Textual Emotional Analysis Based On Figurative Rhetoric

In recent years,the research on the natural language processing of figurative rhetoric has attracted more and more attention,especially the processing of such as metaphor,irony,satire and other rhetoric,and the analysis of the emotional tendency of texts expressing implicit emotions,all of which are of great research value.Based on the deep learning model,this paper studies three aspects: irony recognition,metaphor recognition and implicit affective analysis.These research methods can be used not only for sentiment analysis tasks,but also for other natural language processing tasks.Therefore,this paper has important academic value.In the research on irony recognition of tweets,aiming at the problem that ironic sentences usually contain emotional inversion,this paper proposes to use pre-training CNN which contains three types of emotions,namely,positive and negative emotions,to extract emotional features,and to adopt the method of classification by combining LSTM,CNN and other deep learning models.The experimental results show that the irony recognition F value of this method is higher than the traditional CNN and LSTM methods,indicating that this combined deep learning method can effectively obtain emotional information and specific semantic information in sentences.In the problem of metaphor recognition,the metaphor recognition of Chinese verbs and English word-level metaphors are studied respectively.Aiming at the analysis of verbs and their associated noun entities in Chinese metaphor recognition,this paper proposes a bi-directional long-term and short-term memory network(BiLSTM)model based on self-attention mechanism to extract structured sentence embedding method,and uses the encoder layer based on Transformer to extract sentence features.BiLSTM stores context information effectively through hidden layer,and the attention mechanism of selfconcern can express sentence embedding through two-dimensional matrix,which effectively incorporates sentence-level semantic information.The experimental results show that the proposed method effectively improves the effectiveness of metaphor recognition.Meanwhile,the method won the third place in the Chinese Metaphor Recognition and Emotional Analysis Assessment at the 2018 Chinese Conference on Computational Linguistics..To solve the problem of insufficient use of context information in the process of training word vectors in English metaphor recognition,this paper proposes a method of pre-training two-way word representation based on fine-tuning multi-layer bidirectional transformer encoder(BERT).The model can adjust context information before and after words to generate word representation,and improve the ability to understand long sequence context relations.The experimental results show that this method is effective in metaphor detection of English verbs and other parts of speech(POS).The F value is 4% higher than that of baseline.In implicit emotional analysis,the emotional information expressed in sentences is not significant,and more contextual semantic information needs to be excavated to improve the effect.BERT obtains deep sentence semantic information by constructing deep network representation.This paper proposes using BERT to extract sentence features.The experimental results show that the predictive effect of the model is obvious for the hidden emotional polarity in sentences.Aiming at the lack of training data,this paper uses simple data enhancement technology(EDA)to expand data sets by synonym substitution,random insertion,random exchange and random deletion.The experimental results show that the combination of BERT and EDA achieves the best results in implicit affective analysis tasks,with F value reaching 0.65.