7.1 New software

7.2 New platforms

Over the last couple of years, we have worked towards integrating a variety of our contributions in the domain of Digital Humanities within stable platforms with the idea to make the corresponding environment reusable by other colleagues and institutions. This endeavor is in keeping with the general vision of EU infrastructures such as DARIAH to offer the humanities communities with the best up to date technologies for creating, managing, transforming and disseminating their data. We have put a specific emphasis on the stabilisation of a primary source management platform linking low level handwritten character recognition techniques with final end structured publication, with the following results:

• set up a technical environment within the collaborative project CREMMA to acquire and deploy a GPU based HTR server based upon the eScriptorium/Kraken suite. This environment, shared with the EPHE, LAMOP and Ecole Nationale des Chartes, now allows the latter for instance to provide a high quality environment for all their practical courses on digital epigraphy;
• work on a proposal for adapting the TEI standard to integrate all incoming information from HTR processes within a ready-to-publish document, which so far where only available in a scattered landscape of formats (ALTO XML with various versions, Page XML) 46. With a change of paradigm, where TEI is given more importance earlier in a digitisation workflow compared to traditional automatic transcription formats, we hope to improve the reusability of content at various stages, and hence ease the combination of different software solutions, such as those used for automatic transcription and edition. We also aim for a better propagation of metadata elements.
• Experiment a generic instance of the open source environment TEI Publisher for the online publication of editions integrating the abovementioned standardisation proposal and combining image, HTR results (based on an implementation of the SVG W3C recommendation), annotations and final end edited results. This work has been selected as part of the TEI Publisher Community Challenge.

8.1 Large corpus creation

Participants: Pedro Ortiz Suarez, Julien Abadji, Benoît Sagot.

Since the introduction of large language models in Natural Language Processing, large raw corpora have played a crucial role in Computational Linguistics. However, most of these large raw corpora are either available only for English or not available to the general public due to copyright issues. Nevertheless, there are some examples of freely available multilingual corpora for training Deep Learning NLP models, such as the Paracrawl corpora and our own large-scale multilingual corpus OSCAR 92.15 However, they have quality issues, especially for low-resource languages, an issue investigated in a large-scale experiment in which we have taken part and whose initial publication in 2021 54 will be followed by a publication in the Transactions of the Association for Computational Linguistics. Moreover, recreating or updating these corpora is very complex. In 2021 we have developed Ungoliant 18, a new pipeline that improves in multiple ways over the goclassy pipeline used to create the original OSCAR corpus (known as OSCAR v1 or OSCAR 2019). Ungoliant is faster, modular, parameterizable, and well documented. We used it to create a new version of OSCAR that is larger and based on recent data 18. Moreover, this new version of OSCAR (version 21.09) includes metadata information at the document level. Ungoliant is reseased under an open source license and publish the corpus under a research-only license. Since then, we have achieved a new set of improvements and automatic annotations in order to produce the latest version of OSCAR 51.

8.2 Neural language modelling

Participants: Benoît Sagot, Djamé Seddah, Rachel Bawden, Arij Riabi, Thomas Wang, Roman Castagné.

Pretrained language models are now ubiquitous in Natural Language Processing. Despite their success, most available models have either been trained on English data or on the concatenation of data in multiple languages 71, 85. This makes practical use of such models—in all languages except English—very limited. One of the most visible achievements of the ALMAnaCH team was the training and release of CamemBERT in 2019, a BERT-like 71 (and more specifically a RoBERTa-like) neural language model for French trained on the French section of our large-scale web-based OSCAR corpus 92 (see Section 8.1), together with CamemBERT variants 88 and ELMo models trained on OSCAR corpora for other languages, including French 90, 91.

In 2021, we investigated the impact on language modelling of shifting from token-based or subword-based models to character-based models, which were reported to improve model robustness under certain circumstances. We carried out a number of experiments on a number of such models proposed by other authors. In particular, we investigated the performance of character-based language models on North-African Arabizi (NArabizi), i.e. North-African colloquial dialectal Arabic written using an extension of the Latin script, a low-resource language variety on which ALMAnaCH has been working for several years. We show that a character-based model trained on only 99k sentences of NArabizi and fined-tuned on a small treebank of this language leads to performance close to those obtained with the same architecture pre-trained on large multilingual and monolingual models 27. We also confirmed these results a on much larger data set of noisy French user-generated content.

We also carried out a number of experiments on previously proposed and novel ways to approximate the attention mechanism in Transformer-based language models, which is quadratic, with a linear-time mechanism.

In collaboration with the startup LightOn, we trained and published PAGnol, a collection of GPT-like generative language models for French. Using scaling laws, we efficiently train PAGnol-XL (1.5B parameters) with the same computational budget as CamemBERT, which was 13 times smaller 56. PAGnol-XL is the largest model trained to date for the French language, and the first large-scale generative model for the language. Experiments shown that such a model, when trained on our large web-based corpus OSCAR (2019 version), can produce offensive content. It is one of the reasons that led us to develop detection tools to flag certain types of possibly offensive content in further versions of OSCAR.

Finally, a number of ALMAnaCH members have participated in the BigScience “workshop”,16 an informal 1-year-long international initiative led by the American company HuggingFace and involving 600 researchers from 50 countries and more than 250 institutions. The goal of the project, supported by the French state in the form of large grants of computing power on the Jean Zay public supercomputer, is to create a publicly available very large multilingual neural network language model and a publicly available very large multilingual text dataset. ALMAnaCH members are involved at all levels: working group chair, sub-working group chair, working group active member. First results have already been published. The working group on tokenisation, of which Benoît Sagot is one of the two chairs, has published a survey on the matter 58. The first language model of interest successfully trained and evaluated within the project has also resulted in a publication, to which some of ALMAnaCH's member involved in the project have contributed 29.

8.3 Cross-lingual transfer learning for low-resource non-standard languages

Participants: Djamé Seddah, Benoît Sagot, Benjamin Muller.

Building NLP systems for such highly variable and low-resource languages is a difficult challenge. The recent success of large-scale multilingual pretrained neural language models (including our CamemBERT language model for French) provides us with new modelling tools to tackle it. We have studied the ability of the multilingual version of BERT to model a non-formalised North-African dialect, Arabizi, written in Latin script and often code-mixed with French, which has been a research topic for the team for several years—in particular, we have published in 2019 a treebank of this language variety, the first UGC dataset for an Arabic-related language 111. We have shown in different scenarios that multilingual language models are able to transfer to such an unseen dialect, specifically in two extreme cases: across scripts (Arabic to Latin) and from Maltese, a related language written in the Arabic script, unseen during pretraining. The first results have led to fruitful collaborations with Yanai Elazar from Bar Illan University and Antonios Anastasopoulos from George Mason University. These works, expanding from North-African Arabic to many other languages that share similar properties, focus on the objective of understanding the inner workings of large multilingual neural language models when fine-tuned on cross-language scenarios with various degres of resource availability (from large to extremely scarce). Conducted in 2020, those works led to the publications of two papers in 2021 59, 25, following two publications in 2020 121, 115.

