Identifier les relations discursives implicites en combinant données naturelles et données artificielles

Chloé Braud 1 Pascal Denis 2
1 ALPAGE - Analyse Linguistique Profonde à Grande Echelle ; Large-scale deep linguistic processing
Inria Paris-Rocquencourt, UPD7 - Université Paris Diderot - Paris 7
2 MAGNET - Machine Learning in Information Networks
LIFL - Laboratoire d'Informatique Fondamentale de Lille, Inria Lille - Nord Europe
Abstract : This paper presents the first experiments on French in automatic identification of implicit discourse relations (i.e. relations that lack an overt connective). Our systems exploit hand-labeled implicit examples, along with artificial implicit examples obtained from explicit examples by suppressing their connective, following Marcu et Echihabi (2002). Previous work on English shows that using artificial data for training largely degrades performance on natural data, reflecting important differences in the distribution. This conclusion, that also holds for French, has led us to consider various methods inspired by domain adaptation to better combine the data. We evaluate these methods on the ANNODIS corpus: our best system achieves a 41.7 % accuracy, that is a significant gain of 4.4 % compared to a model using only the natural data. MOTS-CLÉS : structure discursive, relations discursives implicites, apprentissage automatique.
Complete list of metadatas

Cited literature [30 references]  Display  Hide  Download

https://hal.inria.fr/hal-01094346
Contributor : Chloé Braud <>
Submitted on : Friday, December 12, 2014 - 10:40:00 AM
Last modification on : Thursday, February 21, 2019 - 10:52:55 AM
Long-term archiving on : Friday, March 13, 2015 - 10:31:16 AM

File

braud_denis_TAL55.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-01094346, version 1

Citation

Chloé Braud, Pascal Denis. Identifier les relations discursives implicites en combinant données naturelles et données artificielles. Traitement Automatique des Langues, Lavoisier (Hermes Science Publications), 2014, 55 (1), pp.31. ⟨http://www.atala.org/-Volume-55-⟩. ⟨hal-01094346⟩

Share

Metrics

Record views

429

Files downloads

323