Bootstrapping Lexical Choice via Multiple-Sequence Alignment

Bootstrapping Lexical Choice via Multiple-Sequence Alignment Regina Barzilay Department of Computer Science Columbia University New York, NY 10027 USA regina@cs.columbia.edu Lillian Lee Department of Computer Science Cornell University Ithaca, NY 14853 USA llee@cs.cornell.edu Abstract An important component of any generation system is the mapping dictionary, a lexicon of elementary semantic expressions and cor- responding natural language realizations. Typically, labor-intensive knowledge-based methods are used to construct the dictio- nary. We instead propose to acquire it automatically via a novel multiple-pass al- gorithm employing multiple-sequence align- ment, a technique commonly used in bioin- formatics. Crucially, our method lever- ages latent information contained in multi- parallel corpora — datasets that supply several verbalizations of the corresponding semantics rather than just one. We used our techniques to generate natural language versions of computer-generated mathematical proofs, with good results on both a per-component and overall-output basis. For example, in evaluations involv- ing a dozen human judges, our system pro- duced output whose readability and faith- fulness to the semantic input rivaled that of a traditional generation system. 1 Introduction One or two homologous sequences whisper . . . a full multiple alignment shouts out loud (Hubbard et al., 1996). Today’s natural language generation systems typically employ a lexical chooser that translates complex semantic concepts into words. The lex- ical chooser relies on a mapping dictionary that lists possible realizations of elementary seman- tic concepts; [sex:female]]sample entries might be [Parent → mother or love(x,y )→ {x loves y, x is in love with y}.1 To date, creating these dictionaries has involved human analysis of a domain-relevant corpus com- prised of semantic representations and correspond- ing human verbalizations (Reiter and Dale, 2000). The corpus analysis and knowledge engineering work required in such an approach is substantial, pro- hibitively so in large domains. But, since corpus data is already used in building lexical choosers by hand, an appealing alternative is to have the system learn a mapping dictionary directly from the data. Clearly, this would greatly reduce the human effort involved and ease porting the system to new domains. Hence, we address the following problem: given a parallel (but unaligned) corpus consisting of both complex semantic input and corresponding natural language verbalizations, learn a semantics-to-words mapping dictionary automatically. Now, we could simply apply standard statistical machine translation methods, treating verbalizations as “translations” of the semantics. These meth- ods typically rely on one-parallel corpora consist- ing of text pairs, one in each “language” (but cf. Simard (1999); see Section 5). However, learning the kind of semantics-to-words mapping that we desire from one-parallel data alone is difficult even for hu- mans. First, given the same semantic input, differ- ent authors may (and do) delete or insert informa- tion (see Figure 1); hence, direct comparison between a semantic text and a single verbalization may not provide enough information regarding their underly- ing correspondences. Second, a single verbalization certainly fails to convey the variety of potential lin- guistic realizations of the concept that an expressive lexical chooser would ideally have access to. …

Content truncated to 500 words