| Text sentiment analysis is an important task in the field of natural language processing,which aims to mine opinions,emotions,attitudes and perceptions from unstructured emotional texts through collecting,processing,analyzing,summarizing and reasoning.It involves many research fields,such as artificial intelligence,machine learning,data mining,and natural language processing.According to the granularity of the given text,text sentiment analysis tasks can be divided into document level,sentence level and aspect level sentiment analysis.For the first two tasks,most of the existing methods are only limited to the analysis of the aggregate level of the document or sentence,which means it can only provide whether the document or sentence is positive,negative or neutral,and the identification of sentiment polarity is lack of considering human beings’ reading cognitive process,so this task deserves further research.Apart from the above tasks,aspect level sentiment analysis has always been a challenging task in sentiment analysis,and has always been a research hotspot in this field.Especially for the sentiment triplet extraction task proposed in recent years,the purpose of which is to extract target words and opinion words for a given text,and determine the corresponding sentiment polarity at the same time.Among them,opinion words can give the reason for judging sentiment polarity.The sentiment triplet composed of "Target terms-Opinion terms-Sentiment polarity" makes the results of sentiment analysis more complete and interpretable.Therefore,how to make a deeper semantic understanding of the text and how to effectively use external knowledge to improve the performance of ABSA task has also become one of the focuses of this paper.In addition,in recent years,with the development of Machine Reading Comprehension(MRC),the Review Reading Comprehension task involving text comment comprehension,comment opinion extraction and machine question answering has also become a challenging task in the field of text sentiment analysis,which is important for the extraction and understanding of key information of comment texts,and the RRC task has become one of the focuses of this paper.Moreover,considering the impact of news public opinions which reflect investors’ sentiments on the financial market,how to introduce news public opinions after text sentiment calculation into financial time series prediction also raises concerns.Especially for the international crude oil market,there is relatively little research on crude oil news public opinion analysis and crude oil price fluctuations,so this task is also worth exploring.Therefore,based on the above tasks and challenges,utilizing semantic understanding technology,this paper takes text sentiment analysis as the core,and focuses on four important tasks including document level and sentence level sentiment analysis in text sentiment mining,aspect level fine-grained sentiment classification and sentiment triplet extraction,review reading comprehension and the application of affective computing in financial time series prediction.This research consists of the following main contents.(1)For document level and sentence level sentiment analysis tasks,this paper proposes a multi-class fine-grained sentiment analysis method based on multi-level knowledge bases,named “Mi Mu SA”.This method involves multi-level module structures,which aims to mimic human beings’ reading cognitive and language understanding process,such as ambivalence handling process and strength handling process.Specifically,the method involves multi-level modular knowledge bases,including basic knowledge base,negation and special knowledge base,sarcasm rule and adversative knowledge base.In addition,it establishes a sentiment polarity strength knowledge base to understand the degree of sentiment polarity(such as strongly positive,slightly positive)on the basis of identifying aggregate level sentiments.The experiments are carried out on two datasets to demonstrate the effectiveness of the method.(2)For aspect-based sentiment analysis(ASBA)task,considering the significant value of ABSA,this paper selects ABSA and aspect sentiment triplet extraction(ASTE)as the research tasks.Specifically,considering that the performance of BERT is not good enough when directly applied to the ABSA task,this paper introduces a self-supervised sentence pair relationship classification task for ABSA,and proposes a method based on the pre-training language model,called "multi-level semantic relation-enhanced learning network(MSRL-Net)".In MSRL-Net,after transforming the original ABSA task into a sentence semantic matching task,the model utilizes the word dependency information,the relationship between words and sentences to enhance the word-level semantic representation for the sentence semantic matching task,while it uses the sentence semantic relationship and sentence pair relationship to enhance the sentence-level semantic representation.By making full use of the label information in the samples and constructing the two sub-tasks,the model can better understand the semantics and determine the aspect-level sentiment polarity.On the other hand,this paper proposes a novel end-to-end model for ASTE task,called“GCN-EGTS”.It is an enhanced grid tagging scheme(GTS)for ASTE task,which utilizes the syntactic component analysis tree and commonsense knowledge graph information trained based on graph convolution neural network(GCN)models.Specifically,two types of GCNs are used to model these two kinds of information:span GCN is used to capture syntactic constituency parsing information,and Relation GCN(R-GCN)is used to encode commonsense knowledge graph information.In addition,a new loss function in the model is designed,which enhances the original scheme by adding constraints to the original scheme.For the above two research tasks,the effectiveness of the proposed model is demonstrated by the experiments.(3)For the review reading comprehension(RRC)task based on machine reading comprehension(MRC)task,this paper proposes a neural network model based on BERT,called “KMA-Net”.Specifically,this paper uses multiple external knowledge,including task-aware knowledge,domain-aware knowledge and entity-relationship knowledge,to overcome the challenges of RRC tasks.In addition,this paper also uses prior knowledge to enrich the semantic representation through sentence-level information,and adds multi-granularity attention mechanism to explore different levels of interactive information,so as to enhance the information fusion between paragraphs and queries.Finally,the experiments on two datasets show the effectiveness of the proposed model.(4)For the financial time series prediction task with affective computing,due to the great impact of news public opinions on financial time series prediction,the stock price series and international crude oil price series are chosen as the research objectives,and the relevant news public opinion are quantified as affective factors into the time series prediction model through affective computing.Specifically,for stock prices prediction,this paper takes the increase and decrease of stock price prediction as a classification task,and introduces stock technical indicators and news sentiment indexes into different machine learning and deep learning models.The results show that the integration of stock technical indicators and social media sentiment index is the right direction to improve the performance of the model for stock market trend prediction.In addition,in order to explore the impact of information at different levels of stock news on stock price trends,this paper proposes a multiple text embedding construction method,which encodes multiple text features of news,including topic features,sentiment features and semantic features,and constructs a deep learning model to predict the short-term trend of stock prices by using financial data and multi-dimensional features of news text.For the international crude oil price prediction,this paper proposes a model,called "CEEMDANLSTM＿attention-ADD",based on the "decomposition and integration" framework.This model combines CEEMDAN and LSTM with the attention mechanism,based on the constructed crude oil news sentiment index.According to the characteristics of each component after the decomposition of the original sequence,the model utilizes the characteristics of the crude oil price sequence and the news sentiment sequence,then predicts the subsequence,and finally completes the prediction of the original crude oil price series.Through two application scenarios and real-world financial market experiments,this paper proves the effectiveness of text sentiment calculation based on news public opinion for financial time series prediction. |
PDF Full Text Request