PoMo: Generating Entity-Specific Post-Modifiers in Context

Proceedings of NAACL-HLT 2019, pages 826–838 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 826 PoMo: Generating Entity-Specific Post-Modifiers in Context Jun Seok Kang1, Robert L. Logan IV2, Zewei Chu3, Yang Chen3, Dheeru Dua2, Kevin Gimpel4, Sameer Singh2, Niranjan Balasubramanian1 1Stony Brook University, NY, USA 2University of California, Irvine, CA, USA 3University of Chicago, IL, USA 4Toyota Technological Institute at Chicago, IL, USA {junkang,niranjan}@cs.stonybrook.edu {rlogan,ddua,sameer}@uci.edu {zeweichu,yangc1}@uchicago.edu {kgimpel}@ttic.edu Abstract We introduce entity post-modifier generation as an instance of a collaborative writing task. Given a sentence about a target entity, the task is to automatically generate a post-modifier phrase that provides contextually relevant in- formation about the entity. For example, for the sentence, “Barack Obama, , sup- ported the #MeToo movement.”, the phrase “a father of two girls” is a contextually rel- evant post-modifier. To this end, we build PoMo, a post-modifier dataset created auto- matically from news articles reflecting a jour- nalistic need for incorporating entity informa- tion that is relevant to a particular news event. PoMo consists of more than 231K sentences with post-modifiers and associated facts ex- tracted from Wikidata for around 57K unique entities. We use crowdsourcing to show that modeling contextual relevance is necessary for accurate post-modifier generation. We adapt a number of existing generation ap- proaches as baselines for this dataset. Our results show there is large room for improve- ment in terms of both identifying relevant facts to include (knowing which claims are relevant gives a > 20% improvement in BLEU score), and generating appropriate post-modifier text for the context (providing relevant claims is not sufficient for accurate generation). We con- duct an error analysis that suggests promising directions for future research. 1 Introduction The goal of machine-in-the-loop writing systems is to assist human writers by directly augmenting their text. Examples include systems that refine human text for grammar (Rao and Tetreault, 2018), collaborate on story plot generation systems (Clark et al., 2018; Yu and Riedl, 2012), or modify the content for style (Hu et al., 2017; Shen et al., 2017; Yang et al., 2018). In this paper, we introduce Professor Melman ’s arguments appealed to a wide spectrum, attracting unions like the United Automobile Workers and the Machinists Union ... Noam Chomsky , , said Dr. Melman helped mobilize what once was weak and scattered re- sistance to war and other military operations. “The country is a lot different than it was 30 to 40 years ago, and he had a big role in that,” Mr. Chomsky said. Noam Chomsky (Q9049) spouse Carol Chomsky occupation university teacher political ideology anarchism employer MIT notable work “Class Warfare” Input Entity Mention and Context and Claims from Wikidata Output “the MIT professor and antiwar activist” Figure 1: Post-Modifier Generation Task post-modifier generation as an instance of such an assistive writing task in the news domain. Jour- nalists use post-modifiers to introduce background information about entities discussed in news arti- cles. To write these post-modifiers journalists often need to look up relevant facts …

Content truncated to 500 words