Séminaires du CENTAL

CENTAL Louvain-La-Neuve

Les séminaires du CENTAL ont pour but de réunir des enseignants, des étudiants et des chercheurs (du monde académique ou de l'industrie) intéressés par le traitement automatique de langues. Les séminaires sont gratuits et ouverts à tous et ont généralement lieu le mardi de 11h à 12h. Si vous souhaitez être informé par courrier électronique des séminaires que nous organisons et des actualités du CENTAL, nous vous proposons de vous inscrire à la liste de diffusion du CENTAL en indiquant votre adresse électronique dans le formulaire.

 

Organisation 2021-22

Eva Rolin
Xiaoou Wang


Calendrier 2021-22

19 octobre 2021 — Adrien Bibal - Local DOYEN 21
Interpretability and Explainability in Machine Learning

Adrien Bibal, Postdoctoral Researcher in Machine Learning and NLP at UCLouvain

 Lien vers la visioconférence

Abstract :

Machine learning models are becoming more and more complex for the sake of performance. However, in many situations, the way in which the model is computed must be somewhat transparent. For instance, in some countries, the reasons for credit denial must legally be provided. Furthermore, in science, it is often not the predictive performance of the model that is sought, but the knowledge that can be extracted from it. Interpretability is a property of models that characterizes the degree to which models are understandable by their users, while explainability is the capacity the model to be explained, if it is not interpretable. In this seminar, we will introduce these two concepts, as well as the issues related to their implementation and evaluation.

Références :

Bibal, A., & Frenay, B. (2016). Interpretability of machine learning models and representations: an introduction. In European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (pp. 77-82).

Lipton, Z. C. (2018). The mythos of model interpretability: in machine learning, the concept of interpretability is both important and slippery. Queue16(3), 31-57.

Guidotti, R., Monreale, A., Ruggieri, S., Turini, F., Giannotti, F., & Pedreschi, D. (2018). A survey of methods for explaining black box models. ACM Computing Surveys51(5), 1-42.

Bibal, A., Lognoul, M., De Streel, A., & Frénay, B. (2021). Legal requirements on explainability in machine learning. Artificial Intelligence and Law29(2), 149-169.

Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence1(5), 206-215.


09 novembre 2021 — Thomas François - Salle du Conseil FIAL
Le TAL pour l'évaluation automatique de la difficulté de lecture en FLE

Thomas François,  Chargé de cours à l'UCLouvain

 Lien vers la visioconférence

Abstract : 

Cette présentation introduira trois projets du Cental ciblant la question de l’évaluation automatique de la difficulté de documents pour le FLE. Tout d’abord, nous présenterons une nouvelle formule de lisibilité pour évaluer la difficulté des textes de FLE automatiquement. Cette formule intègre les dernières technologies basées sur le Deep Learning et les représentations sémantiques de type BERT. Dans un second temps, nous présenterons le projet CEFRLex, un projet phare du Cental. La présentation décrira ses fondements théoriques et la méthodologie de conception des ressources. Ensuite, nous discuterons comment les informations du projet CEFRLex, couplées avec les référentiels de Beacco pour le CECR, peuvent être utilisées pour prédire automatiquement la connaissance lexicale réceptive d’apprenants du FLE.

Références 

https://cental.uclouvain.be/team/tfrancois/

 


Reporté au deuxième quadrimestre — Matthew Shardlow
Neural Text Simplification: Methods, Evaluation and Application

Matthew Shardlow, Lecturer in computer science at the Manchester Metropilitan University

Abstract :

Text Simplification is the technique of automatically reducing the complexity of a text by altering the syntax or lexicon. Neural text simplification seeks to apply deep learning to the text simplification problem in order to create systems which can automatically generate easier to understand alternatives to complex texts. In this talk, I will look at the current methods for neural text simplification ranging from the application of statistical machine translation software to the transformer methodology and beyond. I will consider how to evaluate text simplification (and other natural language generation methods) with a view to developing gold standard evaluation practices that can be adopted between researchers. I will also discuss some of the varied applications of text simplification, ranging from improving medical language for patients to improving the performance of other NLP tool.

Références :

Shardlow, M. and Nawaz, R., 2019, July. Neural Text Simplification of Clinical Letters with a Domain Specific Phrase Table. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (pp. 380-389).

Cooper, M. and Shardlow, M., 2020, May. CombiNMT: An exploration into neural text simplification models. In Proceedings of the 12th Language Resources and Evaluation Conference (pp. 5588-5594).

Przybyła, P. and Shardlow, M., 2020, December. Multi-Word Lexical Simplification. In Proceedings of the 28th International Conference on Computational Linguistics (pp. 1435-1446).

Shardlow, M., Sellar, S. and Rousell, D., 2021. Collaborative augmentation and simplification of text (CoAST): pedagogical applications of natural language processing in digital learning environments. Learning Environments Research, pp.1-23.

 


07 décembre 2021 — Bill McDowell (DuoLingo) - Online
The Duolinguo CEFR Checker: A multilingual Tool for Adapting Learning Content

Bill McDowell, this seminar is held in the context of the CEFR workshop

 Lien vers la visioconférence

Abstract :

Duolingo is the world's most popular language education platform, with more than 500 million students worldwide. Content creation for the Duolingo app requires adapting text in many languages to target varying levels of proficiency.   To make this process more efficient, we have developed automated multilingual methods for aligning content to the CEFR proficiency standard.  In this talk, I’ll discuss the Duolingo CEFR Checker, a (semi-)language-agnostic tool that aligns text to the CEFR standard using methods that involve transfer learning, multilingual word embeddings, and word frequencies estimated across large corpora.


14 décembre 2021 A.Seza Doğruöz — Online