Global Inference to Chinese Temporal Relation Extraction

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1451–1460, Osaka, Japan, December 11-17 2016. Global Inference to Chinese Temporal Relation Extraction Peifeng Li, Qiaoming Zhu, Guodong Zhou, Hongling Wang School of Computer Science & Technology Soochow University, Suzhou, 215006, China {pfli, qmzhu, gdzhou, hlwang}@suda.edu.cn Abstract Previous studies on temporal relation extraction focus on mining sentence-level information or enforcing coherence on different temporal relation types among various event mentions in the same sentence or neighboring sentences, largely ignoring those discourse-level temporal relations in nonadjacent sentences. In this paper, we propose a discourse-level global inference model to mine those temporal relations be- tween event mentions in document-level, especially in nonadjacent sentences. Moreover, we provide var- ious kinds of discourse-level constraints, which derived from event semantics, to further improve our global inference model. Evaluation on a Chinese corpus justifies the effectiveness of our discourse-level global inference model over two strong baselines. 1 Introduction Temporal relation extraction is to determine the temporal relationship (e.g., Before and After) holding among events. It has been drawing more and more attention due to the crucial importance of temporal information to various natural language processing (NLP) applications, such as language generation, information extraction, summarization, and question answering. The difficulty with this task is that tem- poral information about event mentions is sometimes not stated explicitly and one can only infer from their context. Currently, temporal relation extraction still remains a challenge in corpus construction and inference mechanism. On one hand, although the TimeBank corpus (Pustejovsky et al., 2003), the commonly used corpus in previous studies, has largely promoted the development of temporal relation extraction, it only anno- tates a small subset of easily-identified event mention pairs. Moreover, it largely ignores almost all temporal relations between event mentions in nonadjacent sentences. These lead to fragmented relations and limit its applications to other NLP tasks, such as information extraction, and summarization. Finally, while constructing a fully-annotated corpus is expensive and time-consuming, many NLP tasks are nor- mally interested in specific types of events. For example, a summarization or information extraction system on terrorism attacks may only concern with a few event types (e.g., Attack, Die, and Injure). Therefore, annotating an event-driven fully-annotated temporal relation corpus becomes a crucial issue to the success of real-life applications. On the other hand, previous studies on temporal relation extraction focus on mining sentence-level information or enforcing coherence on different temporal relation types among various mentions in the same sentence or neighboring sentences, largely ignoring those discourse-level temporal relations in nonadjacent sentences. Specifically, only a few studies apply global inference models to exploit tem- poral relations in discourse level. Therefore, how to acquire discourse-level temporal information from those long-distance event mention pairs in nonadjacent sentences becomes another crucial issue to tem- poral relation extraction, especially for Chinese, as a discourse-driven language with a broad range of ellipsis and flexible sentence structures. In this paper, we first annotate an event-driven fully-annotated Chinese temporal relation corpus, on the top of the …

Content truncated to 500 words