialized comparable corpora.We first apply two well-established data selection techniques often used in machine translation that is: Tf-Idf and cross entropy.Then, we propose to exploit BERT for data selection.Overall, all the proposed techniques improve the quality of the extracted bilingual lexicons by a large margin.The best performing model is the cross entropy, obtaining a gain of about 4 points in MAP while decreasing computation time by a factor of 10.

International audience

This paper presents a qualitative study of errors produced by an end-to-end spoken language understanding (SLU) system (speech signal to concepts) that reaches state of the art performance. Different studies are proposed to better understand the weaknesses of such systems: comparison to a classical pipeline SLU system, a study on the cause of concept deletions (the most frequent error), observation of a problem in the capability of the end-to-end SLU system to segment correctly concepts, analysis of the system behavior to process unseen concept/value pairs, analysis of the benefit of the curriculum-based transfer learning approach. Last, we proposed a way to compute embeddings of sub-sequences that seem to contain relevant information for future work.

Article soumis et accepté à la conférence Interspeech - Octobre 2020 - Shangai.

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.

We present an end-to-end approach to extract semantic concepts directly from the speech audio signal. To overcome the lack of data available for this spoken language understanding approach, we investigate the use of a transfer learning strategy based on the principles of curriculum learning. This approach allows us to exploit out-of-domain data that can help to prepare a fully neural architecture. Experiments are carried out on the French MEDIA and PORTMEDIA corpora and show that this end-to-end SLU approach reaches the best results ever published on this task. We compare our approach to a classical pipeline approach that uses ASR, POS tagging, lemmatizer, chunker... and other NLP tools that aim to enrich ASR outputs that feed an SLU text to concepts system. Last, we explore the promising capacity of our end-to-end SLU approach to address the problem of domain portability.

Lack of data can be an issue when beginning a new study on historical handwritten documents. In order to deal with this, we present the character-based decoder part of a multilingual approach based on transductive transfer learning for a historical handwriting recognition task on Italian Comedy Registers. The decoder must build a sequence of characters that corresponds to a word from a vector of letter-ngrams. As learning data, we created a new dataset from untapped resources that covers the same domain and period of our Italian Comedy data, as well as resources from common domains, periods, or languages. We obtain a 97.42% Character Recognition Rate and a 86.57% Word Recognition Rate on our Italian Comedy data, despite a lexical coverage of 67% between the Italian Comedy data and the training data. These results show that an efficient system can be obtained by a carefully selecting the datasets used for the transfer learning.

Recent evaluations on bilingual lexicon extraction from specialized comparable corpora have shown contrasted performance while using word embedding models. This can be partially explained by the lack of large specialized comparable corpora to build efficient representations. Within this context, we try to answer the following questions: First, (i) among the state-of-the-art embedding models, whether trained on specialized corpora or pre-trained on large general data sets, which one is the most appropriate model for bilingual terminology extraction? Second (ii) is it worth it to combine multiple embeddings trained on different data sets? For that purpose, we propose the first systematic evaluation of different word embedding models for bilingual terminology extraction from specialized comparable corpora. We emphasize how the character-based embedding model outperforms other models on the quality of the extracted bilingual lexicons. Further more, we propose a new efficient way to combine different embedding models learned from specialized and general-domain data sets. Our approach leads to higher performance than the best individual embedding model.

Extracting a bilingual terminology for multi-word terms from comparable corpora has not been widely researched. In this work we propose a unified framework for aligning bilingual terms independently of the term lengths. We also introduce some enhancements to the context-based and the neural network based approaches. Our experiments show the effectiveness of our enhancements of previous works and the system can be adapted in specialized domains.

Lack of data can be an issue when beginning a new study on historical handwritten documents. To deal with this, we propose a deep-learning based recognizer which separates the optical and the language models in order to train them separately using different resources. In this work, we present the optical encoder part of a multilingual transductive transfer learning applied to historical handwriting recognition. The optical encoder transforms the input word image into a non-latent space that depends only on the letter-n-grams: it enables it to be independent of the language. This transformation avoids embedding a language model and operating the transfer learning across languages using the same alphabet. The language decoder creates from a vector of letter-n-grams a word as a sequence of characters. Experiments show that separating optical and language model can be a solution for multilingual transfer learning.

