Semantic Frame Recognition Of Chinese Unregistered Verbs Based On Semantic Similarity

In the field of Natural Language Processing, one important way of acquiring semantic information is Semantic Role labelling (SRL).There are two kinds of ways to accomplish the task of Chinese semantic role labeling, one is the supervised algorithm ,another is unsupervised(semi-supervised) way. Because the lack of training corpora, we do the task of Chinese SRL by semi-supervised way. The main disadvantage of the way is the problem of data sparsity. When we construct the system of SRL, we use two lexicons. They are verb lexicon of semantic frame and noun lexicon. The problem of data sparsity is raised when a word is out of pre-defined lexicons, In our system, the verb lexicon of semantic frame is the most important, so we do our best to solve the problem of assigning semantic frame for out-of-vocabulary(OOV) Chinese verbs.In this paper, we propose a method based on semantic similarity to assign a semantic frame for a out-of-vocabulary verb. In our method, we firstly use the word similarity between a OOV verb and a in-vocabulary(IV) verb to assign candidate semantic frames for IV verb, and then we choose a target frame from the candidates through matching semantic frame with syntactic structures of sentence that contains the OOV verb. We have used two algorithms to measure the semantic similarity between OOV and IV verbs, One is based on TongYiCiCiLin, another is based on How-Net. The experiments showed that the labelling accuracy is improved from 72% to 76% when we add a TongYiCiCilin based OOV verb semantic frame assignment module to the SRL system. The experiment show that the labelling accuracy improved from 72% to 79% when we add a How-Net based Semantic Frame assignment module to the SRL system. The results show the efficiency of our way to solve the problem of OOV verbs in SRL task.