8.4 Multimodal word and sentence embeddings

Participants: Matthieu Futeral-Peter, Rachel Bawden, Benoît Sagot.

In collaboration with Ivan Laptev and Cordelia Schmid from the WILLOW project-team, we have investigated how using image data can improve the representation of words and sentences. As part of Matthieu Futeral-Peter’s master’s internship, we study how image captioning data can be used as a pivot to align multilingual embeddings, both on the word and the sentence level, notably exploring extensions to the GlobeTrotter model 117, such as the creation of synthetic data through automatic image captioning and the addition of a nearest neighbours search to expand the number of image-caption pairs. This work is a preliminary step in the path toward multimodal machine translation, which will be one of the main subjects of Matthieu’s PhD.

8.5 NLP for Early Modern French

Participants: Pedro Ortiz Suarez, Rachel Bawden, Benoît Sagot, Laurent Romary, Alix Chagué.

Early Modern French (also known as Modern French or classicial French) represents the French language from the 17th century. With the aim of helping philologists and literary experts study texts from this period, we have been collaborating with several researchers outside Inria (Simon Gabay from the University of Geneva and Philippe Gambette from Université Gustave-Eiffel) to develop various NLP tools adapted to Early Modern French. Aside the differences in topic and word choice, the texts display linguistic differences from contemporary French, including spelling differences, which encompass both typographic differences (such as the use of long s, which has become s in contemporary French) and those illustrative of linguistic change (estoit$\to$était) and classicial influences (sçauoir$\to$savoir). The NLP tools we develop must be adapted to these spelling differences and also be robust to variation, giving that no strict conventions were used during that period.

In 2021, we continued our work on various aspects of the processing of Early Modern French: historical spelling normalisation, named entity recognition, part-of-speech (PoS) tagging. These tasks rely on data from the period, and a major part of our efforts has been in the creation of adapted corpora: the FreEM corpus (short for French Early Modern), which includes a large monolingual corpus (FreEM-max), a smaller parallel corpus of sentences normalised into contemporary French spelling conventions (FreEm-norm) and a dataset annotated for locations 42.

From the FreEM-max data, we have also trained a RoBERTa language model 85 for Early Modern French, which we call D'AlemBERT, which will be used to boost the performance across various tasks. It has already be shown to boost the performance of PoS tagging for Early Modern French. From the FreEM-norm data, we have developed automatic normalisation models, including rule-based and machine-translation (MT)-style approaches. As well as providing normalisation, which can be a useful step before manual analysis or the application of other downstream tasks, we have exploited the normalisation models as a way to generate synthetic Early Modern French data on a large scale, which we then use to analyse the patterns of spelling change across different decades of the 17th century 41.

The NLP tools developed will be integrated into a processing pipeline for Early Modern French texts and encoded in TEI 32. Articles presenting the FreEM-max and FreEM-norm corpora, the D'AlemBERT language model (and results on PoS tagging) and the normalisation models 52 are currently under peer-review at the LREC 2022 conference.

8.6 Cognate prediction vs. low-resource Machine Translation

Participants: Clémentine Fourrier, Benoît Sagot, Rachel Bawden.

In 2021 we resumed our experiments to investigate whether and under which conditions neural networks can be used to learn sound correspondences between two related languages, i.e. for the prediction of cognates of source language words in a related target language. Since the experiments published in 73, we have improved both our way to extract data from our etymological database EtymDB 2.0 74 and our neural systems in multiple ways, leading to improved results published in 22.

More precisely, we investigated whether cognate prediction can benefit from insights from low-resource MT. We compared statistical MT (SMT) and neural MT (NMT) architectures in a bilingual setup, and studied the impact of employing data augmentation techniques commonly seen to give gains in low-resource MT: monolingual pretraining, backtranslation and multilinguality—on these techniques, see for instance 55, a survey on low-resource MT of which a member of the team is one of the main authors,17 as well as 19.

Our experiments on several Romance languages show that cognate prediction behaves only to a certain extent like a standard lowresource MT task. In particular, MT architectures, both statistical and neural, can be successfully used for the task, but using supplementary monolingual data is not always as beneficial as using additional language data, contrarily to what is observed for MT.

8.7 Machine Translation of non-standard texts

Participants: Djamé Seddah, Rachel Bawden, Jose Rosales Nunez, Camille Rey, Lydia Nishimwe, Sonal Sannigrahi.

Machine translation (MT) in an increasingly important research topic for ALMAnaCH, which has been reinforced since the arrival of Rachel Bawden as a Chargée de Recherches in November 2020. Consistent with the team's main research challenges, understanding the impact of language variation on MT models and trying to improve the robustness of these models to language variation are among our key research topics.

In the context of a PhD thesis co-supervised by Djamé Seddah with the Laboratoire de Linguistique Formelle (Université de Paris), we have taken a critical look at the automatic evaluation of user-generated content (UGC). TheWe have shown that measuring the average-case performance using a standard metric on a UGC test set falls far short of giving a reliable image of the UGC translation quality. We therefore introduced PMUMT,18 a new manually annotated dataset for the evaluation of UGC translation. We conducted several experiments on this dataset to measure the impact of different kinds of UGC specificities on translation quality, more precisely than previously possible 28.

Inspired by research on character-level MT and on robustness in character-level language models, we also explored the ability of character-based Neural MT to translate noisy UGC with a strong focus on exploring the limits of such approaches to handle productive UGC phenomena, which cannot be seen at training time. Within a strict zero-shot scenario, we studied the detrimental impact on translation performance of various user-generated content phenomena on a small annotated dataset we developed and showed that such models are indeed incapable of handling unknown characters, which leads to catastrophic translation failure once such characters are encountered. We also showed the importance of reducing the vocabulary size hyper-parameter to increase the robustness of character-based models for machine translation 26. On this topic and in the context of an internship supervised by Rachel Bawden, we are also investigating the robustness of character-level and byte-level models for multilingual MT, whereby we will apply various perturbations to input representations (e.g. transliteration, character-shifting) to compare the impact of lexical sharing in the performance of these models.

