Anaphor resolution in unrestricted texts with partial parsing

Anaphor resolution in unrestricted texts with partial parsing t A. Ferrfindez; M. Palomar Dept. Languages and Information Systems Alicante University - Apt. 99 03080 - Alicante - Spain antonio@dlsi.ua.es mpalomar@dlsi.ua.es L. Moreno Dept. Information Systems and Computation Valencia University of Technology lmoreno@dsic.upv.es Abstract In this paper we deal with several kinds of anaphora in unrestricted texts. These kinds of anaphora are pronominal references, surface- count anaphora and one-anaphora. In order to solve these anaphors we work on the output of a part-of-speech tagger, on which we automatically apply a partial parsing from the formalism: Slot Unification Grammar, which has been implemented in Prolog. We only use the following kinds of intbrmation: [exical (the lemma of each word), morphologic (person, number, gender) and syntactic. Finally we show the experimental results, and the restrictions and preferences that we have used for anaphor resolution with partial parsing. Introduction Nowadays there are two different approaches to anaphor resolution: integrated and alternative. The former is based on the integration of different kinds of knowledge (e.g. syntactic or semantic information) whereas the latter is based on statistical, neural networks or the principles of reasoning with uncertainty: e.g. Connoly (1994) and Mitkov (1997). Our system can be included into the first approach. In these integrated approaches the semantic and domain knowledge information is very expensive in relation to computational processing. As a consequence, current anaphor resolution implementations mainly rely on constraints and preference heuristics which employ information originated from morphosyntactic or shallow semantic analysis, e.g. in Baldwin (1997). These approaches, however, perform remarkably well. In Lappin and Leass (1994) it is described an algorithm for pronominal anaphor resolution with a high rate of correct analyses: 85%. This one operates primarily on syntactic infbrmation only. In Kennedy and Boguraev (1996) it is proposed an algorithm for anaphor resolution which is a modified and extended version of that developed by Lappin and Leass (1994). In contrast to that work, this algorithm does not require in-depth, full, syntactic parsing of text. The modifications enable the resolution process to work from the output of a POS tagger, enriched only with annotations of grammatical function of lexical items in the input text stream. The advantage of this algorithm is that anaphor resolution can be realized within NLP frameworks which do not -or cannot- employ robust and reliable parsing components. Quantitative evaluation shows the anaphor resolution algorithm described here to run at a rate of 75% accuracy. Our framework will allow us a similar approach to that of Kennedy and Boguraev (1996), but we will automatically get syntactic information from partial parsing. Moreover, our proposal will also be applied to other kinds of anaphors such as surface-count anaphora or one-anaphora. There are some other approaches that work on the output of a POS tagger, e.g. that of Mitkov and Stys (1997), in which it is proposed another knowledge-poor approach to resolving pronouns in technical manuals in both English and Polish. This approach is a modification of the reported in Mitkov (1997). Here, the …

Content truncated to 500 words