International audience

Named entity recognition (NER) is among SLU tasks that usually extract semantic information from textual documents. Until now, NER from speech is made through a pipeline process that consists in processing first an automatic speech recognition (ASR) on the audio and then processing a NER on the ASR outputs. Such approach has some disadvantages (error propagation, metric to tune ASR systems sub-optimal in regards to the final task, reduced space search at the ASR output level,...) and it is known that more integrated approaches outperform sequential ones, when they can be applied. In this paper, we explore an end-to-end approach that directly extracts named entities from speech, though a unique neural architecture. On a such way, a joint optimization is possible for both ASR and NER. Experiments are carried on French data easily accessible, composed of data distributed in several evaluation campaigns. The results are promising since this end-to-end approach provides similar results (F-measure=0.66 on test data) than a classical pipeline approach to detect named entity categories (F-measure=0.64). Last, we also explore this approach applied to semantic concept extraction, through a slot filling task known as a spoken language understanding problem, and also observe an improvement in comparison to a pipeline approach.

Bilingual lexicon extraction from comparable corpora is constrained by the small amount of available data when dealing with specialized domains. This aspect penalizes the performance of distributional-based approaches, which is closely related to the reliability of word’s cooccurrence counts extracted from comparable corpora. A solution to avoid this limitation is to associate external resources with the comparable corpus. Since bilingual word embeddings have recently shown efficient models for learning bilingual distributed representation of words, we explore different word embedding models and show how a general-domain comparable corpus can enrich a specialized comparable corpus via neural networks

As the amount of news information available online grows, media are in need of advanced tools to explore the information surrounding specific events before writing their own piece of news, e.g., adding context and insight. While many tools exist to extract information from large datasets, they do not offer an easy way to gain insight from a news collection by browsing, going from article to article and viewing unaltered original content. Such browsing tools require the creation of rich underlying structures such as graph representations. These representations can be further enhanced by typing links that connect nodes, in order to inform the user on the nature of their relation. In this article, we introduce an efficient way to generate links between news items in order to obtain an easily navigable graph, and enrich this graph by automatically typing created links. User evaluations are conducted on real world data in order to assess for the interest of both the graph representation and link typing in a press reviewing task, showing a significant improvement compared to classical search engines.

Comparable corpora are the main alternative to the use of parallel corpora to extract bilingual lexicons. Although it is easier to build comparable corpora, specialized comparable corpora are often of modest size in comparison with corpora issued from the general domain. Consequently, the observations of word co-occurrences which are the basis of context-based methods are unreliable. We propose in this article to improve word co-occurrences of specialized comparable corpora and thus context representation by using general-domain data. This idea, which has been already used in machine translation task for more than a decade, is not straightforward for the task of bilingual lexicon extraction from specific-domain comparable corpora. We go against the mainstream of this task where many studies support the idea that adding out-of-domain documents decreases the quality of lexicons. Our empirical evaluation shows the advantages of this approach which induces a significant gain in the accuracy of extracted lexicons.

Abstract The main work in bilingual lexicon extraction from comparable corpora is based on the implicit hypothesis that corpora are balanced in terms of size. However, the historical context-based projection method is relatively insensitive to the size of each part of the comparable corpus. Within this context, we have carried out a study on the influence of unbalanced specialized comparable corpora and on the quality of bilingual terminology extraction by doing different experiments. Moreover, we have introduced a strategy into the context-based projection method to re-estimate word co-occurrence observations. This is done by using smoothing or prediction techniques that boost the observations of word co-occurrences which are mainly useful for the smallest part of an unbalanced comparable corpus. Our results show that the use of unbalanced specialized comparable corpora results in a significant improvement in the quality of extracted lexicons.

The main work in bilingual lexicon extraction from comparable corpora is based on the implicit hypothesis that corpora are balanced. However, the historical contextbased projection method dedicated to this task is relatively insensitive to the sizes of each part of the comparable corpus. Within this context, we have carried out a study on the influence of unbalanced specialized comparable corpora on the quality of bilingual terminology extraction through different experiments. Moreover, we have introduced a regression model that boosts the observations of word cooccurrences used in the context-based projection method. Our results show that the use of unbalanced specialized comparable corpora induces a significant gain in the quality of extracted lexicons.

Named Entity Recognition (NER) from speech usually involves two sequential steps: transcribing the speech using Automatic Speech Recognition (ASR) and annotating the outputs of the ASR process using NER techniques. Recognizing named entities in automatic transcripts is difficult due to the presence of transcription errors and the absence of some important NER clues, such as capitalization and punctuation. In this paper, we describe a methodology for speech NER which consists of incorporating NER into the ASR process so that the ASR system generates transcripts annotated with named entities. The combination is achieved by adapting ASR language models and pre-annotating the pronunciation dictionary. We evaluate this method on Ester 2 corpus, and show significant improvements over traditional approaches.

