Research On Key Technologies For Intelligent Customer Service In E-commerce

In recent years,Electronic Commerce(E-commerce)based on Internet technology has gradually changed people's living habits and shopping ways.At present,e-commerce in our country is in a stage of intensive innovation and rapid development,and many companies have sprung up.It is important for e-commerce companies to pay more attention to customer services if they aspire to quickly occupy the market and achieve greater success in future.To meet an increasing number of users and products,many e-commerce companies have developed some information systems with a large number of manuals,materials,and financial costs to support customer service.However,there still inevitably arose some problems,such as limited service support of availability,low efficiency,and high cost.Therefore,artificial intelligence(AI)is increasingly used as a human assistant in ecommerce customer service,and it greatly improved the efficiency and quality of the service.However,due to the complexity of e-commerce industry and the urgent needs of enterprises desire for its accuracy and speediness,the techniques still face many problems in commercial application.This dissertation focuses on the technologies of automatic question answering,recommendation system and customer opinions mining,and the main contributions are described as follows:1.Question Answering System(QA)based on Knowledge Graph(KG)is widely used in e-commerce enterprises currently.However,in fact,some attributes of the commodity entity cannot be directly mapped into the 1-hop attribute in the KG,because they belong to the multi-hop attribute of the commodity,and the previous works did not pay more attention to this problem.To tackle the problem,this dissertation proposed a method based on rule inference of ontology which can use an engine to perform the rule inference and dynamically generate new attributes of the entities in the KG.Such kind of operation could convert the multi-hop attributes into 1-hop ones.Finally,experiments show that the proposed method not only solves the problem of multi-hop attribute,but also simplifies the time complexity of the model.In addition,in order to improve the performance of the model,this dissertation proposed an attribute selection model based on the attribute attention mechanism.On the other hand,to verify the effectiveness of the model,we evaluated the model on NLPCC-KBQA dataset and QA dataset in real e-commerce.2.Now existing algorithms of collocation recommendation extract the representing vectors from the title or image of the item and then calculate the matching similarity between the vectors.But in practice,there are some noises and less information in the titles or images.And this dissertation conducted a series of researches on this problem.First,we constructed a KG with the accessible contextual data from e-commerce platforms,such information includes detailed description of items,purchased data and category information of items.Then we improved representation learning method of the Trans H to obtain the entity embedding of items in the KG.On the other hand,it is more complex to handle the title of the item which is a set of disordered words and phrases in Chinese rather than a context-aware word sequence.Aiming at this situation,this dissertation designed a new title encoding model for feature extraction.Experiments showed that the model achieved better results.Then,we designed a collocation recommendation model to integrate the title feature and the item entity feature and calculate the collocation similarity.The model has been evaluated on public datasets and been proved that it outperformed other baselines indeed.3.The research on the review helpfulness prediction has been studied by academy and industry for many years.Previous works predicted the helpful reviews mainly based on the content,because they believed that the reviews are more helpful if they are rich in content.However,these methods ignored a fact that the items and the reviews should match each other.Recently,some studies have considered that the review must be consistent with the product,but they ignored the versatility of review.In response to this problem,this dissertation attempted to identify the helpful reviews from following two aspects:(1)The reviews must contain richer content.(2)The review must match the current product.Given this,we designed a featureaware external memory network.First,the external memory network controls the read/write operation of the memory matrix through a controller,and then it stores the typical features of items with same category into the memory matrix.Further,to better extract the feature of the reviews,this dissertation introduced a self-attention mechanism.On the other hand,when calculating the usefulness of reviews,the model used the soft attention mechanism to read the memories from the matrix to generate the final item features.Then the model obtained the helpfulness score by calculating the similarity between the item features and review features.Finally,experiments on the Amazon review dataset proved the effectiveness of the method proposed in the dissertation.4.The aspect category detection of reviews is a key task of opinion mining.Generally,a review is mainly composed of multiple sentences,and some of which are longer or shorter.And the longer sentences usually contain multiple aspects,while the shorter sentences usually contain a single aspect.Then we found that the word-level attention mechanism can achieve better results in short texts,while the sentence-level attention mechanism is more suitable for long texts.In view of this,the dissertation proposed a joint attention model,which combined the sentence-level and word-level self-attention mechanism to deal with the challenges of long and short sentences respectively.Finally,experiments were conducted on the SemEval-2014 & SemEval-2016 datasets and fully proved the effectiveness of the method proposed in the dissertation.