More recently, we have developed a novel Variational Neural MT (VNMT) architecture with enhanced robustness properties, which we investigated through a detailed case-study addressing noisy French user-generated content (UGC) translation to English. The proposed model, with results comparable or superior to state-of-the-art VNMT, improves performance over UGC translation in a zero-shot evaluation scenario while keeping optimal translation scores on in-domain test sets. This work echoes the work of a PhD student at LIPN on the application and extension of Variational Auto-Encoders, with whom we collaborate 21, 40.

The arrival of Rachel Bawden at the end of 2020, a specialist of MT and specifically low-resource and robust MT, who published in 2021 a survey paper on low-resource MT 55,19 followed by the arrival two new young researchers at the end of 2021, a PhD student and a Master's student, has reinforced the team's interest in these questions and, more generally, on robustness and domain adaptation in MT. In particular, we have started the design and development of a new corpus of English UGC where tweets of varied non-standardness levels are aligned with a normalised version and a (standard) translation in French, in order to submit a new shared task for the WMT 2022 conference.

8.8 New results on text simplification

Participants: Benoît Sagot, Éric Villemonte de La Clergerie, Louis Martin.

The aim of text simplification (TS) is to make a text easier to read and understand by simplifying its grammar and structure while keeping the underlying meaning and information identical. It is therefore an instance of language variation, based on language complexity. It can benefit numerous audiences, such as people with disabilities, language learners and even the general public, for instance when dealing with intrinsically complex texts such as legal documents.

In 2017 we initiated a collaboration with the Facebook Artificial Intelligence Research (FAIR) lab in Paris and with the UNAPEI, the federation of French associations helping people with mental disabilities and their families. The objective of this collaboration is to develop tools to help the simplification of texts aimed at mentally disabled people. More precisely, the aim is to develop a computer-assisted text simplification platform (as opposed to an automatic TS system). In this context, a CIFRE PhD thesis was initiated in collaboration with the FAIR on the TS task, Louis Martin being the PhD student involved.

In 2020 and 2021, this PhD thesis extended the scope of TS to multiple languages using a totally unsupervised method and with state-of-the-art results, even compared to previous supervised models 57. We use a large-scale mining pipeline to extract one billion sentences from the web using Common Crawl. These sentences are used to find millions of paraphrases in three languages: English, French and Spanish. This allows us to train our controllable models 89 in multiple languages, which we then adapt for the TS task. These models push the state-of-the-art further in all three languages and benchmarks. Generated simplifications are considered more fluent, and simpler than previous models.

In 2021, we also developed a French-specific simplification model based on our pre-trained language model CamemBERT 4 and compared its results with the above-mentioned unsupervised models and weakly supervised variants thereof. We also investigated a new way to address the issue of the evaluation of sentence simplification, which had been the focus of previous efforts and publications 87. This new method is based on previous work on the evaluation of automatic summarisation models based on the ability of the models output to successfully allow for a question answering model to answer questions the input text would allow to answer 109. Adapting this approach to the evaluation of sentence simplification proved successful 60.

This research enables us to exploit our previous works and apply them to the Cap'FALC project, whose aim is to assist the simplification of French documents by people with intellectual disabilities.

2021 was also an important year for this line of research, since Louis Martin defended his PhD in September. The Cap'FALC project, however, continues, and we hope to resume working on this socially important and scientifically challenging topic.

8.9 Threads Constitution

Participants: Lionel Tadjou Tadonfouet, Éric Villemonte de La Clergerie, Laurent Romary.

In the context of a CIFRE PhD with Orange, we have developed a corporate corpus that will be used as a reference for modelling and computing threads from conversations generated using communication and collaboration tools. The overall goal of the reconstruction of threads is to higlight the important parts of a running discussion, reviewing the upcoming commitments or deadlines, etc. Since, to our knowledge, there were no available corporate corpus for the French language, we developed a method for building such corpora, including a pseudo-anonymisation step, in compliance with the GDPR 30.

8.10 Hate speech detection

Participants: Djamé Seddah, Syrielle Montariol, Arij Riabi.

In order to support the fight against radicalization and thus prevent future terrorist attacks from taking place, the CounteR H2020 project brings data from disperse sources into an analysis and early alert platform for data mining and prediction of critical areas (e.g. communities), aiming to be a frontline community policing tool which looks at the community and its related risk factors rather than targeting and surveilling individuals. ALMAnaCH is responsible of the most part of the NLP component of the project. However, issues regarding the availability of the input data have prevented us from working specifically on CounteR data. We have therefore focused on a task of huge social importance that is similar to CounteR's needs, namely hate speech detection, with a focus on zero-shot cross-lingual transfer.

In our work in this direction, unpublished to date, we investigate a key limitation of cross-lingual transfer approaches to such tasks, namely the fact that they involve numerous linguistic and cultural differences and discrepencies from one language or one country to the other. We investigated how training on multilingual auxiliary tasks (sentiment analysis, named entity recognition, and tasks relying on syntactic information) impacts the zero-shot transfer of hate speech detection models across languages. We showed the positive impact of these tasks, particularly named entity recognition, for bridging the gap between languages. But we also shown that, in some cases, the language model training data prevents hate speech detection models from benefiting from the knowledge proxy brought by auxiliary tasks fine-tuning.

8.11 Similar case detection for the Cour de Cassation

Participants: Thibault Charmet, Rachel Bawden, Benoît Sagot.

As part of our LabIA project, funded by the DINUM, in collaboration with the Cour de Cassation, we have developed tools for (i) the automatic generation of keyword sequences (titrages), which are currently used by experts at the Cour to detect rulings that are similar in their application of the law, and (ii) the development of various similarity measures, which can be used to find similar documents. The similarity measures we use also integrate the keyword sequences predicted by our models, with improved correlations to expert judgments of similarity. This is important because it enables us to increase the coverage of these keyword sequences, which are currently only available for about 20% of all rulings, and to increase the number of sequences per ruling, since the granularity of the sequences and the word choice is highly variable, which otherwise limits the recall of a similarity search process. To validate results, similarity measures are compared against similarity judgements between pairs of keyword sequences that we manually collected with experts from the Cour. The motivations, methods and results have been submitted for peer review at the LREC 2022 conference.

8.12 Models for the representation of lexical content

Participants: Laurent Romary.

For several years, the ALMAnaCH team has taken a leadership role in defining standards for representing lexical content, either as a result of digitising legacy dictionaries or through the creation of new lexical resources serving as a basis for computational linguistics processes.

