Assisting Verb Selection for ESL: Considering Error Patterns Produced by Learners

澤井 悠 (1151054)


Selecting appropriate verbs based on the sentence context is difficult particularly for ESL learners. When treating this task as classification problem, restricting possible class label by confusion sets is crucial, because the number of possible candidate is quite large. In previous works, they have not used confusion sets reflecting error patterns. In this presentation, I propose a new technique for detecting and suggesting alternatives for verb selection errors using multi-class classification to address these issues. To broaden the coverage of verb selection system, I construct the confusion sets of verbs from a large scale web derived learner corpus. The proposed model is trained using both large scale native corpus and the web derived ESL corpus via domain adaptation technique. Experiments are carried out on a public available ESL dataset, the CLC-FCE dataset. The results show that proposed verb error confusion sets and domain adaptation technique increased the coverage of learners' errors. It also gains the performance of correction candidate suggestion. An inter-corpus evaluation on the KJ corpus shows that the proposed method works generally effective. For those contributions, I especially focus on the effect of proposed methods for suggestion.