Deep Natural Language Processing for intent classification: an application to the classification of emotions and dialog act

ALEXIS VIGNARD, Pierre Personnat

Abstract:

This paper forcuses on the implementation of a Natural Language Processing (NLP) model that performs contextual emotion classification on a dialogue dataset sourced from the popular TV show "Friends". The primary goal is to account for the context of each sentence in the conversations through a self-attention mechanism. To achieve this objective, the authors propose constructing a hierarchical attention network that mimics the structure of a corpus of dialogues, which has a hierarchical structure between the conversation, the sentences within it, and the words in each sentence.

The aim is to capture a dual context, namely, that the importance of words varies depending on the sentence or conversation, and not every sentence in a conversation carries equal importance in the context of the dialogue. Therefore, the meaning and, consequently, the label of an utterance will vary depending on its location within the conversation.