Document-Level Event Factuality Identification via Adversarial Neural Network

Proceedings of NAACL-HLT 2019, pages 2799–2809 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2799 Document-Level Event Factuality Identification via Adversarial Neural Network Zhong Qian1, Peifeng Li1 2, Qiaoming Zhu1 2 and Guodong Zhou1 2 1School of Computer Science and Technology, Soochow University, Suzhou, China 2AI Research Institute, Soochow University, Suzhou, China qianzhongqz@163.com, {pfli, qmzhu, gdzhou}@suda.edu.cn Abstract Document-level event factuality identification is an important subtask in event factuality and is crucial for discourse understanding in Natural Language Processing (NLP). Previous studies mainly suffer from the scarcity of suit- able corpus and effective methods. To solve these two issues, we first construct a corpus annotated with both document- and sentence- level event factuality information on both En- glish and Chinese texts. Then we present an LSTM neural network based on adversarial training with both intra- and inter-sequence at- tentions to identify document-level event fac- tuality. Experimental results show that our neural network model can outperform various baselines on the constructed corpus. 1 Introduction Document-level event factuality identification is the task of deciding the commitment of relevant sources towards the factual nature of an event, and to determine whether an event is a fact, a possi- bility, or an impossible situation from the view of document. Identifying document-level factuality of events requires comprehensive understanding of documents. As illustrated in Figure 1 where events are in bold, the event “reach” (including its other forms) have various factuality values in different sentences. For example, in paragraph 2, “reach” is impossible/CT- according to the nega- tive word “denied”, while in paragraph 3, “reach” is possible/PS+ due to the speculative word “may”. The main contents of this document is “Mexico de- nied that they will reach an agreement with the U.S. on the new trade deal”, and the document- level factuality of the event “reach” is CT-. Document-level event factuality identification is fundamental for document-level NLP applica- tions, such as machine reading comprehension, which aims to have machines read a text passage According to Politico.com, it is said the United States will reach(CT+) an agreement with Mexico on the new trade deal that will replace North American Free Trade Agreement (NAFTA) before December, 2017. However, Mexican Economy Minister Ildefonso Gua- jardo denied that they plan to reach(CT-) any agreement with the U.S. on the trade deal talks. “We are not going to sacrifice the quality of an agreement because of pressure of time. We will keep en- gaged.” he said. Just two days ago, Guajardo said the two sides may reach(PS+) an agreement within hours. The government has not been informed that any agree- ment will be reached(CT-) yet, said another two Mexi- can officials. During the past few weeks, the U.S. has been ne- gotiating with Mexico on the new trade deal and has achieved much progress. Thus, some media speculate that they will possibly reach(PS+) an agreement. But now it seems that the negotiations will continue before they can get a good deal. (Time: November, 2017) (Document-level factuality of …

Content truncated to 500 words