PESE: Event Structure Extraction using Pointer Network based Encoder-Decoder Architecture

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1091–1100 November 20–23, 2022. ©2022 Association for Computational Linguistics 1091 PESE: Event Structure Extraction using Pointer Network based Encoder-Decoder Architecture Alapan Kuila IIT Kharagpur, India alapan.cse@iitkgp.ac.in Sudeshna Sarkar IIT Kharagpur, India sudeshna@cse.iitkgp.ac.in Abstract The task of event extraction (EE) aims to find the events and event-related argument informa- tion from the text and represent them in a struc- tured format. Most previous works try to solve the problem by separately identifying multiple substructures and aggregating them to get the complete event structure. The problem with the methods is that it fails to identify all the interde- pendencies among the event participants (event- triggers, arguments, and roles). In this paper, we represent each event record in a unique tuple format that contains trigger phrase, trigger type, argument phrase, and corresponding role infor- mation. Our proposed pointer network-based encoder-decoder model generates an event tu- ple in each time step by exploiting the interac- tions among event participants and presenting a truly end-to-end solution to the EE task. We evaluate our model on the ACE2005 dataset, and experimental results demonstrate the effec- tiveness of our model by achieving competitive performance compared to the state-of-the-art methods. 1 Introduction Event extraction (EE) from text documents is one of the crucial tasks in natural language processing and understanding. Event extraction deals with the identification of event-frames from natural lan- guage text. These event-frames have a complex structure with information regarding event-trigger, event type, event-specific arguments, and event- argument roles. For example, In Baghdad, a cameraman died when an American tank fired on the Palestine ho- tel. In this sentence died and fired are the event trig- gers for the event types Die and Attack respectively. The sentence contains entities phrases: Baghdad, a cameraman, an American tank and Palestine ho- tel. Some of these entities play a specific role in these mentioned events and termed as event argu- ments. For event type Die, (argument; role) pairs are: (Baghdad; Place), (A cameraman; victim), (American tank; instrument). Whereas, for Attack event, (argument; role) pairs are: (Baghdad; Place), (A cameraman; target), (American tank; instru- ment), and (Palestine Hotel; Target). Apparently, a sentence may contain multiple events; an entity may be shared by multiple event frames; more- over, a specific argument may play different roles in different event frames. Therefore an ideal event extraction system will identify all the trigger words, classify the correct event types, extract all the event- specific arguments and correctly predict the event- argument roles. Each of these subtasks is equally important and challenging. Most existing works decompose the EE task into these predefined subtasks and later aggregate those outputs to get the complete event frames. Some of these models follow a pipelined approach where triggers and corresponding arguments are identified in separate stages. In contrast, others rely on joint modeling that predicts triggers and relevant argu- …

Content truncated to 500 words