Semantic Parsing of Ambiguous Input through Paraphrasing and Verification

Semantic Parsing of Ambiguous Input through Paraphrasing and Verification Philip Arthur, Graham Neubig, Sakriani Sakti, Tomoki Toda, Satoshi Nakamura Graduate School of Information Science, Nara Institute of Science and Technology, Japan {philip.arthur.om0, neubig, ssakti, tomoki, s-nakamura}@is.naist.jp Abstract We propose a new method for semantic pars- ing of ambiguous and ungrammatical input, such as search queries. We do so by build- ing on an existing semantic parsing framework that uses synchronous context free grammars (SCFG) to jointly model the input sentence and output meaning representation. We gener- alize this SCFG framework to allow not one, but multiple outputs. Using this formalism, we construct a grammar that takes an ambigu- ous input string and jointly maps it into both a meaning representation and a natural lan- guage paraphrase that is less ambiguous than the original input. This paraphrase can be used to disambiguate the meaning representa- tion via verification using a language model that calculates the probability of each para- phrase.1 1 Introduction Semantic parsing (SP) is the problem of parsing a given natural language (NL) sentence into a meaning representation (MR) conducive to further processing by applications. One of the major challenges in SP stems from the fact that NL is rife with ambiguities. For example, even the simple sentence “Where can we eat a steak in Kobe?” contains syntactic ambi- guities (“eat in Kobe” or “steak in Kobe”?), quan- tifier scope ambiguities (do we all eat one steak, or each eat one steak?), and word sense ambigui- ties (is Kobe a city in Japan; or an NBA basketball 1Tools to replicate our experiments can be found at http://isw3.naist.jp/~philip-a/tacl2015/index.html. player?). Previous works using statistical models along with formalisms such as combinatorial cat- egorial grammars, synchronous context free gram- mars, and dependency based compositional seman- tics have shown notable success in resolving these ambiguities (Zettlemoyer and Collins, 2005; Wong and Mooney, 2007; Liang et al., 2011; Kwiatkowski et al., 2013). Much previous work on SP has focused on the case of answering natural language queries to a database of facts, where the queries generally take the form of full sentences such as “What is the height of Kobe Bryant?” While answering these ques- tions provides an excellent first step to natural lan- guage information access, in many cases the input is not a full sentence, but something more underspec- ified and ungrammatical. For example, this is the case for keyword-based search queries (Sajjad et al., 2012) or short dialogue utterances (Zettlemoyer and Collins, 2007). Specifically taking the example of search queries, users tend to omit some of the function words and grammatical constructs in the language to make a more concise query. The first column of Table 1 illustrates several search queries of the pattern “Kobe X” where X is another word. From these queries and their MRs in column two, we can see that there are several kinds of ambiguity, including not only the distinction between Kobe as city or a basketball player as in the previous example, but also more …

Content truncated to 500 words