Redefining senses for better paraphrase substitution in context

Marianna Apidianaki (LIMSI, CR), mardi 19 septembre à 11h30 en salle de conférence.

The role of senses in NLP has been questioned due to the high performance of vector space models in semantic tasks. These models deliver state-of-the-art performance without explicitly accounting for senses which have even been shown to be harmful for some tasks. In this talk, I will show how sense representations tailored to the task can improve the results of vector-based lexical substitution models. I will discuss two aspects related to paraphrase substitution, namely their clusterability into senses and their substitutability in context. Finally, I will present preliminary results on core sense detection through a multi-view approach to paraphrase semantic analysis.