The bridge between the general ISO standardisation work and the community-based TEI guidelines has been further developed in the context of the DARIAH working group on lexical resources, whose objective consists in identifying univocal TEI-based constructs for a variety of lexical features encountered in machine readable dictionaries. The work carried out by the group (see web site), whose quality was recognised by the TEI community as a whole with the Rahtz Prize for TEI Ingenuity, has been and implemented in several major dictionary endeavours worldwide as exemplified in 79 and, in 2021, 20. In the context of the ELEXIS project, we also took part in the publication of a reference document on best practices in electronic lexicographic resource development, the Lexicographic Data Seal of Compliance 61.

Regarding our long-lasting efforts on the encoding of etymological information in the TEI has resulted in the publication of a description of TEI Lex-0 Etym, the etymological component of the TEI Lex-0 initiative, which aims at defining a terser subset of the TEI guidelines for the representation of etymological features in dictionary entries 53. Going beyond the basic provision of etymological mechanisms in the TEI guidelines, TEI Lex-0 Etym proposes a systematic representation of etymological and cognate descriptions by means of embedded constructs based on the (for etymologies) and (for etymons and cognates) elements. In particular, given that all the potential contents of etymons are highly analogous to those of dictionary entries in general, the contents presented herein heavily re-use many of the corresponding features and constraints introduced in other components of the TEI Lex-0 to the encoding of etymologies and etymons. The TEI Lex-0 Etym model is also closely aligned to ISO 24613-3 on modelling etymological data and the corresponding TEI serialisation available in ISO 24613-4.

8.13 Information extraction from specialised collections

Participants: Laurent Romary, Alix Chagué, Floriane Chiffoleau, Lucas Terriel, Hugo Scheithauer, Tanti Kristanti, Yves Tadjo.

Building up on our long-standing contribution to the development of the GROBID suite, the ALMAnaCH team has pursued its effort in extending the scope and performance of the GROBID modules in relation to various ongoing collaboration schemes.

In the continuity of Mohamed Khemakhem's PhD thesis 82, which has demonstrated the possibility to parse lexical entries from legacy dictionary content in a very fine-grained way, we investigated the applicability of the same architecture on catalogue-like textual objects, as previsouly demonstrated in 81 and in 2021 in 23 on the example of exhibition catalogues. The DataCatalog project, jointly led with the Bibliothèque nationale de France (BnF) and the Institut national d’histoire de l’art (INHA), was launched in late 2021. It aims to develop a GROBID module for automatically structuring catalogues in TEI-XML, and to publish files produced with it on an interface and allowing visualizations and queries on segmented zones. We hope to give researchers a new way to access these documents where accessing the information is made easier and more accurate. The BnF and the INHA provide a vast collection of sale catalogues on which the custom TEI model created for the project is based.

The occurrence of the COVID pandemic in Spring 2020 has also been an opportunity to start a project with the Parisian hospital network (APHP). One of the goals of the project (and the action that ended up being the main focus of our attention) was to see how the GROBID suite could be further expanded to parse the variety of medical reports and documents associated with a patient so that doctors can trace the precise relevant information (anamnesis, symptoms, treatments etc.). The excellent results obtained in 2020 have led us to build up a stable collaboration with the APHP to make the GROBID workflow an essential building block of their document processing chain.

Following the provision of a direct grant from the Ministry of Higher Education and Research (MESRI - DAHN project) and benefiting from a collaborative framework with the French National Archives (Lectaurep project 33, 34), we have explored how our experience in extracting information from legacy documents could be made part of wider understanding of the components of a generic digitisation workflow of documents initially available as images 39, 36. In particular, in collaboration with colleagues from the ÉPHÉ (École Pratique des Hautes Études), we contributed to the further development of the eScriptorium platform, which aims to provide an annotation, training and transcription environment for handwritten text recognition (HTR) tasks, for any language and any writing system, down to the understanding of how to encode the data at hand in the TEI format 46 and publish it in the TEI Publisher environment 47.

We have made contributions in a number of complementary directions, including the following:

• We have defined methods for the coherent management and sharing of ground truth, i.e reference annotated data with potential interest across a variety of time periods, languages, writing systems and domains. To this end, HTR-United (web site) aims at gathering various annotated corpora along with their metadata for the creation of HTR models 37;
• We have worked on the development of several refined HTR models and carefully evaluated its performance levels thanks to KaMI (Kraken as Model Inspector) 45, 48;
• We have worked in the context of the NER4archives project on semi-automatic methods and tools to recognise named entities in archival research instruments encoded in XML/EAD 38;
• We have worked on the automatic transcription and edition of digitised sources on work in the textile industry 43; and of the digitised WWI correspondance of Paul d'Estournelles de Constant 39. We have also updated and complemented the encoding and edition of several corporas including the Berliner Intelektuelle and testimonies from the Holocaust (web site).

9.1 Bilateral contracts with industry

Ongoing contracts:

• Verbatim Analysis
  Verbatim Analysis is an Inria start-up co-created in 2009 by Benoît Sagot. It uses some of ALMAnaCH's free NLP software (SxPipe) as well as a data mining solution co-developed by Benoît Sagot, VERA, for processing employee surveys with a focus on answers to open-ended questions.
• opensquare
  was co-created in December 2016 by Benoît Sagot with 2 senior specialists of HR (human resources) consulting. It is dedicated to designing, carrying out and analysing employee surveys as well as HR consulting based on these results. It uses a new employee survey analysis tool, enqi, which is still under development. This tool being co-owned by opensquare and Inria, both parties have signed a Software Licence Agreement in exchange for a yearly fee paid by opensquare to ALMAnaCH based on its turnover. Benoît Sagot currently contributes to opensquare, under the “Concours scientifique” scheme.
• Facebook

  A collaboration on text simplification (“français Facile À Lire et à Comprendre”, FALC) is ongoing with Facebook's Parisian FAIR laboratory. It involved a co-supervised (CIFRE) PhD thesis in collaboration with UNAPEI, the largest French federation of associations defending and supporting people with special needs and their families. The PhD thesis was defended in 2021. This collaboration is part of a larger initiative called Cap'FALC involving (at least) these three partners as well as the relevant ministries. Funding received as a consequence of the CIFRE PhD thesis: 60,000 euros

  Moreover, two new CIFRE PhD theses on other topics (1. language modelling applied to speech conversational data; 2. sentence-level vector representations) have started in 2021.

