Semantic role labeling (SRL) is a fundamentaltask in natural language processing to find a sentence-levelsemantic representation. At present, the mainstream studiesof semantic role labeling focus on the use of a variety ofstatistical machine learning techniques. But it difficult toobtain high quality labeled data. To solve the problem, weproposed a novel prototype patterns selection algorithmbased on semi-supervised learning in this paper. There aretwo main innovations in this article: firstly, order parameterevolution is introduced to expand training data. Thestrongest order parameter will win by competition anddesired pattern will be selected. Secondly, the must-linksand cannot-links constraints exist in the train data is used toreduce the noise of extend data. The experiment resultsshow the proposed method has a higher performance forsemantic role labeling.
展开▼