TY - GEN
T1 - A Unified Information Extraction System Based on Role Recognition and Combination
AU - Zhang, Yadong
AU - Lan, Man
N1 - Publisher Copyright:
© 2021, Springer Nature Switzerland AG.
PY - 2021
Y1 - 2021
N2 - In this paper, we propose a unified information extraction system, which handles event extraction (EE) and relation extraction (RE) tasks. Given context and schema, event extraction aims to extract the events and the specific roles in the events, and relation extraction extracts all SPO triples. We formulate event extraction and relation extraction as one extraction schema, that is, role recognition and role combination. We use Multi-Label Pointer Network (MLPN) to recognize composite roles that contain both event/relation and role information and simultaneously train a Co-occurrence Matrix (CM) to determine the co-occurrence relationship of composite roles, i.e., whether two roles describe the same event/relation. Using such a Unified model based on Role Recognition and Combination (URRC) and corresponding combination strategy, we implement three tasks: sentence-level event extraction, document-level event extraction, and relation extraction. In LIC 2021, our model achieved 6th in the Multi-format Information Extraction racing track with an average F1 score of 77.44% in the final test dataset of three subtasks.
AB - In this paper, we propose a unified information extraction system, which handles event extraction (EE) and relation extraction (RE) tasks. Given context and schema, event extraction aims to extract the events and the specific roles in the events, and relation extraction extracts all SPO triples. We formulate event extraction and relation extraction as one extraction schema, that is, role recognition and role combination. We use Multi-Label Pointer Network (MLPN) to recognize composite roles that contain both event/relation and role information and simultaneously train a Co-occurrence Matrix (CM) to determine the co-occurrence relationship of composite roles, i.e., whether two roles describe the same event/relation. Using such a Unified model based on Role Recognition and Combination (URRC) and corresponding combination strategy, we implement three tasks: sentence-level event extraction, document-level event extraction, and relation extraction. In LIC 2021, our model achieved 6th in the Multi-format Information Extraction racing track with an average F1 score of 77.44% in the final test dataset of three subtasks.
KW - Document-level event extraction
KW - Event extraction
KW - Relation extraction
UR - https://www.scopus.com/pages/publications/85118167691
U2 - 10.1007/978-3-030-88483-3_36
DO - 10.1007/978-3-030-88483-3_36
M3 - 会议稿件
AN - SCOPUS:85118167691
SN - 9783030884826
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 447
EP - 459
BT - Natural Language Processing and Chinese Computing - 10th CCF International Conference, NLPCC 2021, Proceedings
A2 - Wang, Lu
A2 - Feng, Yansong
A2 - Hong, Yu
A2 - He, Ruifang
PB - Springer Science and Business Media Deutschland GmbH
T2 - 10th CCF Conference on Natural Language Processing and Chinese Computing, NLPCC 2021
Y2 - 13 October 2021 through 17 October 2021
ER -