• Orange
• Winespace
  The collaboration with this start-up company, dedicated to information extraction from wine descriptions to develop a wine recommendation system, was carried out in 2020 following previous discussions, in collaboration with Inria Bordeaux's “InriaTech” structure. In 2021, we designed and prepared a second step for this collaboration, which will involve a dedicated research engineer who will begin their work in 2022. We also presented the results of the collaboration's first step at a regional data science event, Dataquitaine 31.
• INPI
  A collaboration with the Institut National de la Propriété Industrielle (France's patent office) started in 2021. A research engineer was hired for a one-year contract to work on patent classification. This project informally collaborates with the qatent startup, on which see below.
• Cour de cassation
  A LabIA project started in early 2021. A research engineer was hired for a one-year contract to work on the automatic analysis of Cour de Cassation rulings and on the automatic assessment of the similarity between two rulings.

9.2 Active collaborations without a contract

• Science Miner
  ALMAnaCH (following ALPAGE) has collaborated since 2014 with this company founded by Patrice Lopez, a specialist in machine learning techniques and initiator of the Grobid and NERD (now entity-fishing) suites. Patrice Lopez provides scientific support for the corresponding software components in the context of the Parthenos, EHRI and Iperion projects, as well as in the context of the Inria anHALytics initiative, aiming to provide a scholarly dashboard on scientific papers available from the HAL national publication repository.
• qatent
  The startup company qatent, dedicated to computer-aided patent creation, was created in 2021. It is supported by the Inria Startup Studio and by the ALMAnaCH team. Regular interactions take place between qatent founders and members on the one hand and ALMAnaCH members on the other hand, in particular those involved in the collaboration with INPI. This creates a stimulating informal collaboration between all three entities around NLP for patents, which might foster further activity in this domain (e.g. a future PhD thesis).
• LightON
  LightON builds Optical Processor Units, a specialized line of processor able to outperform GPU on certains tasks. We're working with them to see if we can use their technology to speed up the training of large language models. We recently submitted a grant proposal to access 250k gpu hours on Jean Zay in order to scale their algorithms. This informal collaboration with this company has already resulted in the design, training and publication of the PAGnol generative language model for French (cf. PAGnol's web site).
• AXA ReV
  AXA ReV is the R&D Lab of the Axa Insurance group, located in Paris. This collaboration focuses on neural models interpretability and establishing "explainable" benchmarks as a end-goal for research on question answering.

10.1 European initiatives

10.2 National initiatives

10.3 Regional initiatives

11.1 Promoting scientific activities

11.2 Teaching - Supervision - Juries

11.3 Popularization

12.1 Major publications

• 1 articleD.Darja Fišer and B.Benoît Sagot. Constructing a poor man's wordnet in a resource-rich world.Language Resources and Evaluation4932015, 601-635
  HALDOIback to text
• 2 inproceedings G.Ganesh Jawahar, B.Benoît Sagot and D.Djamé Seddah. What does BERT learn about the structure of language? ACL 2019 - 57th Annual Meeting of the Association for Computational Linguistics Florence, Italy July 2019
  HAL
• 3 inproceedingsP.Patrice Lopez and L.Laurent Romary. HUMB: Automatic Key Term Extraction from Scientific Articles in GROBID.SemEval 2010 WorkshopACL SigLex eventUppsala, SwedenJuly 2010, 248-251
  HAL
• 4 inproceedingsL.Louis Martin, B.Benjamin Muller, P. J.Pedro Javier Ortiz Suárez, Y.Yoann Dupont, L.Laurent Romary, É.Éric Villemonte de La Clergerie, D.Djamé Seddah and B.Benoît Sagot. CamemBERT: a Tasty French Language Model.ACL 2020 - 58th Annual Meeting of the Association for Computational LinguisticsSeattle / Virtual, United StatesJuly 2020
  HALDOIback to text
• 5 inproceedingsP. J.Pedro Javier Ortiz Suárez, B.Benoît Sagot and L.Laurent Romary. Asynchronous Pipeline for Processing Huge Corpora on Medium to Low Resource Infrastructures.7th Workshop on the Challenges in the Management of Large Corpora (CMLC-7)Cardiff, United KingdomLeibniz-Institut für Deutsche SpracheJuly 2019
  HALDOI
• 6 inproceedingsC.Corentin Ribeyre, É.Éric Villemonte de La Clergerie and D.Djamé Seddah. Because Syntax does Matter: Improving Predicate-Argument Structures Parsing Using Syntactic Features.Conference of the North American Chapter of the Association for Computational Linguistics: Human Language TechnologiesProceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language TechnologiesDenver, USA, United StatesJune 2015
  HAL
• 7 articleL.Laurent Romary. TEI and LMF crosswalks.JLCL - Journal for Language Technology and Computational Linguistics3012015
  HAL
• 8 inproceedingsB.Benoît Sagot. The Lefff, a freely available and large-coverage morphological and syntactic lexicon for French.7th international conference on Language Resources and Evaluation (LREC 2010)Valletta, MaltaMay 2010
  HALback to text
• 9 inproceedingsB.Benoît Sagot and É.Éric Villemonte de La Clergerie. Error mining in parsing results.The 21st International Conference of the Association for Computational Linguistics (ACL 2006)Sydney, AustraliaJuly 2006, 329-336
  HALback to text
• 10 inproceedingsD.Djamé Seddah, B.Benoît Sagot, M.Marie Candito, V.Virginie Mouilleron and V.Vanessa Combet. The French Social Media Bank: a Treebank of Noisy User Generated Content.COLING 2012 - 24th International Conference on Computational LinguisticsKay, Martin and Boitet, ChristianMumbai, IndeDecember 2012, URL: http://hal.inria.fr/hal-00780895
• 11 inproceedingsR.Reut Tsarfaty, D.Djamé Seddah, Y.Yoav Goldberg, S.Sandra Kübler, Y.Yannick Versley, M.Marie Candito, J.Jennifer Foster, I.Ines Rehbein and L.Lamia Tounsi. Statistical Parsing of Morphologically Rich Languages (SPMRL) What, How and Whither.Proceedings of the NAACL HLT 2010 First Workshop on Statistical Parsing of Morphologically-Rich LanguagesÉtats-Unis Los AngelesAssociation for Computational Linguistics2010, 1--12
• 12 articleR.Reut Tsarfaty, D.Djamé Seddah, S.Sandra Kübler and J.Joakim Nivre. Parsing Morphologically Rich Languages: Introduction to the Special Issue.Computational Linguistics391March 2013, 8
  HALDOI
• 13 inproceedingsÉ.Éric Villemonte de La Clergerie. Improving a symbolic parser through partially supervised learning.The 13th International Conference on Parsing Technologies (IWPT)Naria, JapanNovember 2013
  HAL

