Research On The Key Techniques Of Implicit Sentiment Analysis

As the major form of network information,texts allow researchers to obtain a large number of true and reliable sentiment domain resource,which provides necessary data resource for sentiment analysis.Sentiment analysis can be divided into explicit and implicit by whether the sentiment is expressed through explicit sentiment words.The task of implicit sentiment analysis is one of the core problems in natural language processing for its implicative expression.The sentiment of humor and offsensive speech is often hidden behind the potential semantics of the text,because they often express their sentiment by describing objective facts.A sarcastic sentence may carry positive surface sentiment,but negative real sentiment.Therefore,how to make the computer better detect humor,sarcasm,and offensive speech becomes the key problem in implicit sentiment analysis.This dissertation focuses on the key technologies of implicit text mining based on deep learning,and studies on three tasks:humor detection,sarcasm detection,and offensive speech dectection.For the task of humor detection,although phonetics and ambiguity have been introduced by previous studies,existing recognition methods still lack suitable feature design for neural networks.Phonetics structure and ambiguity associated with confusing words are needed to be learned for their own representations via the neural network.Then,Phonetics and Ambiguity Comprehension Gated Attention network is proposed to learn phonetic structures and semantic representation for humor recognition.Afterward,incongruity and ambiguity are discussed in detail and then an internal and external attention neural network is proposed for humor detection.The method integrates two types of attention mechanisms to capture the incongruity and ambiguity in humor text.The experimental results show that the proposed model not only achieves state-of-the-art performance but also has better interpretability.For the task of sarcasm detection,existing sarcasm detection approaches suffer from a lack of context and rhetoric information.To address this problem,an multiple semantic fusion is proposed for sarcasm detection,which can make full use of the context features,pard of speech features,and stylistic features.The linguistic characters of sarcastic text were usually not well represented and fused.To alleviate this problem,the latent sematic information of sarcastic text from style,sentiment,inconsistency,and context is represented.Besides,considering the contribution of various linguistic features is different,a hierarchical attention network is proposed to measure the importance of different linguistic information.Experimental results demonstrate the effectiveness of multidimensional semantic representation and hierarchical attention mechanisms.For the task of offensive detection,the most offensive detection uses dirty words dictionary,or extract sentiment features through the data set itself,which had low performance.To address the problem,a novel sentiment transfer and gate attention neural network is proposed for offensive detection.The method consists of several features extraction units,each of which is composed of the semantic and spelling understanding module.They share the parameters between the task of offensive detection and sentiment analysis,thereby transferring the sentimental features in the sentiment anlysis to offensive detection.Finally,a gated attention mechanism is used to fuse features obtained by different units.The experimental results show that the proposed model achieves state-of-the-art performance.