Using ‘Low-cost’ Learning Features for Pronoun Resolution

Using ‘Low-cost’ Learning Features for Pronoun Resolution Ramon Ré Moya Cuevas and Ivandré Paraboni University of São Paulo, Escola de Artes, Ciências e Humanidades (USP / EACH) Av. Arlindo Bettio, 1000, São Paulo, Brazil. {fusion, ivandre} @usp.br Abstract. We investigate a machine learning approach to Portuguese pronoun resolution. We presently focus on so-called ‘low-cost’ learning features readily obtainable from the output of a part-of-speech tagger, and we largely bypass deep syntactic and semantic analysis. Preliminary results show significant improvement in resolution precision and recall, and are comparable to existing rule-based approaches for the Portuguese language spoken in Brazil. Keywords: Anaphora resolution, Machine Learning. 1. Introduction The computational resolution of anaphoric expressions lies at the heart of a variety of NLP applications, including text understanding, Machine Translation, text summarization and many others. Although it has received a great deal of attention for many years now, anaphora resolution remains a computational problem yet to be overcome (Mitkov, 2002), and a challenge that is considerably increased if we speak of languages for which basic NLP resources (such as parsers, taggers or large corpora) are still under development, or may have only recently become available. This is the case, for instance, of Portuguese, one of the most widely-spoken languages in the world, and which still lacks somewhat behind as a relatively resource-poor language in NLP. In this work we extend our previous investigation on learning approaches to Portuguese personal pronoun resolution in (Cuevas et. al., 2008.) In doing so, we focus on so-called ‘low- cost’ learning features, that is, we will limit the proposed solution to the knowledge readily obtainable from basic NLP tools such as part-of-speech taggers, and we will largely bypass deep syntactic or semantic analysis. In this sense, our work resembles the knowledge-poor approach in Kennedy & Boguraev (1996), which consists of a re-interpretation of the ‘classic’ algorithm proposed in Lappin & Leass (1994) using shallow rather than in-depth analysis. In addition to that, as we do not intend to explicitly write any anaphora resolution algorithms or rules (but rather induce them automatically) our work is mainly related to machine learning approaches such as Soon et. al. (2001), McCarthy and Lehnert (1995) and Ng & Cardie (2002). However, in discussing a possible ‘low-cost’ learning approach to Portuguese third person plural pronouns (“Eles/Elas”), we will focus more on the choice of learning features, and less on the results of a particular machine learning approach, which are to be discussed elsewhere. The rest of this paper is structured as follows. Section 2 reviews previous work taken as the basis for our present investigation. Section 3 proposed an extended set of features for the problem at hand. Results of a standard decision-tree induction algorithm using the new features are presented in Section 4. Finally, Section 5 draws a number of comparisons with related work in Portuguese pronoun resolution and Section 6 describes our future work. 377 22nd Pacific Asia Conference on Language, Information and Computation, pages 377–383 2. Previous Work As in Cuevas et. …

Content truncated to 500 words