In this paper, we present HAMEX, a new public dataset that contains mathematical expressions available in their on-line handwritten form and in their audio spoken form. We have designed this dataset so that, given a mathematical expression, its handwritten signal and its audio signal can be used jointly to design multimodal recognition systems. Here, we describe the different steps that allowed us to acquire this dataset, from the creation of the mathematical expression corpora (including expressions from Wikipedia pages) to the segmentation and the transcription of the collected data, via the data collection process itself. Currently, the dataset contains 4 350 on-line handwritten mathematical expressions written by 58 writers, and the corresponding audio expressions (in French) spoken by 58 speakers. The ground truth is also provided both for the handwritten expressions (as INKML files with the digital ink, the symbol segmentation, and the MATHML structure) and for the audio expressions (as XML files with the transcriptions of the spoken expressions).

The traditional weighting schemes used in text categorization for the vector space model (VSM) cannot exploit information intrinsic to texts obtained through online handwriting recognition or any OCR process. Especially, top n (n > 1) recognition candidates could not be used without flooding the resulting text with false occurrences of spurious terms. In this paper, an improved weighting scheme for text categorization, that estimates the occurrences of terms from the posterior probabilities of the top n candidates, is proposed. The experimental results show that the categorization performances increase for texts with high error rates.

With the growth of on-line handwriting technologies, managing facilities for handwritten documents, such as retrieval of documents by topic, are required. These documents can contain graphics, equations or text for instance. This work reports experiments on categorization of on-line handwritten documents based on their textual contents. We assume that handwritten text blocks have been extracted from the documents, and as a first step of the proposed system, we process them with an existing handwritten recognition engine. We analyse the effect of the word recognition rate on the categorization performances, and we compare them with those obtained with the same texts available as ground truth. Two categorization algorithms (kNN and SVM) are compared in this work. The handwritten texts are a subset of the Reuters-21578 corpus collected from more than 1500 writers. Results show that there is no significant categorization performance loss when the word error rate stands below 22%.

Handwriting is an alternative method for entering texts composing Short Message Services. However, a whole new language features the texts which are produced. They include for instance abbreviations and other consonantal writing which sprung up for time saving and fashion. We have collected and processed a significant number of such handwriting SMS, and used various strategies to tackle this challenging area of handwriting recognition. We proposed to study more specifically three different phenomena: consonant skeleton, rebus, and phonetic writing. For each of them, we compare the rough results produced by a standard recognition system with those obtained when using a specific language model.

This paper highlights the interest of a language model in increasing the performances of on-line handwriting recognition systems. Models based on statistical approaches, trained on written corpora, have been investigated. Two kinds of models have been studied: n-gram models and n-class models. In the latter case, the classes result either from a syntactic criteria or a contextual criteria. In order to integrate it into small capacity systems (mobile device), an n-class model has been designed by combining these criteria. It outperforms bulkier models based on n-gram. Integration into an online handwriting recognition system demonstrates a substantial performance improvement due to the language model. 1.

Hypernym links acquired through an information extraction procedure are projected on multi-word terms through the recognition of semantic variations.The quality of the projected links resulting from corpus-based acquisition is compared with projected links extracted from a technical thesaurus.1 Motivation In the domain of corpus-based terminology, there are two main topics of research: term acquisition--the discovery of candidate terms-and automatic thesaurus construction--the addition of semantic links to a term bank.Several studies have focused on automatic acquisition of terms from corpora (Bourigault, 1993;Justeson and Katz, 1995;Daille, 1996).The output of these tools is a list of unstructured multi-word terms.On the other hand, contributions to automatic construction of thesauri provide classes or links between single words.Classes are produced by clustering techniques based on similar word contexts (Schiitze, 1993) or similar distributional contexts (Grefenstette, 1994).Links result from automatic acquisition of relevant predicative or discursive patterns (Hearst, 1992;Basili et al., 1993;Riloff, 1993).Predicative patterns yield predicative relations such as cause or effect whereas discursive patterns yield non-predicative relations such as generic/specific or synonymy links.* The experiments presented in this paper were performed on [AGRO], a 1.3-million word French corpus of scientific abstracts in the agricultural domain.The termer used for multi-word term acquisition is ACABIT (Daille, 1996).It has produced 15,875 multi-word terms composed of 4,194 single words.For expository purposes, some examples are taken from [MEDIC], a 1.56million word English corpus of scientific abstracts in the medical domain.