A Novel Context-Aware Model for Relation Extractions from Conversational Texts using Graph Convolutional Networks

Conversational text or dialogue is one of the most common types of texts. Examples include forum posts, parliament Hansard, and movie transcripts. Extracting relations among the entities in a conversation contributes to understanding the dynamics of the conversation. Although conversations exhibit cross-sentence relations, most of the existing solutions focus on texts from formal genres such as academic articles or news reports. One of the unique characteristics of conversation is the speaker information that could be leveraged to provide contexts to the relation extraction tasks. Depending on the nature of the conversational texts, the characteristics of a speaker may include gender, position held in an organisation, or the role played in a movie transcript. Besides, the context of a conversation, which includes the themes and intents of the conversation is important to enabling accurate relation extraction. Unfortunately, existing solutions do not provide a holistic solution to perform cross sentence relation extraction on conversational texts. Besides, most of these existing solutions do not take into account the attributes of the speakers when modelling relation extraction. This research project aims to overcome these gaps and propose a context-aware model that is able to provide a holistic solution for cross-sentence relation extraction from conversational texts. The model will be proposed based on graph convolutional networks, which has been proven effective in many other natural language processing tasks. Ultimately, the output of this research will enable computational understanding of relations among the entities in a conversational text, which can be used to facilitate information extraction, summarisation and visualisation.