Text2Event: Controllable Sequence-to-Structure Generation for End-to-end Event Extraction

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 2795–2806 August 1–6, 2021. ©2021 Association for Computational Linguistics 2795 TEXT2EVENT: Controllable Sequence-to-Structure Generation for End-to-end Event Extraction Yaojie Lu1,3, Hongyu Lin1, Jin Xu4,∗, Xianpei Han1,2,∗, Jialong Tang1,3, Annan Li1,3, Le Sun1,2, Meng Liao4, Shaoyi Chen4 1Chinese Information Processing Laboratory 2State Key Laboratory of Computer Science Institute of Software, Chinese Academy of Sciences, Beijing, China 3University of Chinese Academy of Sciences, Beijing, China 4Data Quality Team, WeChat, Tencent Inc., China {yaojie2017,jialong2019,liannan2019}@iscas.ac.cn {hongyu,xianpei,sunle}@iscas.ac.cn {jinxxu,maricoliao,shaoyichen}@tencent.com Abstract Event extraction is challenging due to the com- plex structure of event records and the semantic gap between text and event. Traditional meth- ods usually extract event records by decompos- ing the complex structure prediction task into multiple subtasks. In this paper, we propose TEXT2EVENT, a sequence-to-structure gener- ation paradigm that can directly extract events from the text in an end-to-end manner. Specif- ically, we design a sequence-to-structure net- work for unified event extraction, a constrained decoding algorithm for event knowledge injec- tion during inference, and a curriculum learning algorithm for efficient model learning. Exper- imental results show that, by uniformly mod- eling all tasks in a single model and univer- sally predicting different labels, our method can achieve competitive performance using only record-level annotations in both super- vised learning and transfer learning settings. 1 Introduction Event extraction is an essential task for natural language understanding, aiming to transform the text into event records (Doddington et al., 2004; Ahn, 2006). For example, in Figure 1, mapping “The man returned to Los Angeles from Mexico following his capture Tuesday by bounty hunters.” into two event records {Type: Transport, Trigger: returned, Arg1 Role: Artifact, Arg1: The man, Arg2 Role: Destination, Arg2: Los Angeles, ... } and {Type: Arrest-Jail, Trigger: capture, Arg1 Role: Person, Arg1: The man, Arg2 Role: Agent, Arg2: bounty hunters, ... }. Event extraction is challenging due to the com- plex structure of event records and the semantic gap between text and event. First, an event record contains event type, trigger, and arguments, which ∗Corresponding authors. The man returned to Los Angeles from Mexico following his capture Tuesday by bounty hunters. Controllable Generation Event Schema Arrest-Jail • Person • Crime • Agent • Time Transport • Destination • Origin • Artifact • Vehicle • Time Event Type Arrest-Jail Trigger capture Person The man Time Tuesday Agent bounty hunters Event Type Transport Trigger returned Artifact The man Destination Los Angeles Origin Mexico Sequence-to-Structure Network Constraint Figure 1: The framework of TEXT2EVENT. Here, TEXT2EVENT takes raw text as input and generates a Transport event and an Arrest-Jail event. form a table-like structure. And different event types have different structures. For example, in Figure 1, Transport and Arrest-Jail have entirely different structures. Second, an event can be ex- pressed using very different utterances, such as diversified trigger words and heterogeneous syntac- tic structures. For example, both “the dismission of the man” and “the man departed his …

Content truncated to 500 words