Graph Based Network with Contextualized Representations of Turns in Dialogue

Dialogue-based relation extraction (RE) aims to extract relation(s) between\ntwo arguments that appear in a dialogue. Because dialogues have the\ncharacteristics of high personal pronoun occurrences and low information\ndensity, and since most relational facts in dialogues are not supported by any\nsingle sentence, dialogue-based relation extraction requires a comprehensive\nunderstanding of dialogue. In this paper, we propose the TUrn COntext awaRE\nGraph Convolutional Network (TUCORE-GCN) modeled by paying attention to the way\npeople understand dialogues. In addition, we propose a novel approach which\ntreats the task of emotion recognition in conversations (ERC) as a\ndialogue-based RE. Experiments on a dialogue-based RE dataset and three ERC\ndatasets demonstrate that our model is very effective in various dialogue-based\nnatural language understanding tasks. In these experiments, TUCORE-GCN\noutperforms the state-of-the-art models on most of the benchmark datasets. Our\ncode is available at https://github.com/BlackNoodle/TUCORE-GCN.\n