Neural-Driven Multi-criteria Tree Search for Paraphrase Generation - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Communication Dans Un Congrès Année : 2020

Neural-Driven Multi-criteria Tree Search for Paraphrase Generation

Résumé

A good paraphrase is semantically similar to the original sentence but it must be also well formed, and syntactically different to ensure diversity. To deal with this tradeoff, we propose to cast the paraphrase generation task as a multi-objectives search problem on the lattice of text transformations. We use BERT and GPT2 to measure respectively the semantic distance and the correctness of the candidates. We study two search algorithms: Monte-Carlo Tree Search For Paraphrase Generation (MCPG) and Pareto Tree Search (PTS) that we use to explore the huge sets of candidates generated by applying the PPDB-2.0 edition rules. We evaluate this approach on 5 datasets and show that it performs reasonably well and that it outperforms a state-of-the-art edition-based text generation method.
Fichier principal
Vignette du fichier
neural_driven_multi_criteria_tree_search_for_paraphrase_generation.pdf (2.72 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03127865 , version 1 (01-02-2021)

Identifiants

  • HAL Id : hal-03127865 , version 1

Citer

Betty Fabre, Tanguy Urvoy, Jonathan Chevelu, Damien Lolive. Neural-Driven Multi-criteria Tree Search for Paraphrase Generation. Learning Meets Combinatorial Algorithms (LMCA) Workshop at NeurIPS 2020, Dec 2020, online, France. ⟨hal-03127865⟩
73 Consultations
182 Téléchargements

Partager

Gmail Facebook X LinkedIn More