12.2 Publications of the year

12.3 Other

12.4 Cited publications

• 64 articleM. J.María Jesús Aranzabe, A. D.Arantza Díaz De Ilarraza and I.Itziar Gonzalez-Dios. Transforming complex sentences using dependency trees for automatic text simplification in Basque.Procesamiento del lenguaje natural502013, 61--68
  back to text
• 65 inproceedingsE. M.Emily M. Bender, T.Timnit Gebru, A.Angelina McMillan-Major and S.Shmargaret Shmitchell. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?Proceedings of the 2021 ACM Conference on Fairness, Accountability, and TransparencyFAccT '21New York, NY, USAVirtual Event, CanadaAssociation for Computing Machinery2021, 610–623URL: https://doi.org/10.1145/3442188.3445922
  DOIback to text
• 66 articleO.Olivier Bonami and B.Benôit Sagot. Computational methods for descriptive and theoretical morphology: a brief introduction.Morphology2742017, 1-7
  HALDOIback to text
• 67 articleA.A. Bouchard-Côté, D.D. Hall, T.T.L. Griffiths and D.D. Klein. Automated Reconstruction of Ancient Languages using Probabilistic Models of Sound Change.Proceedings of the National Academy of Sciences1102013, 4224--4229
  back to text
• 68 inproceedingsJ. C.Jackie Chi Kit Cheung and G.Gerald Penn. Utilizing Extra-sentential Context for Parsing.Proceedings of the 2010 Conference on Empirical Methods in Natural Language ProcessingEMNLP '10Cambridge, Massachusetts2010, 23--33
  back to text
• 69 inproceedingsM.Matthieu Constant, M.Marie Candito and D.Djamé Seddah. The LIGM-Alpage Architecture for the SPMRL 2013 Shared Task: Multiword Expression Analysis and Dependency Parsing.Fourth Workshop on Statistical Parsing of Morphologically Rich LanguagesSeattle, United StatesOctober 2013, 46-52
  HALback to text
• 70 inproceedingsS.Spencer Desrochers, C.Chad Paradis and V. M.Vincent M. Weaver. A Validation of DRAM RAPL Power Measurements.Proceedings of the Second International Symposium on Memory SystemsMEMSYS '16New York, NY, USAAlexandria, VA, USAAssociation for Computing Machinery2016, 455–470URL: https://doi.org/10.1145/2989081.2989088
  DOIback to text
• 71 inproceedingsJ.Jacob Devlin, M.-W.Ming-Wei Chang, K.Kenton Lee and K.Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers)2019, 4171--4186URL: https://www.aclweb.org/anthology/N19-1423/
  back to textback to text
• 72 articleY.Yuan Fang and M.-W.Ming-Wei Chang. Entity Linking on Microblogs with Spatial and Temporal Signals.TACL22014, 259--272URL: https://tacl2013.cs.columbia.edu/ojs/index.php/tacl/article/view/323
  back to text
• 73 inproceedingsC.Clémentine Fourrier and B.Benôit Sagot. Comparing Statistical and Neural Models for Learning Sound Correspondences.LT4HALA 2020 : First Workshop on Language Technologies for Historical and Ancient LanguagesDue to the COVID-19 pandemic, the workshop will not take place. However, the proceedings are published online.Marseille, FranceMay 2020
  HALback to text
• 74 inproceedingsC.Clémentine Fourrier and B.Benôit Sagot. Methodological Aspects of Developing and Managing an Etymological Lexical Resource: Introducing EtymDB 2.0.LREC 2020 - 12th Language Resources and Evaluation ConferenceDue to COVID19 pandemic, the 12th edition is cancelled. The LREC 2020 Proceedings are available at http://www.lrec-conf.org/proceedings/lrec2020/index.htmlMarseille, FranceMay 2020
  HALback to text
• 75 articleJ. E.James E Hoard, R.Richard Wojcik and K.Katherina Holzhauser. An automated grammar and style checker for writers of Simplified English.Computers and Writing: State of the Art1992, 278--296
  back to text
• 76 inproceedingsD.Dirk Hovy and T.Tommaso Fornaciari. Increasing In-Class Similarity by Retrofitting Embeddings with Demographic Information.Proceedings of the 2018 Conference on Empirical Methods in Natural Language ProcessingBrussels, BelgiumAssociation for Computational Linguistics2018, 671--677URL: http://aclweb.org/anthology/D18-1070
  back to text
• 77 articleD.D.J. Hruschka, S.S. Branford, E.E.D. Smith, J.J. Wilkins, A.A. Meade, M.M. Pagel and T.T. Bhattacharya. Detecting Regular Sound Changes in Linguistics as Events of Concerted Evolution.Current Biology1252015, 1--9
  back to text
• 78 inproceedingsG.Ganesh Jawahar, B.Benjamin Muller, A.Amal Fethi, L.Louis Martin, É.Éric Villemonte de La Clergerie, B.Benôit Sagot and D.Djamé Seddah. ELMoLex: Connecting ELMo and Lexicon features for Dependency Parsing.CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal DependenciesBrussels, BelgiumOctober 2018
  HALDOIback to text
• 79 inproceedingsF.Fahad Khan, L.Laurent Romary, A.Ana Salgado, J.Jack Bowers, M.Mohamed Khemakhem and T.Toma Tasovac. Modelling Etymology in LMF/TEI: The Grande Dicionário Houaiss da Língua Portuguesa Dictionary as a Use Case.LREC 2020 - 12th Language Resources and Evaluation ConferenceDue to COVID19 pandemic, the 12th edition is cancelled. The LREC 2020 Proceedings are available at http://www.lrec-conf.org/proceedings/lrec2020/index.htmlMarseille, FranceMay 2020
  HALback to text
• 80 inproceedingsM.Mohamed Khemakhem, L.Luca Foppiano and L.Laurent Romary. Automatic Extraction of TEI Structures in Digitized Lexical Resources using Conditional Random Fields.electronic lexicography, eLex 2017Leiden, NetherlandsSeptember 2017
  HALback to textback to text
• 81 inproceedingsM.Mohamed Khemakhem, S.Simon Gabay, B.Béatrice Joyeux-Prunel, L.Laurent Romary, L.Léa Saint-Raymond and L.Lucie Rondeau Du Noyer. Information Extraction Workflow for Digitised Entry-based Documents.DARIAH Annual event 2020Zagreb / Virtual, CroatiaMay 2020
  HALback to text
