Learning Textual Entailment Classification from a Chinese RITE Dataset Specialized for Linguistic Phenomena

This paper proposed a textual entailment classification system developed based on a dataset focusing on individual entailment-related linguistic phenomena. Identical and synonymous terms in the text pair were aligned and ignored. Several groups of classification rules have been proposed with respect to the difference between the sentences in the text pair. The set of Wikipedia redirected titles became an important resource of synonyms. A SVM classifier was also constructed based on several groups of features. The experimental results showed that a hybrid system achieved the best performance with a macro-averaged F1-measure of 48.50% and an accuracy of 49.33% which outperformed the best systems in the NTCIR-11 RITE-VAL System Validation Chinese subtasks.