Scalable Semantic Parsing with Partial Ontologies

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1311–1320, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Scalable Semantic Parsing with Partial Ontologies Eunsol Choi Tom Kwiatkowski† Luke Zettlemoyer Computer Science & Engineering University of Washington eunsol@cs.washington.edu, tomkwiat@google.com, lsz@cs.washington.edu Abstract We consider the problem of building scal- able semantic parsers for Freebase, and present a new approach for learning to do partial analyses that ground as much of the input text as possible without requiring that all content words be mapped to Freebase concepts. We study this problem on two newly introduced large-scale noun phrase datasets, and present a new semantic pars- ing model and semi-supervised learning approach for reasoning with partial onto- logical support. Experiments demonstrate strong performance on two tasks: refer- ring expression resolution and entity at- tribute extraction. In both cases, the par- tial analyses allow us to improve precision over strong baselines, while parsing many phrases that would be ignored by existing techniques. 1 Introduction Recently, significant progress has been made in learning semantic parsers for large knowledge bases (KBs) such as Freebase (FB) (Cai and Yates, 2013; Berant et al., 2013; Kwiatkowski et al., 2013; Reddy et al., 2014). Although these methods can build general purpose meaning representations, they are typically evaluated on question answering tasks and are designed to only parse questions that have complete ontological coverage, in the sense that there exists a logical form that can be executed against Freebase to get the correct answer.1 In this paper, we instead consider the problem of learning semantic parsers for open domain text containing †Now at Google, NY. 1To ensure all questions are answerable, the data is man- ually filtered. For example, the WebQuestions dataset intro- duced by Berant et al. (2013) contains only the 7% of the originally gathered questions. Wikipedia Haitian human rights activists Art museums and galleries in New York School buildings completed in 1897 Olympic gymnasts of Norway Appos. the capital of quake-hit Sichuan Province a major coal producing province the relaxed seaside capital of Mozambique Figure 1: Example noun phrases from Wikipedia category labels and appositives in newswire text. concepts that may or may not be representable us- ing the Freebase ontology. Even very large knowledge bases have two types of incompleteness that provide challenges for se- mantic parsing algorithms. They (1) have partial ontologies that cannot represent the meaning of many English phrases and (2) are typically miss- ing many facts. For example, consider the phrases in Figure 1. They include subjective or otherwise unmodeled phrases such as “relaxed” and “quake- hit.” Freebase, despite being large-scale, contains a limited set of concepts that cannot represent the meaning of these phrases. They also refer to enti- ties that may be missing key facts. For example, a recent study (West et al., 2014) showed that over 70% of people in FB have no birth place, and 99% have no ethnicity. In our work, …

Content truncated to 500 words