• 82 phdthesisM.Mohamed Khemakhem. Standard-based Lexical Models for Automatically Structured Dictionaries.Université de ParisOctober 2020
  HALback to text
• 83 inproceedingsS.Sandra Kübler, M.Matthias Scheutz, E.Eric Baucom and R.Ross Israel. Adding Context Information to Part Of Speech Tagging for Dialogues.NEALT Proceedings Series92010, 115-126
  back to text
• 84 articleA.-L.Anne-Laure Ligozat, C.Cyril Grouin, A.Anne Garcia-Fernandez and D.Delphine Bernhard. Approches à base de fréquences pour la simplification lexicale.TALN-RÉCITAL 20132013, 493
  back to text
• 85 articleY.Yinhan Liu, M.Myle Ott, N.Naman Goyal, J.Jingfei Du, M.Mandar Joshi, D.Danqi Chen, O.Omer Levy, M.Mike Lewis, L.Luke Zettlemoyer and V.Veselin Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach.arXiv preprint arXiv:1907.116922019
  back to textback to text
• 86 inproceedingsH.Héctor Mart\'inez Alonso, D.Djamé Seddah and B.Benôit Sagot. From Noisy Questions to Minecraft Texts: Annotation Challenges in Extreme Syntax Scenarios.2nd Workshop on Noisy User-generated Text (W-NUT) at CoLing 2016Osaka, JapanDecember 2016
  HALback to text
• 87 inproceedingsL.Louis Martin, S.Samuel Humeau, P.-E.Pierre-Emmanuel Mazaré, A.Antoine Bordes, É.Éric Villemonte de La Clergerie and B.Benôit Sagot. Reference-less Quality Estimation of Text Simplification Systems.1st Workshop on Automatic Text Adaptation (ATA)Tilburg, NetherlandsNovember 2018
  HALback to text
• 88 unpublishedL.Louis Martin, B.Benjamin Muller, P. J.Pedro Javier Ortiz Suárez, Y.Yoann Dupont, L.Laurent Romary, É.Éric Villemonte de La Clergerie, D.Djamé Seddah and B.Benôit Sagot. CamemBERT: a Tasty French Language Model.October 2019, Web site: https://camembert-model.fr
  HALback to text
• 89 unpublishedL.Louis Martin, B.Benôit Sagot, É.Éric Villemonte de La Clergerie and A.Antoine Bordes. Controllable Sentence Simplification.October 2019, Code and models: https://github.com/facebookresearch/access
  HALback to text
• 90 inproceedingsP. J.Pedro Javier Ortiz Suárez, Y.Yoann Dupont, B.Benjamin Muller, L.Laurent Romary and B.Benôit Sagot. Establishing a New State-of-the-Art for French Named Entity Recognition.LREC 2020 - 12th Language Resources and Evaluation ConferenceDue to COVID19 pandemic, the 12th edition is cancelled. The LREC 2020 Proceedings are available at http://www.lrec-conf.org/proceedings/lrec2020/index.htmlMarseille, FranceMay 2020
  HALback to text
• 91 inproceedingsP. J.Pedro Javier Ortiz Suárez, L.Laurent Romary and B.Benôit Sagot. A Monolingual Approach to Contextualized Word Embeddings for Mid-Resource Languages.ACL 2020 - 58th Annual Meeting of the Association for Computational LinguisticsSeattle / Virtual, United StatesJuly 2020
  HALDOIback to text
• 92 inproceedingsP. J.Pedro Javier Ortiz Suárez, B.Benôit Sagot and L.Laurent Romary. Asynchronous Pipeline for Processing Huge Corpora on Medium to Low Resource Infrastructures.7th Workshop on the Challenges in the Management of Large Corpora (CMLC-7)Cardiff, United KingdomLeibniz-Institut für Deutsche SpracheJuly 2019
  HALDOIback to textback to text
• 93 phdthesisJ.J. Pyssalo. System PIE: the Primary Phoneme Inventory and Sound Law System for Proto-Indo-European.University of Helsinki2013
  back to text
• 94 inproceedingsL.Luz Rello, R.Ricardo Baeza-Yates, S.Stefan Bott and H.Horacio Saggion. Simplify or help?: text simplification strategies for people with dyslexia.Proceedings of the 10th International Cross-Disciplinary Conference on Web AccessibilityACM2013, 15
  back to text
• 95 inproceedingsL.Luz Rello, R.Ricardo Baeza-Yates, L.Laura Dempere-Marco and H.Horacio Saggion. Frequent words improve readability and short words improve understandability for people with dyslexia.IFIP Conference on Human-Computer InteractionSpringer2013, 203--219
  back to text
• 96 inproceedingsC.Corentin Ribeyre, M.Marie Candito and D.Djamé Seddah. Semi-Automatic Deep Syntactic Annotations of the French Treebank.The 13th International Workshop on Treebanks and Linguistic Theories (TLT13)Proceedings of TLT 13Tübingen UniversitätTübingen, GermanyDecember 2014
  HALback to text
• 97 inproceedingsL.Laurent Romary, M.Mohamed Khemakhem, F.Fahad Khan, J.Jack Bowers, N.Nicoletta Calzolari, M.Monte George, M.Mandy Pet and P.Piotr Bański. LMF Reloaded.AsiaLex 2019: Past, Present and FutureIstanbul, TurkeyJune 2019
  HALback to text
• 98 inproceedingsA. M.Alexander M. Rush, R.Roi Reichart, M.Michael Collins and A.Amir Globerson. Improved Parsing and POS Tagging Using Inter-sentence Consistency Constraints.Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language LearningEMNLP-CoNLL '12Jeju Island, Korea2012, 1434--1444
  back to text
• 99 inproceedingsA.Anders S\ogaard and Y.Yoav Goldberg. Deep multi-task learning with low level tasks supervised at lower layers.Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)Berlin, Germany2016, 231--235
  back to text
• 100 inproceedingsB.Benôit Sagot. DeLex, a freely-avaible, large-scale and linguistically grounded morphological lexicon for German.Language Resources and Evaluation ConferenceEuropean Language Resources AssociationReykjavik, IcelandMay 2014
  HALback to text
• 101 techreportB.Benôit Sagot. External Lexical Information for Multilingual Part-of-Speech Tagging.RR-8924Inria ParisJune 2016
  HALback to text
• 102 inproceedingsB.Benôit Sagot. Extracting an Etymological Database from Wiktionary.Electronic Lexicography in the 21st century (eLex 2017)Leiden, NetherlandsSeptember 2017, 716-728
  HALback to textback to text
• 103 inproceedingsB.Benôit Sagot and H.Héctor Mart\'inez Alonso. Improving neural tagging with lexical information.15th International Conference on Parsing TechnologiesPisa, ItalySeptember 2017, 25-31
