STORIUM: A Dataset and Evaluation Platform for Machine-in-the-Loop Story Generation

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 707–717, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 707 Incremental Event Detection via Knowledge Consolidation Networks Pengfei Cao1,2, Yubo Chen1,2, Jun Zhao1,2, and Taifeng Wang3 1 National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China 2 School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, 100049, China 3 Ant Group, Hangzhou, 310013, China {pengfei.cao, yubo.chen, jzhao}@nlpr.ia.ac.cn, taifeng.wang@antgroup.com Abstract Conventional approaches to event detection usually require a fixed set of pre-defined event types. Such a requirement is often challenged in real-world applications, as new events con- tinually occur. Due to huge computation cost and storage budge, it is infeasible to store all previous data and re-train the model with all previous data and new data, every time new events arrive. We formulate such challeng- ing scenarios as incremental event detection, which requires a model to learn new classes in- crementally without performance degradation on previous classes. However, existing in- cremental learning methods cannot handle se- mantic ambiguity and training data imbalance problems between old and new classes in the task of incremental event detection. In this pa- per, we propose a Knowledge Consolidation Network (KCN) to address the above issues. Specifically, we devise two components, pro- totype enhanced retrospection and hierarchi- cal distillation, to mitigate the adverse effects of semantic ambiguity and class imbalance, re- spectively. Experimental results demonstrate the effectiveness of the proposed method, out- performing the state-of-the-art model by 19% and 13.4% of whole F1 score on ACE bench- mark and TAC KBP benchmark, respectively. 1 Introduction Event detection (ED) is an important task of in- formation extraction, which aims to identify event triggers and classify them into specific types (Ahn, 2006; Chen et al., 2015). For instance, in the sen- tence “A man died when a tank fired on the ho- tel”, an ED model should be able to recognize two events: a Die event triggered by the word “died” and an Attack event triggered by the word “fired”. Existing ED methods can only handle a fixed number of event classes (also called event types) and perform once-and-for-all training on a fixed Learn Attack Event Data for Attack Data for Die Continually learn new events Data for Meet Learn Meet Event Learn Die Event Figure 1: Incremental event detection: a model con- tinually learns new event classes, when data for them becomes available. At any time, the model is able to perform prediction for all event classes observed so far. dataset. Although the evaluation of this setting is straightforward, it clearly limits the usage of these methods in practical applications, as new events continually emerge in the real world. A practical ED system should be able to incremen- tally learn new classes, instead of requiring a fixed set of pre-defined event classes. Therefore, we con- sider a more realistic incremental learning setting (also called continual learning or lifelong learning) (Ring, 1994; Thrun, 1998), where a learning …

Content truncated to 500 words