Domain-specific natural language understanding technology is one of the core technology of vertical search engines

domain-specific question answering system and other applications. This research focus on a novel constrained semantic grammar and its automatic learning methods based on an existing domain-specific question answering system. An error-driven learning method of semantic grammar is proposed. The method first partially parses the ungrammatical sentence based on the core semantic grammar

then it attempts to build a complete parse tree

including predicting the top-level node of the partial parsing tree

generating and verifying hypotheses of new grammar rules. Learnability metrics is used to filter sentences in the training corpus to improve the overall quality and efficiency of grammar extending algorithm. The proposed algorithm is applied to two domains of different scales. In the interactive learning paradigm

learning efficiency are compared in different domains. In the batch learning paradigm

the paper compares the accuracy

MRR and recognition rate of the extended grammar and core grammar on twodatasets. The test results show that the proposed method is effective.