Research On Application Of Semantic Similarity Calculation Method In Financial Intelligent Customer Service

Nowadays,the traditional form of customer service mainly based on manual services is constantly advancing towards the direction of intelligence and multi-channel,especially in the Internet finance industry where customer service is intensive.Intelligent customer service can realize the automation and intelligence of customer service work,reduce a lot of labor costs,and improve the user experience.Semantic similarity calculation is a key technology in intelligent customer service.The customer service system of large enterprises involves a wide range of products and services.Most of the questions asked by users appear in the form of irregular short text,which makes the user ask questions and the standard question library Semantic precise matching has become a problem.Solving the task of computing semantic similarity through artificial intelligence and natural language processing technology is one of the current research hotspots.This paper aims to study the semantic similarity calculation method applied to intelligent customer service,starting from the Chinese word segmentation of text pre-processing,part-of-speech tagging,and text word vector representation.Focusing on the neural network-based semantic similarity calculation method,in view of some problems of the existing methods,two improved semantic similarity calculation methods are proposed,which are the gated recurrent network method based on word segmentation correction and combined constituency parsing and dilated convolution network method.Specifically,in the pre-processing stage,considering that word segmentation in professional fields is prone to word segmentation errors,a word segmentation correction model for financial vocabulary is proposed.In order to solve the problem of partial semantic loss caused by the interception and zero-filling operation of sentences that are too long or too short before input to the neural network,it is proposed to use constituency parsing to design rules to supplement important semantic components in sentences.In the semantic feature learning stage,the network structure was improved.Two sets of single-layer and double-layer gated recurrent networks were designed to extract shallow and deep-level semantic features,respectively,and the two sets of difference vectors and cosine distances were combined in various ways.Stitching to highlight the difference in features between sentences.Aiming at the traditional convolution pooling structure,it is proposed to use the dilated convolution hole structure to capture the semantic association information contained in the spacers in the sentence.Finally,compare the differences in sentence features learned by the network to obtain the semantic similarity score.In the experiment part,the intelligent customer service data set provided by Ant Financial Services in the financial field and the public data set "Microsoft Research Paraphrase Corpus"for semantic similarity calculation were selected to verify the effectiveness of the proposed method.The experimental results show that the precision and F1-score of the method in the two data sets have been improved to a certain extent,and it also has good stability.And the method of this article is embedded in a simple intelligent customer service application,which can achieve a good semantic matching effect,and has certain application and practical value.