Improving Event Detection via Open-domain Trigger Knowledge

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5887–5897 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 5887 Improving Event Detection via Open-domain Trigger Knowledge Meihan Tong1, Bin Xu1, Shuai Wang2, Yixin Cao3∗, Lei Hou1, Juanzi Li1 and Jun Xie4 1Knowledge Engineering Laboratory, Tsinghua University, Beijing, China 2SLP Group, AI Technology Department, JOYY Inc, China 3National University of Singapore, Singapore 4SPPD Group, Tencent Inc, China tongmeihan@gmail.com, xubin@tsinghua.edu.cn wangshuai1@yy.com, caoyixin2011@gmail.com greener2009@gmail.com, lijuanzi@tsinghua.edu.cn stiffxie@tencent.com Abstract Event Detection (ED) is a fundamental task in automatically structuring texts. Due to the small scale of training data, previous meth- ods perform poorly on unseen/sparsely la- beled trigger words and are prone to overfit- ting densely labeled trigger words. To ad- dress the issue, we propose a novel Enrich- ment Knowledge Distillation (EKD) model to leverage external open-domain trigger knowl- edge to reduce the in-built biases to frequent trigger words in annotations. Experiments on benchmark ACE2005 show that our model outperforms nine strong baselines, is espe- cially effective for unseen/sparsely labeled trigger words. The source code is released on https://github.com/shuaiwa16/ekd.git. 1 Introduction Event Detection (ED) aims at detecting trigger words in sentences and classifying them into pre- defined event types, which shall benefit numer- ous applications, such as summarization (Li et al., 2019) and reading comprehension (Huang et al., 2019). For instance, in S1 of Figure 1, ED aims to identify the word fire as the event trigger and classify its event type as Attack. Mainstream re- searches (Chen et al., 2015; Liu et al., 2017, 2018b; Liao and Grishman, 2010b; Zhao et al., 2018; Liu et al., 2018a) focus on the second step event type disambiguation via lexical and contextual features. However, it is also crucial to identify trigger words correctly as the preliminary step. Trigger word identification is a non-trivial task, which suffers from the long tail issue. Take the benchmark ACE2005 as an example: trigger words with frequency less than 5 account for 78.2% of the ∗Corresponding author. Densely Labeled Triggers Unseen/Sparsely Labeled Triggers S1: Now we 're hearing the boom of Iraqi guns as they fireAttack towards our positions . S2: Troops were trying to break up stone-throwing protests , but did not use live fireAttack. S4: The intifadaAttack exploded in September 2000 S3: A man was hackedAttack to death by the criminal Trigger identified by Open-domain Trigger Knowledge Figure 1: Examples of ED. fire is the densely labeled trigger for Attack event in ACE2005. Hacked and in- tifada are the unseen/sparsely labeled triggers in the training corpus. The red ones illustrate the triggers identified by open-domain trigger knowledge. total. The long tail issue makes supervised methods (Li et al., 2013; Yang et al., 2019) prone to overfit- ting and perform poorly on unseen/sparsely labeled triggers (Lu et al., 2019). Automatically generat- ing more training instances seems to be a solution: expanding more instances by bootstrapping (Fer- guson et al., 2018; Zhang et al., 2019; Cao et al., 2019) and expending more data from distantly …

Content truncated to 500 words