Construction of predicate argument structure annotator based on event type thesaurus

抄録

In this research we construct a predicate argument structure annotator that can disambiguate predicate's semantic concept and semantic role types of its arguments in a running texts. The semantic concepts are manually constructed and systematically organized by event type of predicates according to linguistically-motivated lexical semantic structure. For example the following predicates ``get'', ``take'', ``buy'', ``obtain'', ``take'', ``rob'' and ``rent'' must contain a shared meaning of 'Getting from something'; and the difference of ``get'' and ``rent'' must be with/without ownership of the obtained object. We describe this semantic shared/different meaning with abstracted LCS-base structure in our Thesaurus, e.g., The semantic structure of ``get'' might be ([Agent] CAUSE BECOME [Theme] BE AT [Goal=person]), and ``rent'' might be more detailed structure as [Agent] CAUSE (BY MEANS OF [Agent] renting [Theme]) BECOME [Theme] BE AT Goal=person]). This thesaurus is continuously constructed taking into account what are the base semantic structure among all Japanese predicates. Thus by making predicate argument structure annotator (we call ASA) based on thesaurus, we can see how the thesaurus's concepts can catch the same expressions of natural language. For this purpose, ASA can recognized idiom and equivalent expressions of predicates e.g., ``X ga Y ni hone-wo oru (X gives oneself trouble about Y)'' and ``X ga Y ni kurou-suru (X has difficulty with Y)'' with disambiguating verb senses ``hone-wo oru (gives oneself trouble about/break one's back)''. By constructing predicate argument structure with systematic thesaurus, we try to convert running texts into more semantically controllable descriptions.