人間の学習モデルCORESによる代名詞と省略語の処理

抄録

The purpose of this study is to make CORES (Contextual and Relative Existence System), which can learn conceptual relationships based on only a small number of instances under various contexts and relativity, as the first step of making an artificial intelligence model of human natural language processing by applying it first to pronoun resolution and inference of ellipses. Based on many psychological findings concerning pronoun resolution and ellipsis inference, it was hypothesized that early incremental interpretation should be necessary, that evaluation should be based on lexical information, temporal proximity, and verbs' relational information, and that continuation of every process to the end should be indispensable. These hypotheses were first tested and supported by a psychological experiment with 84 subjects. CORES was then improved based on the experimental results and adopted some new components and variables such as affinity nodes, temporal proximity nodes, psychological distances, and activation levels. Finally, the validity of new CORES was confirmed by simulations using the experimental materials. These results were discussed in terms of the necessity of utilization of world knowledge or common sense and that of further exploration and introduction of heuristics.