Brockmann (2002):
Evaluating and Combining Approaches to
Selectional Preference Acquisition

Brockmann, Carsten (2002). Evaluating and combining approaches to selectional preference acquisition. Master's thesis, Universität des Saarlandes, Saarbrücken, Germany.

Summary

Selectional preferences are graded constraints that predicates impose upon their arguments. They can be used for tasks like parse ranking or word sense disambiguation, or can be included in dictionary entries. In the last decade, various methods have been suggested to automatically acquire selectional preference information from corpora.

There have been many evaluations of the automatically acquired selectional preference information. However, most have been conducted for English, only for the verb/direct object relation, and looking at algorithms in isolation. Open questions are whether the methods are valid cross-linguistically, how they perform on other grammatical relations, and how they perform in combination. This information can justify or rebut the usefulness of state of the art techniques for selectional preference acquisition.

The research presented in this thesis attempts to fill these gaps. I conducted a magnitude estimation experiment over the world-wide web to elicit human judgments on the acceptability of 90 German sentences with intransitive and transitive verbs, and verbs subcategorizing for a PP object. These judgments were compared to the preferences computed by five approaches to automatic selectional preference acquisition: Two frequency-based measures, selectional association (Resnik 1993), tree cut models (Li and Abe 1998), and the similarity-class measure (Clark and Weir 2001).

I found that there exist significant linear correlations between the human judgments and different single algorithms, depending on the grammatical relation being observed. Furthermore, when methods are combined by a principal components factor analysis and multiple regression, they outperform the results of a single approach when judging all of the test sentences.

BibTeX Entry

@MastersThesis{b02:ecaspa,
  author =	 {Brockmann, Carsten},
  title =	 {Evaluating and Combining Approaches to Selectional
                  Preference Acquisition},
  school =	 {Universit{\"{a}}t des Saarlandes},
  year =	 2002,
  address =	 {Saarbr{\"{u}}cken, Germany}
}