Em busca do plausível: desafios e contribuições da Geração Automática de Língua Natural

Abstract

consideram tanto modelos de locutores individuais como de grupos de indivíduos e, de modo especial, baseados em personalidade humana. Como resultado deste trabalho, espera-se que novos pesquisadores e estudantes (e.g., de Ciência da Computação eáreas relacionadas) sintam-se motivados a abordar o problema computacional de GLN de modo plausível -seja com base na perspectiva do ouvinte, do locutor, ou perfis destes -e a desenvolver suas próprias metodologias, auxiliando assim no crescimento da linha de pesquisa em GLN de modo geral, e de forma mais específica para o caso da geração de texto em português do Brasil. ABSTRACT Natural Language Generation (NLG) systems -which typically generate text descriptions from non-linguistic input data -are employed when canned text is not sufficient, that is, when greater linguistic or stylistic variation are expected in the output text, or when the application requires closer proximity to human performance. This work addresses the computational problem of plausible NLG based on two perspectives -hereby called the speaker's and the hearer's perspectives -by presenting a number of examples of studies on content selection, surface realisation and other NLG tasks, and by focusing on their research challenges and contributions. The initial discussion is based on examples of studies that follow a rather traditional (or perhaps more intuitive) view on plausibility based on the speaker's perspective.Next, it is argued that in speaker-oriented studies of this kind further research questions will generally require collecting new data. As an alternative to this view, examples of heareroriented studies are discussed, which may turn out to be highly reusable and, in some cases, provide contributions that are otherwise difficult to obtain by means of speaker-oriented studies. Finally, the speaker and hearer perspectives are generalised so as to cover the issue of human variation in language production by discussing examples of studies that taken into account models of individual speakers or groups and, in particular, using models of human personality. As a result, this text is expected to help motivate new researchers and students (e.g., in Computer Science and related fields) to address the computational problem of plausible natural language generation -either from a speaker's or hearer's perspectiveand to develop their own methodologies, hence contributing to the expansion of the NLG research field in general, and particularly so for the case of the Brazilian Portuguese text generation.