Acquiring Predicate-Argument Mapping Information from Multilingual Texts

Acquiring Predicate-Argument Mapping Information from Multilingual Texts Chinatsu Aone, Douglas McKee Systems Research and Applications (SRA) 2000 15th Street North Arlington, VA 22201 aonec@sra, corn, mckeed@sra.com Abstract This paper discusses automatic acquisition of predicate-argument mapping in- formation from multilingual texts. The lexicon of our NLP system abstracts the language-dependent portion of predicate-argument mapping information from the core meaning of verb senses (i.e. semantic concepts as defined in the knowledge base). We represent this mapping information in terms of cross-linguistically generalized mapping types called situation types and word sense-specific idiosyncrasies. This representation has enabled us to automatically acquire predicate-argument map- ping information, specifically situation types and idiosyncrasies, for verbs in English, Spanish, and Japanese texts. 1 Introduction Lexicons for a natural language processing (NLP) system that perform syntactic and semantic analysis require more than purely syntactic (e.g. part-of-speech information) and semantic information (e.g. a concept hierarchy). Language understanding requires mapping from syntactic structures into conceptual representation (henceforth predicate- argument mapping), while language generation requires the inverse mapping. That is, grammatical functions in the syntactic structures (e.g. subject, object, etc.) should be mapped to thematic roles in the semantic structures (e.g. agent, theme, etc.). In this paper, we discuss how we acquire such predicate-argument mapping information from multilingual texts automatically (cf. Zernik and Jaeobs work on collecting thematic roles [20]). As discussed in Aone and Mckee [1], the lexicon of our NLP system abstracts the language-dependent portion of predicate-argument mapping information from the core meaning of verb senses (i.e. semantic concepts as defined in the knowledge base). We represent this mapping information in terms of cross-linguistically generalized mapping types called situation types and word sense-specific idiosyncrasies. This representation has enabled us to automatically acquire predicate-argument mapping information, specifically situation types and idiosyncrasies, for verbs in English, Spanish, an.d Japanese texts. In the following sections, we first describe how we represent the predicate-mapping information. Then, we discuss how we acquire situation type and idiosyncrasy information automatically from multilingual texts and show some results. 2 Predicate-Argument Mapping Representation Each lexical sense of a verb in our lexicon encodes its default predicate-argument mapping type (i.e. situation type), any word-specific mapping exceptions (i.e. idiosyncrasies), and 107 # of required NP or S arguments CAUSED-PROCESS 2 PROCESS-OR-STATE 1 AGENTIVE-ACTION 1 INVEKSE-STATE 2 default thematic roles prohibited thematic roles Asent Theme Theme Agent Agent Goal Theme Agent Table 1: Definitions of Situation Types English Spanish Japanese CAUSED-PROCESS kill matar, mirar korosu, miru PROCESS-OH-STATE die morir shibousuru AGENTIVE-ACTION look bailar odoru INVERSE-STATE see vet mieru Table 2: Situation Types and Verbs in Three Languages its semantic meaning (i.e. semantic concept) in addition to its morphological and syntactic information. In the following, we discuss these three levels in detail. 2.1 Situation Types Each of a verb's lexical senses is classified into one of the four default predicate-argument mapping types called situation types. As shown in Table 1, situation types of verbs are defined by two kinds of information: 1) the number of subcategorized NP or S arguments and 2) the …

Content truncated to 500 words