Similarity Recognition Of Customer Service Question Based On Deep Learning

With the continuous development and popularization of mobile Internet technology and the diversification of application software,traditional customer service can`t meet the current business needs.Intelligent customer service based on language technology emerges as the times require.In the intelligent customer service question answering system,the questions posed by users have such problems as complex consulting intentions,weak context correlation,diverse questions,lack of reference,and serious colloquialism,which results in the low accuracy of the similarity calculation of the questions by intelligent customer service and the easy occurrence of non-answering situations.Traditional text similarity calculation based on word matching,ignoring lexical semantic information.Although the method based on similarity of word vectors can effectively express the semantic relationship between words,it ignores the interaction between two sentences in a specific semantic environment.In order to solve the above problems,this paper proposes to introduce some models of deep learning into the similarity recognition of intelligent customer service problems.First,a multi-interactive attention Convolution Neural Network MA-CNN is proposed.MA-CNN comprehensively considers the deep semantic information at the word level and sentence level between two sentences through two different attention mechanisms,to help intelligent customer service understand user's questions at multi-level,multi-angle and multi-granularity.This improves the similarity calculation method based on word vector,which only pays attention to the relationship between words in sentences level and ignores the semantic connection between different sentences.Secondly,on the basis of the research on the interactive attention mechanism,BMA-GRU(Bilateral Multi-Attention-GRU)model is proposed.Convolutional Neural Network can only acquire one direction text information,and can`t describe text information from both forward direction and backward direction.BMA-GRU can use bilateral GRU to represent text.Convolutional neural network has no memory,and word order will not affect convolutional neural network.GRU can add the sequence and memory of text,which makes text representation more reasonable.In MA-CNN,only using interactive attention can`t extract useful information from each sentence,while BMA-GRU can have different weights for different semantic information in each sentence,so as to better understand the text.Finally,we compare the two models with the traditional model based on word vector and other deep learning models.The results show that the proposed models are reasonable and effective.