Automatic Semantic Role Labeling using Selectional Preferences with Very Large Corpora

OF PhD THESIS Automatic Semantic Role Labeling using Selectional Preferences with Very Large Corpora Determinación Automática de Roles Semánticos usando Preferencias de Selección sobre Corpus muy Grandes Graduated: Hiram Calvo Center for Research in Computing (CIC) National Polytechnic Institute (IPN) Mexico City, Mexico, 07738 [email protected] [email protected] Graduated on June 19th, 2006 Advisor: Dr. Alexander Gelbukh Computing Research Center (CIC) National Polytechnic Institute (IPN) Mexico City, Mexico, 07738 www.gelbukh.com Abstract We present a method for recognizing semantic roles for Spanish sentences. This method is based on dependency parsing using heuristic rules to infer dependency relationships between words, and word co-occurrence statistics (learnt in an unsupervised manner) to resolve ambiguities such as prepositional phrase attachment. If a complete parse cannot be produced, a partial structure is built with some (if not all) dependency relations identified. Evaluation shows that in spite of its simplicity, the parser’s accuracy is superior to the available existing parsers for Spanish. Though certain grammar rules, as well as the lexical resources used, are specific for Spanish, the suggested approach is language-independent. A particularly interesting ambiguity which we have decided to analyze deeper, is the Prepositional Phrase Attachment Disambiguation. The system uses an ordered set of simple heuristic rules for determining iteratively the relationships between words to which a governor has not been yet assigned. For resolving certain cases of ambiguity we use cooccurrence statistics of words collected previously in an unsupervised manner, whether it be from big corpora, or from the Web (through a search engine such as Google). Collecting these statistics is done by using Selectional Preferences. In order to evaluate our system, we developed a Method for Converting a Gold Standard from a constituent format to a dependency format. Additionally, each one of the modules of the system (Selectional Preferences Acquisition and Prepositional Phrase Attachment Disambiguation), is evaluated in a separate and independent way to verify that they work properly. Finally we present some Applications of our system: Word Sense Disambiguation and Linguistic Steganography.