Key Technologies For Sentiment Analysis Towards Question-Answering

In recent years,sentiment analysis towards review texts has been a hot research issue in the Natural Language Processing community due to the rapid development of mobile internet and e-commerce.Existing studies on sentiment analysis mainly focus on individual non-interactive reviews.Unlike previous studies,in this paper,we perform sentiment analysis on interactive Question-Answering(QA)style reviews.Compared to traditional non-interactive reviews,interactive QA style reviews contain fewer fake reviews because answer providers are randomly selected from the customers who have purchased the product.From this regard,QA style reviews could be beneficial to provide more accurate user information for the industries/service providers to better understand the pros and cons of the products.Therefore,it's worthwhile to perform sentiment analysis on QA style review for real-world applications.Sentiment analysis towards question-answering could be divided into three tasks according to different task granularity(one coarse-grained task and two fine-grained tasks).In summary,the three tasks are listed as follows:1)Coarse-grained task:sentiment classification towards QA,which aims to identify the sentiment polarity for the QA text pairs.2)Fine-grained task:aspect extraction towards QA,which aims to identify the aspect categories and extract the aspect terms inside the QA text pairs.3)Fine-grained task:aspect sentiment classification towards QA,which aims to predict the sentiment polarities for specific aspects from QA text pairs.In what follows,we introduce the three tasks in detail.1)Sentiment Classification towards QA(SC-QA).In this paper,we first construct a large-scale annotated corpus for SC-QA.On this basis,we further study the approach to SC-QA.Specifically,to address the semantic matching challenge between question and answer,we propose a Hierarchical Matching Attention Network(HMAN)approach to SC-QA.This HMAN approach consists of three steps.First,we segment both the question and answer texts into sentences,and then construct<Q-sentence,A-sentence>units.Second,we leverage a bidirectional matching attention network to capture the semantic matching information between Q-sentence and A-sentence.Finally,we leverage a self-matching attention network to capture the importance degrees of all<Q-sentence,A-sentence>units.Experimental results demonstrate that the proposed HMAN approach could significantly outperform several strong baselines in the task of SC-QA.2)Aspect Extraction towards QA(AE-QA).In this paper,the proposed AE-QA task consists of two relevant sub-tasks,i.e.,Aspect Category Classification towards QA(ACC-QA)and Aspect Term Extraction towards QA(ATE-QA).To incorporate the relevant information between the above two sub-tasks,we propose a multi-task learning approach to jointly learn these two sub-tasks simultaneously.Specifically,we first propose a matching attention approach to ACC-QA.This approach could effectively employ the semantic matching between question and answer so as to further improve the performance of ACC-QA.Second,we propose a gated attention approach to ATE-QA.This approach could only extract the aspect terms inside the question,which are matched with the answer.Finally,we propose a Multi-Task Learning(MTL)approach to jointly learn the two tasks simultaneously.Experimental results demonstrate that the proposed MTL approach could significantly outperform the strong baselines in both ACC-QA and ATE-QA respectively.3)Aspect Sentiment Classification towards QA(ASC-QA).In this paper,we propose a Hierarchical Reinforced Attention Network(HRAN)approach to tackle two inherent challenges in ASC-QA,i.e.,semantic matching between question and answer,and data noise.This HRAN approach consists of four steps.First,we segment both the question and answer texts into sentences,and then construct<Q-sentence,A-sentence>units.Second,we propose a sequence selector to alleviate the effects of noisy information for a specific aspect.Third,we employ a sequence selector to filter the noisy words inside both the Q-sentence and A-sentence,and propose a reinforced bidirectional attention network to capture the semantic matching information between the filtered Q-sentence and A-sentence.Finally,we employ a sequence selector to filter the noisy<Q-sentence,A-sentence>units,and propose a reinforced bidirectional attention network to capture the importance degrees of all filtered<Q-sentence,A-sentence>units.Experimental results demonstrate that the proposed HRAN approach could significantly outperform several strong baselines in the task of ASC-QA.