Artificial Intelligence & Language

Description

Context

Modern natural language processing (NLP) systems are excessively dependent on the availability of annotated resources (data addiction) which (1) increases the digital gap between high and low resource languages or dialects and (2) increases the risk of machine bias where empirically trained models reproduce controversial societal inequalities such as gender, racial bias, etc. Simultaneously, improving algorithms for automatic analysis of text and speech create new opportunities for basic and applied research in language science (e.g. descriptive and theoretical linguistics, sociolinguistics, study of language development) but these algorithms must be able to learn from few examples since corpora collected by linguists are of limited size. From these observations, the objective of the chair is to make NLP less data addicted (and therefore fairer) as well as to contribute to a methodological turn in language-related social science research by leveraging machine learning and modern natural language processing techniques.

Goals

We aim to build models that 1/ can learn to process language from as little data as a learning child and 2/ are free of the social biases included in the data. For this, human labeling is replaced by weaker signals in the form of multimodal information, prior knowledge, inductive biases, cross language (or task) similarities, context.

Research avenues

Due to the cost of data annotation, an open issue in NLP is the design of learning methods that are data-efficient (generalize from a few examples) and can leverage diverse types of knowledge. In this context, we propose to use:

  1. modeling based on the same information shared between languages or tasks (e.g. multilingual or multitask learning) ;
     
  2. zero-shot methods that do not need annotated data (e.g. but use representations learnt from speech or text without supervision) ;
     
  3. expert knowledge in empirical systems (include priors in bayesian or neural models, use typological features) ;
     
  4. multimodal data for semantic supervision (models of visually grounded speech and language) ;
     
  5. parsimonious models (known to have a better explanatory predictive power) ;
     
  6. inductive biases (from psycholinguistic work on language acquisition) ;
     
  7. innovative language data collection methodologies (via crowdsourcing, mobile apps).

Activities

2 PhDs have started in January 2020:

  • Brooke Stephenson: Incremental (low latency) TTS (co-supervision with chaire P. Perrier)
  • Lorenzo Lupo: Document level neural machine translation

2 M2R have been supervised:

  • End-to-end speech parsing (Ousama Gasmi)
  • Analysis of contextualized language models’ lexical functions (Vincent Bellue)

Collaborative work on self supervised text representation learning - Release of FlauBERT (language model for French, trained with Jean Zay supercomputer): https://github.com/getalp/Flaubert 

Chair events

Organization of ALPS (Advanced Language Processing School) Winter school that will take place (virtually)
from Sunday, January 17th to Friday 22nd 2021 - http://lig-alps.imag.fr

Scientific publications

  • Odette Scharenborg, Laurent Besacier, Alan Black, Mark Hasegawa-Johnson, Florian Metze, et al.. Speech technology for unwritten languages. IEEE/ACM Transactions on Audio, Speech and Language Processing Journal, Institute of Electrical and Electronics Engineers, 2020
     
  • Ha Nguyen, Fethi Bougares, Natalia Tomashenko, Yannick Estève, Laurent Besacier. Investigating Self-supervised Pre-training for End-to-end Speech Translation. Interspeech 2020, Oct 2020, Shangai (Virtual Conf), China.
     
  • Vaishali Pal, Fabien Guillot, Manish Shrivastava, Jean-Michel Renders, Laurent Besacier. Modeling ASR Ambiguity for Neural Dialogue State Tracking. Interspeech 2020, Oct 2020, Shangai (Virtual Conf), China.
     
  • Ewan Dunbar, Julien Karadayi, Mathieu Bernard, Xuan-Nga Cao, Robin Algayres, et al.. The Zero Resource Speech Challenge 2020: Discovering discrete subword and word units. Interspeech 2020, Oct 2020, Shangai (Virtual Conf), China.
     
  • Brooke Stephenson, Laurent Besacier, Laurent Girin, Thomas Hueber. What the Future Brings: Investigating the Impact of Lookahead for Incremental Neural TTS. Interspeech 2020, Oct 2020, Shangai (Virtual Conf), China.
     
  • Maha Elbayad, Laurent Besacier, Jakob Verbeek. Efficient Wait-k Models for Simultaneous Machine Translation. Interspeech 2020, Oct 2020, Shangai (Virtual Conf), China.
     
  • Hang Le, Juan Pino, Changhan Wang, Jiatao Gu, Didier Schwab, Laurent Besacier. Dual-decoder Transformer for Joint Automatic Speech Recognition and Multilingual Speech Translation. The 28th International Conference on Computational Linguistics (COLING 2020).
     
  • Maha Elbayad, Michael Ustaszewski, Emmanuelle Esperança-Rodier, Francis Brunet Manquat, Laurent Besacier. Online Versus Offline NMT Quality: An In-depth Analysis on English-German and German-English. The 28th International Conference on Computational Linguistics (COLING 2020).
     
  • Eric Lefferand, Steven Bird, Laurent Besacier. Enabling Interactive Transcription in an Indigenous Community. The 28th International Conference on Computational Linguistics (COLING 2020).
     
  • Jerin Philip, Alexandre Bérard, Laurent Besacier, Matthias Gallé. Monolingual Adapters for Zero-Shot Neural Machine Translation. EMNLP (Empirical Methods for Natural Language Processing), Nov 2020, Virtual, France.
     
  • William Havard, Laurent Besacier, Jean-Pierre Chevrot. Catplayinginthesnow: Impact of Prior Segmentation on a Model of Visually Grounded Speech. Conference on Natural Language Learning (CoNLL), Nov 2020, Virtual, France.