Research On Algorithm Of Natural Language Semantic Understanding In Smart Speakers

With the advent of the smart age,the technology related to artificial intelligence has developed rapidly.Voice interaction has become one of the most important ways for human-computer interaction.Whether the Internet giants or the traditional hardware manufacturers have focused on smart speakers,and proposed their new products,trying to make smart speakers as a voice portal to provide users with content services and take over the hub of smart home.After several decades of development,the natural language has made certain achievements.But for now,intelligent voice products have proposed new challenges to semantic understanding.The accuracy of semantic understanding has become the most important standard when evaluate the technical capabilities of a voice platform and the quality of a voice product in intelligent voice products.Therefore,in a smart speaker,its most core part its semantic understanding and the most important function is music.Based on the usage scenarios of smart speakers,this article has studied the semantic understanding in the musical domain.The main tasks are as follows:Through analyzing the usage scenarios and requirements of products,I have concluded that the requirements of the musical semantic understanding algorithm are information extraction and intention recognition.Combining the related technologies and principles of natural language processing,analyzing the advantages and disadvantages of previous related studies in semantic understanding of music world,and basing on the current state of the requirements for information extraction and intention recognition,as well as messy music resources,I have proposed the idea that we can combine the knowledge base with searching to realize the semantic understanding of the musical domain in smart speakers.I have also designed a musical semantic understanding algorithm including query preprocessing,searching and sorting,field extraction and scoring.Taking use of mature grammar matching,searching and other natural language processing technologies,the algorithm has realized the natural language understanding based on the rules of usage scenarios in music world.I have proposed the idea of using sorting learning algorithm based on machine to replace the original ordering rule,this action optimize the search returned collection in searching and sorting module to provide a better collection of reference for field extraction module,and it also optimizes the accuracy of information extraction namely the attribute and its value accuracy in semantic understanding.Comparing with different sorting learning algorithms from different perspectives,I have selected the LambdaMART algorithm according to the results and intended use of the algorithms.Finally,the LambdaMART-based music sorting algorithm has improved the attribute and attribute values accuracy of semantic understanding by 1 percentage point.Based on the purpose of scoring module intention recognition and the essence of distinguishing music or non-music,I have proposed the idea of using GBDT-based scoring algorithm to replace the original rule scoring,the advantage of that is to optimize the scoring in scoring module to improve the accuracy of intention recognition in semantic understanding algorithm.The GBDT model is optimized by continuously optimizing the selected features and processing the feature values.Finally,the scoring algorithm based on GBDT model has increased the intention accuracy rate of the semantic understanding algorithm by 1 percentage point.