Mining Paraphrasal Typed Templates from a Plain Text Corpus

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1913–1923, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Mining Paraphrasal Typed Templates from a Plain Text Corpus Or Biran Columbia University orb@cs.columbia.edu Terra Blevins Columbia University tlb2145@columbia.edu Kathleen McKeown Columbia University kathy@cs.columbia.edu Abstract Finding paraphrases in text is an impor- tant task with implications for genera- tion, summarization and question answer- ing, among other applications. Of par- ticular interest to those applications is the specific formulation of the task where the paraphrases are templated, which provides an easy way to lexicalize one message in multiple ways by simply plugging in the relevant entities. Previous work has fo- cused on mining paraphrases from parallel and comparable corpora, or mining very short sub-sentence synonyms and para- phrases. In this paper we present an ap- proach which combines distributional and KB-driven methods to allow robust mining of sentence-level paraphrasal templates, utilizing a rich type system for the slots, from a plain text corpus. 1 Introduction One of the main difficulties in Natural Language Generation (NLG) is the surface realization of messages: transforming a message from its inter- nal representation to a natural language phrase, sentence or larger structure expressing it. Often the simplest way to realize messages is though the use of templates. For example, any message about the birth year and place of any person can be ex- pressed with the template “[Person] was born in [Place] in [Year]”. Templates have the advantage that the genera- tion system does not have to deal with the inter- nal syntax and coherence of each template, and can instead focus on document-level discourse co- herence and on local coreference issues. On the other hand, templates have two major disadvan- tages. First, having a human manually compose a template for each possible message is costly, espe- cially when a generation system is relatively open- ended or is expected to deal with many domains. In addition, a text generated using templates often lacks variation, which means the system’s output will be repetitive, unlike natural text produced by a human. In this paper, we are concerned with a task aimed at solving both problems: automatically mining paraphrasal templates, i.e. groups of tem- plates which share the same slot types and which, if their slots are filled with the same entities, re- sult in paraphrases. We introduce an unsupervised approach to paraphrasal template mining from the text of Wikipedia articles. Most previous work on paraphrase detection fo- cuses either on a corpus of aligned paraphrase candidates or on such candidates extracted from a parallel or comparable corpus. In contrast, we are concerned with a very large dataset of tem- plates extracted from a single corpus, where any two templates are potential paraphrases. Specifi- cally, paraphrasal templates can be extracted from sentences which are not in fact paraphrases; for example, the sentences “The population of Mis- souri includes more than 1 million African Ameri- cans” and “Roughly 185,000 Japanese Americans reside in Hawaii” can …

Content truncated to 500 words