CCG Supertagging Using Morphological and Dependency Syntax Information

Luyện Ngọc Lê , Yannis Haralambous (1, 2, 3, 4, 5)

1 Lab-STICC_IMTA_CID_DECIDE
2 IMT Atlantique - INFO - Département Informatique
3 Lab-STICC_DECIDE - Equipe DECIDE
4 IMT Atlantique - IMT Atlantique
5 Lab-STICC - Laboratoire des sciences et techniques de l'information, de la communication et de la connaissance

After presenting a new CCG supertagging algorithm based on morphological and dependency syntax information, we use this algorithm to create a CCG French Tree Bank corpus (20,261 sentences) based on the FTB corpus by Abeillé et al. We then use this corpus, as well as the Groningen Tree Bank corpus for the English language, to train a new BiLSTM+CRF neural architecture that uses (a) morphosyntactic input features and (b) feature correlations as input features. We show experimentally that for an inflected language like French, dependency syntax information allows significant improvement of the accuracy of the CCG supertagging task, when using deep learning techniques.

Mots clés: Dependency Syntax; Supertagging; Combinatory Categorial Grammar; Deep learning

Domaines: Interface homme-machine [cs.HC]; Réseaux et télécommunications [cs.NI]; Traitement du texte et du document; Linguistique

Luyện Ngọc Lê, Yannis Haralambous. CCG Supertagging Using Morphological and Dependency Syntax Information. International Conference on Computational Linguistics and Intelligent Text Processing, Apr 2019, La Rochelle, France. pp.608-621, ⟨10.1007/978-3-031-24337-0_43⟩. ⟨hal-02110790⟩