Research On Short Utterance Semantic Recognition Method Based On Cascaded Conditional Random Fields

With the increasing maturity of speech recognition technology and the advances of natural language processing,it has become a trend that people use mobile devices by the means of speech interaction.This study focused on how to make a mobile intelligent terminal understand the semantic content of users' spoken operating instructions,and presented a recognition method of short utterance semantic chunks based on cascaded conditional random fields.This method divided semantic understanding of spoken language into two processes: classification of operating instructions and extraction of semantic chunk information,which both utilized the technology of conditional random fields(CRFs).By obtaining the middle semantic representation(IF)of chunks,semantic tree expressions of operating tasks can been gained.Intelligent terminal performed the appropriate operation of task according to the final semantic tree,thus achieving the purpose of semantic comprehension on spoken operation instructions.The main work of this paper is as follows:1)The corpus used in this article were collected by respondents,who are groups of students on campus.Then the processed and analyzed corpus were divided into ten kinds of instruction type,after that we adopted two different word segmentation and POS tagging tools to build the corpus,obtaining a variety of operating classification corpus and semantic chunking recognition corpus,which will be used in classification of tasks and extraction of chunk information respectively.2)According to the characteristic of research object,we designed two kinds of templates: the operating instruction classification template and the semantic chunk recognition template.Using template and CRFs to train the training corpus in operating classification corpus,we can get the operation instruction classification model(m_A).Using the CRFS to train the training corpus in semantic chunking recognition corpus,we can get other ten different operating instruction semantics chunks recognition model(m_N).3)The model m_A which has been trained was used to classify some testing corpus in the operating instructions classification corpus.The recognition result is operating instructions N.Then by the model m_N,we accomplished semantic chunk information extraction of operation instruction N.More than,we adopted middle semantic representation(IF)to indicate operating instruction N,obtaining semantic tree of operating instruction N.Eventually,machine's comprehension on operating instruction N have been realized.In my experiments,the method employed was HanLP word tagging method,whose accuracy rate of classification reaches 91.67%,and the method employed was BosonNLP,whose accuracy rate of classification reaches 94.79%.The method employed was HanLP word tagging method,whose accuracy rate of semantic annotation chunk extract reaches 88.19%,while the method employed was BosonNLP reaches 91.25%.