Key Technologies For Aspect Extraction Towards Question-answering Style Reviews

In recent years,with the rapid development of mobile Internet and the implementation of the “Internet +” national strategy,e-commerce ushered in a golden age of rapid development,and sentiment information extraction method for e-commerce platform user reviews has gradually become a hot research issue in the Natural Language Processing field.Existing studies mainly focus on Aspect Extraction(AE)towards the traditional non-interactive user reviews.Unlike existing studies,this paper studies the aspect extraction methods on Question-Answering(QA)style reviews.The QA style review consists of a question part and an answer part whose provider is randomly invited from the customers who have purchased the product.Thus,compared to traditional non-interactive user reviews,it is more difficult to generate a large number of faker reviews in QA style reviews.Therefore,it is worthwhile to research the key technologies of aspect-based sentiment analysis towards QA style reviews due to its important academic value in sentiment analysis and great value of application in e-commerce platforms.This paper focuses on two key sub-tasks in aspect extraction: 1)Aspect Category Classification for QA style reviews: this task aims to identify all valid aspect categories involved in a given QA style review;2)Aspect Term Extraction: this task aims to extract all aspect terms that appear in a given QA style review.In the following part,we introduce the main studies of this paper:Firstly,in this paper,we first define the aspect category classification for QA style reviews as a multi-label classification problem.For the data sparseness problem caused by the incomplete short text and question-answering semantic matching challenge in the multi-label aspect category classification task,we propose a Bidirectional Attention Neural Network(BANN)approach to perform the aspect category classification towards QA style reviews.This BANN approach consists of three steps: First,we introduce the pre-trained language model,BERT,to capture the semantic information of the context and generate context-related word representations,and at the same time we use two different strategies for word vector fusion.Second,we construct a bidirectional attention network to capture the matching information between the question and the answer,to help the model capture the aspect-relevant information.Finally,we leverage the Sigmoid function and the binary cross-entropy loss function to decode the classification result and address the multi-label aspect category classification.Experimental results demonstrate that the proposed BANN approach can significantly outperform on aspect category classification towards QA style reviews.Secondly,for the problems of data sparseness and ambiguity caused by the incomplete short text and the problems of question-answering semantic matching and invalid aspect noise caused by chaotic description,we propose a Hierarchical Graph Attention Network(H-GAN)approach to perform aspect term extraction towards QA style reviews.The H-GAN approach consists of three steps: First,the question text and the answer text are encoded respectively using the BERT model.Second,we construct a hierarchical graph attention network to learn the context information inside the question and answer individually,and the matching information between the question and answer.In this step,we filter the aspect noise information by establishing the direct connection between the characters and calculating their attention weight.Finally,we leverage the CRF model to learn the dependence and constraint relationships between the word tags.The experimental results show that the H-GAN approach is superior to other benchmark baselines for aspect term extraction towards QA style reviews.Finally,to make full use of the relevance between aspect category classification task and aspect term extraction task,we propose a Multi-Task Learning(MTL)approach to jointly learn these two tasks simultaneously which can improve the performance of these two tasks by sharing the relevant information and parameters between tasks.The MTL approach consists of three steps: First,we learn a common text representation model of the two tasks by sharing the BERT model and the weight matrix in the word vector fusion.Second,for aspect term extraction task,we construct a Hierarchical Graph Attention Network to capture the context information and question-answering matching information which is helpful in aspect term extraction,and generate a word representation with the attention information,and then use the CRF model to decode the result sequence.Finally,for the aspect category classification task,we use a Bidirectional Attention Neural Network(BANN)to model the matching information between question and answer,and then we merge the word representation learned from these two tasks to achieve a better performance of multi-label aspect category classification.Experimental results demonstrate that the MTL approach can further improve the performance of the aspect category classification task and aspect term extraction task towards QA style reviews simultaneously.