Research On Implicit Text Emotion Analysis Based On BERT And Deep Neural Network

With the continuous improvement of Internet technology and the continuous advancement of the information technology revolution,various social networks,forums and e-commerce platforms(such as Weibo,Facebook,Zhihu and Tiktok)have emerged one after another.At the same time,massive text data has also been generated,including information about user sharing,communication and purchase evaluation.These text data can not only reflect the true emotional information of users,but also provide a broad research space for emotional analysis.Therefore,emotional analysis of these text data can well mine and monitor social perspectives,and has strong application value.For the government,it is possible to keep abreast of people’s livelihood at any time.For businesses,it is possible to attract more consumers by improving service quality and product quality.Emotional analysis of text data requires using text processing techniques such as natural language processing to represent text features and classify them emotionally.As a sub field of natural language processing,text emotion analysis has become one of the current research hotspots.According to the subjectivity and objectivity of the text,sentences in emotional text are divided into two categories: explicit emotional sentences and implicit emotional sentences.Currently,research on the classification of explicit affective sentences has achieved significant results.However,implicit emotional sentences are sentences that do not have any clear emotional words to state facts,and lack emotional words as clues,resulting in increased difficulty in classifying implicit emotional sentences.In particular,most existing emotional classification methods cannot well mine the deep semantic features of implicit emotional sentences and do not fully utilize contextual information.For implicit emotional texts,sentences themselves provide little emotional information,and most of the emotional information is concentrated in the context.In order to solve these problems,this paper conducts research work from two aspects: implicit affective sentences themselves and contextual information.The main contributions are as follows:(1)In order to solve the problem that most existing emotion classification models cannot well extract the deep grammatical,semantic,and word vector features of implicit emotion sentences,this paper proposes a Chinese implicit emotion analysis model BERT-CNNBi LSTM-Attention(BCBA)that integrates multiple neural networks.This model combines the BERT model,bidirectional short-term and short-term memory network,convolutional neural network,and attention mechanism.Firstly,using BERT pre training model instead of traditional word vector generation technology to effectively extract text features and capture underlying semantic and grammatical information;Then,the extracted word vector features are input to the CNN network to obtain local important information about implicit emotional sentences.The output results are obtained through the dual channel network of Bi LSTM to obtain temporal information between words and solve the problem of implicit emotional sentence length dependence.Then,an attention mechanism is introduced to calculate the emotional weight to capture important text emotional information.Finally,the final emotional category judgment is made through the softmax layer.The model was tested on the Chinese Implicit Emotion Analysis and Evaluation(SMP-ECISA)dataset,and the results showed that the model achieved better performance in Chinese implicit emotion classification tasks compared to some deep neural networks.(2)When conducting implicit affective analysis research,it was found that the emotional polarity of many target affective sentences is not obvious,and only by placing them in the corresponding context can they better reflect their emotional categories.Therefore,this paper proposes a new implicit emotion analysis model BCBA-CF,based on the BCBA model and incorporating contextual semantic features.Bi GRU and attention mechanism are used to extract the contextual semantic features of emotional target sentences,then BCBA model is used to extract the deep features of emotional target sentences,and interactive attention mechanism is used to represent these two features.Finally,softmax layer is used for final emotional category judgment.The model was tested on the SMP-ECISA public dataset,and the results showed that compared to the previous model,the performance of the model was significantly improved.