When two texts have an inclusion relation, the relationship between them is called entailment. The task of mechanically distinguishing such a relation is called recognising textual entailment (RTE), which is basically a kind of semantic analysis. A variety of methods have been proposed for RTE. However, when the previous methods were combined, the performances were not clear. So, we utilized each method as a feature of machine learning, in order to combine methods. We have dealt with the binary classification problem of two texts exhibiting inclusion, and proposed a method that uses machine learning to judge whether the two texts present the same content. We have built a program capable to perform entailment judgment on the basis of word overlap, i.e. the matching rate of the words in the two texts, mutual information, and similarity of the respective syntax trees (Subpath Set). Word overlap was calclated by utilizing BiLingual Evaluation Understudy (BLEU). Mutual information is based on co-occurrence frequency, and the Subpath Set was determined by using the Japanise WordNet. A Confidence- Weighted Score of 68.6% was obtained in the mutual information experiment on RTE. Mutual information and the use of three methods of SVM were shown to be effective.
展开▼
