Abstract: With the development of artificial intelligence, speech interaction technology has made great progress in smart home, intelligent assistant and other fields. The semantic understanding technology of oral short text has become the research focus. There are some characteristics in spoken text, such as nonstandard syntax, less key information and small differences between sentences, which affect the accuracy of semantic recognition. In this paper, the hierarchical theory is combined with the voice interaction application in smart home field to form a multi-layer semantic analysis framework, which realizes the transformation from text data to structured knowledge. On the basis of a small amount of text data, the training time is less than 10 seconds, and the average accuracy rate is more than 96%. A distributed model training system is designed to meet the industrial application requirements.

Key words: Voice interaction, Oral semantics, A small amount of data, Hierarchical