Broad Coverage Multilingual Deep Sentence Generation with a Stochastic Multi-Level Realizer

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 98–106, Beijing, August 2010 Broad Coverage Multilingual Deep Sentence Generation with a Stochastic Multi-Level Realizer Bernd Bohnet1, Leo Wanner1,2, Simon Mille1, Alicia Burga1 1Department of Information and Communication Technologies Pompeu Fabra University 2Instituci´o Catalana de Recerca i Estudis Avanc¸ats (ICREA) {first-name.last-name}@upf.edu Abstract Most of the known stochastic sentence generators use syntactically annotated corpora, performing the projection to the surface in one stage. However, in full-fledged text generation, sentence realization usually starts from semantic (predicate-argument) structures. To be able to deal with semantic structures, stochastic generators require semantically annotated, or, even better, multilevel an- notated corpora. Only then can they deal with such crucial generation issues as sentence planning, linearization and mor- phologization. Multilevel annotated cor- pora are increasingly available for multi- ple languages. We take advantage of them and propose a multilingual deep stochastic sentence realizer that mirrors the state-of- the-art research in semantic parsing. The realizer uses an SVM learning algorithm. For each pair of adjacent levels of anno- tation, a separate decoder is defined. So far, we evaluated the realizer for Chinese, English, German, and Spanish. 1 Introduction Recent years saw a significant increase of inter- est in corpus-based natural language generation (NLG), and, in particular, in corpus-based (or stochastic) sentence realization, i.e., that part of NLG which deals with mapping of a formal (more or less abstract) sentence plan onto a chain of in- flected words; cf., among others, (Langkilde and Knight, 1998; Oh and Rudnicky, 2000; Bangalore and Rambow, 2000; Wan et al., 2009). The advan- tage of stochastic sentence realization over tradi- tional rule-based realization is mainly threefold: (i) it is more robust, (ii) it usually has a signifi- cantly larger coverage; (iii) it is per se language- and domain-independent. Its disadvantage is that it requires at least syntactically annotated corpora of significant size (Bangalore et al., 2001). Given the aspiration of NLG to start from numeric time series or conceptual or semantic structures, syn- tactic annotation even does not suffice: the cor- pora must also be at least semantically annotated. Up to date, deep stochastic sentence realization was hampered by the lack of multiple-level an- notated corpora. As a consequence, available stochastic sentence generators either take syntac- tic structures as input (and avoid thus the need for multiple-level annotation) (Bangalore and Ram- bow, 2000; Langkilde-Geary, 2002; Filippova and Strube, 2008), or draw upon hybrid models that imply a symbolic submodule which derives the syntactic representation that is then used by the stochastic submodule (Knight and Hatzivas- siloglou, 1995; Langkilde and Knight, 1998). The increasing availability of multilevel anno- tated corpora, such as the corpora of the shared task of the Conference on Computational Natu- ral Language Learning (CoNLL), opens new per- spectives with respect to deep stochastic sentence generation—although the fact that these corpora have not been annotated with the needs of genera- tion in mind, may require additional adjustments, as has been, in fact, in the case of our work. 98 …

Content truncated to 500 words