  HALback to text
• 104 inproceedingsB.Benôit Sagot, D.Damien Nouvel, V.Virginie Mouilleron and M.Marion Baranes. Extension dynamique de lexiques morphologiques pour le français à partir d'un flux textuel.TALN - Traitement Automatique du Langage NaturelLes sables d'Olonne, FranceJune 2013, 407-420
  HALback to text
• 105 inproceedingsC.Carolina Scarton, M.Matheus De Oliveira, A.Arnaldo Candido Jr, C.Caroline Gasperin and S. M.Sandra Maria Aluísio. SIMPLIFICA: a tool for authoring simplified texts in Brazilian Portuguese guided by readability assessments.Proceedings of the NAACL HLT 2010 Demonstration SessionAssociation for Computational Linguistics2010, 41--44
  back to text
• 106 inproceedingsY.Yves Scherrer and B.Benôit Sagot. A language-independent and fully unsupervised approach to lexicon induction and part-of-speech tagging for closely related languages.Language Resources and Evaluation ConferenceEuropean Language Resources AssociationReykjavik, IcelandMay 2014
  HALback to text
• 107 inproceedingsS.Sebastian Schuster, É.Éric Villemonte de La Clergerie, M.Marie Candito, B.Benôit Sagot, C. D.Christopher D. Manning and D.Djamé Seddah. Paris and Stanford at EPE 2017: Downstream Evaluation of Graph-based Dependency Representations. EPE 2017 - The First Shared Task on Extrinsic Parser EvaluationProceedings of the 2017 Shared Task on Extrinsic Parser EvaluationPisa, ItalySeptember 2017, 47-59
  HALback to textback to text
• 108 articleR.Roy Schwartz, J.Jesse Dodge, N. A.Noah A. Smith and O.Oren Etzioni. Green AI.Commun. ACM6312nov 2020, 54–63URL: https://doi.org/10.1145/3381831
  DOIback to text
• 109 inproceedingsT.Thomas Scialom, P.-A.Paul-Alexis Dray, S.Sylvain Lamprier, B.Benjamin Piwowarski, J.Jacopo Staiano, A.Alex Wang and P.Patrick Gallinari. QuestEval: Summarization Asks for Fact-based Evaluation.Proceedings of the 2021 Conference on Empirical Methods in Natural Language ProcessingOnline and Punta Cana, Dominican RepublicAssociation for Computational LinguisticsNovember 2021, 6594--6604URL: https://aclanthology.org/2021.emnlp-main.529
  DOIback to text
• 110 inproceedingsD.Djamé Seddah and M.Marie Candito. Hard Time Parsing Questions: Building a QuestionBank for French.Tenth International Conference on Language Resources and Evaluation (LREC 2016)Proceedings of the 10th edition of the Language Resources and Evaluation Conference (LREC 2016)Portoroż, SloveniaMay 2016
  HALback to text
• 111 inproceedingsD.Djamé Seddah, F.Farah Essaidi, A.Amal Fethi, M.Matthieu Futeral, B.Benjamin Muller, P. J.Pedro Javier Ortiz Suárez, B.Benôit Sagot and A.Abhishek Srivastava. Building a User-Generated Content North-African Arabizi Treebank: Tackling Hell.ACL 2020 - 58th Annual Meeting of the Association for Computational LinguisticsSeattle / Virtual, CanadaJuly 2020
  HALDOIback to text
• 112 inproceedingsD.Djamé Seddah, B.Benôit Sagot, M.Marie Candito, V.Virginie Mouilleron and V.Vanessa Combet. The French Social Media Bank: a Treebank of Noisy User Generated Content.COLING 2012 - 24th International Conference on Computational LinguisticsKay, Martin and Boitet, ChristianMumbai, IndiaDecember 2012
  HALback to text
• 113 inproceedingsD.Djamé Seddah, B.Benôit Sagot and M.Marie Candito. The Alpage Architecture at the SANCL 2012 Shared Task: Robust Pre-Processing and Lexical Bridging for User-Generated Content Parsing.SANCL 2012 - First Workshop on Syntactic Analysis of Non-Canonical Language , an NAACL-HLT'12 workshopMontréal, CanadaJune 2012
  HALback to text
• 114 articleM.Matthew Shardlow. A survey of automated text simplification.International Journal of Advanced Computer Science and Applications412014, 58--70
  back to text
• 115 inproceedingsA.Abhishek Srivastava, B.Benjamin Muller and D.Djamé Seddah. Unsupervised Learning for Handling Code-Mixed Data: A Case Study on POS Tagging of North-African Arabizi Dialect.EurNLP - First annual EurNLPPosterOctober 2019
  HALback to text
• 116 inproceedingsE.Emma Strubell, A.Ananya Ganesh and A.Andrew McCallum. Energy and Policy Considerations for Deep Learning in NLP.Proceedings of the 57th Annual Meeting of the Association for Computational LinguisticsFlorence, ItalyAssociation for Computational LinguisticsJuly 2019, 3645--3650URL: https://aclanthology.org/P19-1355
  DOIback to textback to text
• 117 miscD.Dídac Surís, D.Dave Epstein and C.Carl Vondrick. Globetrotter: Unsupervised Multilingual Translation from Visual Alignment.2020
  back to text
• 118 inproceedingsÉ.Éric Villemonte de La Clergerie. Jouer avec des analyseurs syntaxiques.TALN 2014ATALAMarseilles, FranceJuly 2014
  HALback to text
• 119 inproceedingsÉ.Éric Villemonte de La Clergerie, B.Benôit Sagot and D.Djamé Seddah. The ParisNLP entry at the ConLL UD Shared Task 2017: A Tale of a #ParsingTragedy.Conference on Computational Natural Language LearningProceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal DependenciesVancouver, CanadaAugust 2017, 243-252
  HALDOIback to textback to text
• 120 inproceedingsG.Géraldine Walther and B.Benôit Sagot. Speeding up corpus development for linguistic research: language documentation and acquisition in Romansh Tuatschin.Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and LiteratureProceedings of the Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and LiteratureVancouver, CanadaAugust 2017, 89 - 94
  HALDOIback to text
• 121 unpublishedWhen Being Unseen from mBERT is just the Beginning: Handling New Languages With Multilingual Language Models.
  back to text