Generation by Inverting a Semantic Parser that Uses Statistical Machine Translation

Proceedings of NAACL HLT 2007, pages 172–179, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Generation by Inverting a Semantic Parser That Uses Statistical Machine Translation Yuk Wah Wong and Raymond J. Mooney Department of Computer Sciences The University of Texas at Austin 1 University Station C0500 Austin, TX 78712-0233, USA {ywwong,mooney}@cs.utexas.edu Abstract This paper explores the use of statisti- cal machine translation (SMT) methods for tactical natural language generation. We present results on using phrase-based SMT for learning to map meaning repre- sentations to natural language. Improved results are obtained by inverting a seman- tic parser that uses SMT methods to map sentences into meaning representations. Finally, we show that hybridizing these two approaches results in still more accu- rate generation systems. Automatic and human evaluation of generated sentences are presented across two domains and four languages. 1 Introduction This paper explores the use of statistical machine translation (SMT) methods in natural language gen- eration (NLG), specifically the task of mapping statements in a formal meaning representation lan- guage (MRL) into a natural language (NL), i.e. tacti- cal generation. Given a corpus of NL sentences each paired with a formal meaning representation (MR), it is easy to use SMT to construct a tactical gener- ator, i.e. a statistical model that translates MRL to NL. However, there has been little, if any, research on exploiting recent SMT methods for NLG. In this paper we present results on using a re- cent phrase-based SMT system, PHARAOH (Koehn et al., 2003), for NLG.1 Although moderately effec- 1We also tried IBM Model 4/REWRITE (Germann, 2003), a word-based SMT system, but it gave much worse results. tive, the inability of PHARAOH to exploit the for- mal structure and grammar of the MRL limits its ac- curacy. Unlike natural languages, MRLs typically have a simple, formal syntax to support effective au- tomated processing and inference. This MRL struc- ture can also be used to improve language genera- tion. Tactical generation can also be seen as the inverse of semantic parsing, the task of mapping NL sen- tences to MRs. In this paper, we show how to “in- vert” a recent SMT-based semantic parser, WASP (Wong and Mooney, 2006), in order to produce a more effective generation system. WASP exploits the formal syntax of the MRL by learning a trans- lator (based on a statistical synchronous context- free grammar) that maps an NL sentence to a lin- earized parse-tree of its MR rather than to a flat MR string. In addition to exploiting the formal MRL grammar, our approach also allows the same learned grammar to be used for both parsing and genera- tion, an elegant property that has been widely ad- vocated (Kay, 1975; Jacobs, 1985; Shieber, 1988). We present experimental results in two domains pre- viously used to test WASP’s semantic parsing abil- ity: mapping NL queries to a formal database query language, and mapping NL soccer coaching instruc- tions to a formal robot command language. WASP−1 is shown to produce a more …

Content truncated to 500 words