Using bilingual corpora for the construction of contrastive generation grammars: issues and problems

This paper reports on the use of corpora for the construction of a computational grammar of Spanish, contrastive with English, in the application context of Multilingual Natural Language Generation (MLG). The theoretical framework for this work is Systemic Functional Linguistics (SFL) and the computational context provided by KPML (Komet Penman Multilingual), an extensive grammar development environment and generation engine that supports large-scale multilingual development (Bateman 1997). The initial phenomena which are being investigated contrastively belong to three different functional regions of the grammar, i.e., particular subareas of the grammar that are concerned with particular areas of meanings. These regions are transitivity (ideational meaning), thematicity (textual meaning) and mood (interpersonal meaning). The present study concentrates on textual meaning (thematicity and focus) as an illustration. Following what has now established itself as a standard methodology for empirically-based Natural Language Generation (Bateman 1998a, Reiter and Dale 1997), the following steps were carried out: first, a bilingual corpus (English-Spanish) was selected. For Spanish, a sample of spoken texts from the MacroCorpus of the educated linguistic standard of the main cities of the Spanish-speaking world was used, while for English a comparable sample was selected from the British National Corpus Sampler. This was motivated by the need to provide a realistic account of the behaviour of the linguistic phenomena investigated in unplanned and spontaneous contexts of use. The second step was to carry out a contrastive analysis of the phenomena mentioned before. Finally, the results of the analysis were coded up as resources/processes for generation. In the case of Spanish, these had to be created anew. In the case of English, as the KPML already includes an English generation grammar, this last step consisted on checking the coverage of the existing specifications and extending them when could not cover the instances found in the corpus, and adapting them for effective MLG. Given the nature of the NLG process, which typically converts communicative goals expressed in some internal representation into surface forms, the kind of information that is most readily usable for NLG are statements of mappings from functions to forms. Therefore, the corpus analysis phase for NLG usually includes an explicit, and usually quite lengthy linguistic analysis where the analyst seeks possible realisations of communicative functions, which restricts the size of the corpus that can be realistically considered. This paper describes the different steps carried out for the generation of the linguistic phenomena mentioned above, …