Automatic Pronoun Resolution for Swedish Automatic Pronoun Resolution for Swedish

This thesis describes SwePron, an algorithm for automatic resolution of pronouns in Swedish, using Mitkov’s algorithm for pronoun resolution in English (MARS) as a starting point. There are several changes made to the algorithm in order to adapt it to the Swedish language. One key language difference is that Swedish has two “neutral” genders, whereas English only has one. A second important difference is that the word order varies more in Swedish than it does in English. A number of other modifications, not directly related to language differences, are also investigated. It is argued that incorporating lexical information is an important element in increasing the accuracy of the algorithm. One successful example is the use of lists of “human role nouns”, i.e. nouns that normally refer to human beings, although this fact is not reflected in the grammatical gender. However, other aspects of lexical information are also used. Several attempts are made to utilize collocation information to improve the accuracy of the algorithm. The results are disappointing, but this is probably caused by the text database not being large enough. Furthermore, machine learning techniques could be used to improve the evaluation of the collocation information. It is argued that the emphasis on the distance between the anaphor and its antecedent should probably be increased compared to what it is in MARS. The algorithm is implemented in Java. It makes use of two existing applications, Granska ́s Text Analyzer (GTA) and MaltParser/sweMalt. These two applications together provide the syntactic parsing information that is used by the algorithm. The implemented algorithm is evaluated for third person singular pronouns. It will be quite straightforward to extend the algorithm to also resolve plural pronouns, and some guidelines for this are given. The performance of SwePron seems to be roughly similar to that of MARS, although a detailed comparison is difficult due to different coverage, and also differences in syntactic parsing information. It is suggested that some of the modifications tried in this thesis could perhaps be evaluated also on pronoun resolution for English; two options are the human role noun list, and increasing the weight of the referential distance factor.