TY - GEN
T1 - M-HGN
T2 - 17th International Conference on Knowledge Science, Engineering and Management, KSEM 2024
AU - Gao, Xiaoqian
AU - Zhou, Xiabing
AU - Cao, Rui
AU - Zhang, Min
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024.
PY - 2024
Y1 - 2024
N2 - Multi-party dialogue reading comprehension aims to comprehend the dialogue context and answer the questions. Using explicit contextual information effectively and trying to capture the implicit information behind the dialogue lie at the heart of the task, which is still under-explored by previous works. In this paper, we propose a multi-information enhanced heterogeneous graph network to hierarchically model the dialogue context at different granularity levels and the multi-grained interactive relations among them. In detail, we construct a heterogeneous graph by regarding speakers, discourse structures, and questions as different types of nodes to model the information flows among them. In addition, we consider commonsense knowledge as a type of node in order to comprehend additional implicit information. Furthermore, we design a proper interaction module to incorporate these multi-dimensional features together. Experimental results show that our model achieves significant improvements on Molweni dataset. Also, extensive analysis indicates that our model can effectively capture multi-grained information and enhance the understanding of multi-party dialogue.
AB - Multi-party dialogue reading comprehension aims to comprehend the dialogue context and answer the questions. Using explicit contextual information effectively and trying to capture the implicit information behind the dialogue lie at the heart of the task, which is still under-explored by previous works. In this paper, we propose a multi-information enhanced heterogeneous graph network to hierarchically model the dialogue context at different granularity levels and the multi-grained interactive relations among them. In detail, we construct a heterogeneous graph by regarding speakers, discourse structures, and questions as different types of nodes to model the information flows among them. In addition, we consider commonsense knowledge as a type of node in order to comprehend additional implicit information. Furthermore, we design a proper interaction module to incorporate these multi-dimensional features together. Experimental results show that our model achieves significant improvements on Molweni dataset. Also, extensive analysis indicates that our model can effectively capture multi-grained information and enhance the understanding of multi-party dialogue.
KW - Commonsense knowledge
KW - Heterogeneous graph network
KW - Multi-party dialogue
KW - Reading comprehension
UR - https://www.scopus.com/pages/publications/85200765521
U2 - 10.1007/978-981-97-5495-3_28
DO - 10.1007/978-981-97-5495-3_28
M3 - 会议稿件
AN - SCOPUS:85200765521
SN - 9789819754946
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 371
EP - 383
BT - Knowledge Science, Engineering and Management - 17th International Conference, KSEM 2024, Proceedings
A2 - Cao, Cungeng
A2 - Chen, Huajun
A2 - Zhao, Liang
A2 - Arshad, Junaid
A2 - Wang, Yonghao
A2 - Asyhari, Taufiq
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 16 August 2024 through 18 August 2024
ER -