141 / 2019-06-13 09:35:08

Application of Term Library Construction Based on Machine Learning and Statistical Method in Intelligent Power Grid Custom Service

machine learning, position identification, statistical method, term component, term library

Power grid enterprise custom service center plays a fundamental role of communication between enterprises and customers, the core technology of intelligent custom service mainly includes functions based on deep learning such as intelligent speech analysis, knowledge management, intelligent custom service and service robot, which makes the construction of intelligent custom service to be the primary problem of intelligent application of power grid enterprise custom service.
Power grid intelligent custom service(Hotline:95598) mainly deals with problems arising from the supply and use of electricity by residents and enterprises. At present, custom service staffs mainly accept customer business through answering 95598 calls, WeChat or mobile phone APPs, and then manually fill in the repair report, which is proposed to the relevant business departments to deal with. There are two major problems when custom service staffs are dealing with work orders. Firstly, the business acceptance involves a wide range, which covers all areas of the grid, making the custom service staffs need to know everything about the power grid. Secondly, the geographical coverage is wide, as the custom service center centralized accepts power grid service from the whole nation, which is greatly influenced by dialect. Therefore, custom service staffs need to make use of intelligent speech analysis during business processing in order to provide users with convenient and quick services.
In this paper, a term recognition method based on machine learning and statistical information is proposed to collect unregistered term components. Taking the advantage of both “conditional random field sequence labeling word segmentation algorithm” and “word-formation rules & improved mutual information & boundary entropy”, this method indentifies terminology through the combination of correlated calculation and comprehensive evaluation. It is used to construct the intelligent power grid custom service term library, and ultimately to perform extraction experiments utilizing the power grid custom service work order, which lays the foundation for custom service artificial intelligence.
The experimental results show that the term extraction algorithm in this paper is effective and can obtain stable results for work order text in the field of power services. The main reason why the accuracy rate of “accepted content” is lower than “processed content” is that some names of reviewers are mixed in the results. The explanation for the relatively lowoverall recall rateis that the occurrence frequency of some terms is less than 3, which is a lot of terminology, according to Zipf'slaw. Considering the relatively small number (no more than 100) of stop words and filter rules we have added so far, better results can be achieved with targeted filtering and parameter adjustment.
Based on the term library, intelligent voice interaction and robot online services can be widely developed, and extremely good effect has been obtained in the practical application. However, the segmentation method proposed in this paper has some limitation to recognize new words in some public areas. It worth further research to improve the adaptability of the method in different fields.