Research on automatic semantic classification of human-interaction instructions

Abstract

Instructions extraction extracts structured information from unstructured natural language instruction text, is an application of information extraction in the field of human-computer interaction. For a natural language instruction text, if we want to extract structural information which can able to describe the text semantic completely, it is critical to position these words or phrases and mark one description which belongs to their own semantic description. This paper first try to a solution which is semantic classification based on dictionary. Because of some shortcomings of the dictionary itself, the semantic classification results are poor. Through the analysis of dictionary-based semantic classification results, this paper proposes a semantic classification method which combining CRF, self-training and Dictionary. Use this method to conduct experiments in the field of vehicle. The experiment results show that our method can be effective in semantic classification for the natural language instruction text; the overall correct rate is 92%. Semantic classification is prepared for the following work of structured information extraction.