Unsupervised Semantic Parsing

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1–10, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Unsupervised Semantic Parsing Hoifung Poon Pedro Domingos Department of Computer Science and Engineering University of Washington Seattle, WA 98195-2350, U.S.A. {hoifung,pedrod}@cs.washington.edu Abstract We present the first unsupervised approach to the problem of learning a semantic parser, using Markov logic. Our USP system transforms dependency trees into quasi-logical forms, recursively induces lambda forms from these, and clusters them to abstract away syntactic variations of the same meaning. The MAP semantic parse of a sentence is obtained by recur- sively assigning its parts to lambda-form clusters and composing them. We evalu- ate our approach by using it to extract a knowledge base from biomedical abstracts and answer questions. USP substantially outperforms TextRunner, DIRT and an in- formed baseline on both precision and re- call on this task. 1 Introduction Semantic parsing maps text to formal meaning representations. This contrasts with semantic role labeling (Carreras and Marquez, 2004) and other forms of shallow semantic processing, which do not aim to produce complete formal meanings. Traditionally, semantic parsers were constructed manually, but this is too costly and brittle. Re- cently, a number of machine learning approaches have been proposed (Zettlemoyer and Collins, 2005; Mooney, 2007). However, they are super- vised, and providing the target logical form for each sentence is costly and difficult to do consis- tently and with high quality. Unsupervised ap- proaches have been applied to shallow semantic tasks (e.g., paraphrasing (Lin and Pantel, 2001), information extraction (Banko et al., 2007)), but not to semantic parsing. In this paper we develop the first unsupervised approach to semantic parsing, using Markov logic (Richardson and Domingos, 2006). Our USP sys- tem starts by clustering tokens of the same type, and then recursively clusters expressions whose subexpressions belong to the same clusters. Ex- periments on a biomedical corpus show that this approach is able to successfully translate syntac- tic variations into a logical representation of their common meaning (e.g., USP learns to map active and passive voice to the same logical form, etc.). This in turn allows it to correctly answer many more questions than systems based on TextRun- ner (Banko et al., 2007) and DIRT (Lin and Pantel, 2001). We begin by reviewing the necessary back- ground on semantic parsing and Markov logic. We then describe our Markov logic network for un- supervised semantic parsing, and the learning and inference algorithms we used. Finally, we present our experiments and results. 2 Background 2.1 Semantic Parsing The standard language for formal meaning repre- sentation is first-order logic. A term is any ex- pression representing an object in the domain. An atomic formula or atom is a predicate symbol ap- plied to a tuple of terms. Formulas are recursively constructed from atomic formulas using logical connectives and quantifiers. A lexical entry de- fines the logical form for a lexical item (e.g., a word). The semantic parse of a sentence is de- rived by starting …

Content truncated to 500 words