Enhancing dialogue relation extraction ability by incorporating the dialogue structure into the PLM-based encoder

Dialogue relation extraction, as a novel and significant task in recent years, aims to identify the relationship between a pair of subject and object within a conversation. This task serves as a fundamental technical basis for research fields such as dialogue generation and dialogue understanding, offering significant theoretical and practical value. Considering the complicated logical structure based on speaker interactions in multi-party dialogues, which entail abundant information, we aim to enhance the contextual representation of dialogue texts by integrating the dialogue structure into encoders on the basis of pretrained language models. Specifically, we define two types of dialogue structures and introduce an attention correction module into the self-attention layer of pretrained language models. This module parameterizes the dialogue structure into trainable neural network layers, calculates attention biases on the basis of the dialogue structure, and uses these biases to adjust the standard attention scores, thereby integrating knowledge of the dialogue structure into the encoding process. Our approach can be seamlessly integrated into any dialogue model on the basis of pretrained language models. We conduct comprehensive experiments on two dialogue relation extraction datasets, DialogRE and DDRel, which achieve significantly improved results compared with the competitive baselines.