Research On Key Technologies Of Sentiment Analysis For Short Text

Social networks and e-commerce platforms have become huge public information centers,and the mining and value empowerment of massive information data have been going on all the time.The proposal of the fourth paradigm of data science provides a theoretical basis for deep learning to play a great role in the fields of big data and Artificial Intelligence(AI).Natural Language Processing(NLP)also benefits greatly from it and develops rapidly.Using massive Internet data to analyze people's feelings and opinions has important scientific research value and social value.Sentiment analysis in NLP is one of the most active research areas,and has expanded from computer science to management and sociology,such as marketing,finance,politics,history and even health care.Opinion is almost the core of all human activities and the key factor affecting people's behavior.How to use NLP technology to conduct sentiment analysis of subjective opinion text has been paid more and more attention by researchers.In NLP,different from traditional semantic analysis,sentiment analysis focuses more on sentiment semantics related to viewpoints,including the issue of polarity classification of text sentiment,etc.,which requires deeper understanding and modeling.In addition,texts and product reviews on social networks are short,highly targeted and often contain richer sentiment information.Compared with traditional machine learning algorithms,deep learning does not rely on artificial construction of features,but has the self-learning ability of features,which is very suitable for the abstract,high-dimensional and complex features of language text.In this dissertation,effective deep learning solutions are designed according to different task stages,scenarios and granularity for short-text sentiment analysis.The main work and achievements of this dissertation are as follows:1.A hybrid word embedding based interactive attention network(HWE-IAN)is proposed to solve the problem of insufficient semantic expression of text sentiment and the semantic gap between different domains.The BERT(Bidirectional Encoder Representations from Transformers)algorithm is used to supplement the ability of traditional word embedding to represent sentiment semantics,and to fully explore the sentiment semantic information contained in the text during task preprocessing.The semantic features such as part-of-speech,position and n-gram are integrated into the model,which makes the model have more abundant sentiment semantic expression.The model also uses the attention mechanism to make various features interact with each other and abstract the deeper semantic associations within the context,so as to improve the performance in sentiment classification.Finally,the experimental results on two public English sentiment corpora show that the HWE-IAN model is better than other comparative models and can effectively improve the performance of sentiment classification.2.A memory network based on multi-head attention(MAMN)is proposed,which solves the performance bottleneck of multi-head attention mechanism and memory network,so as to further fully explore the sentiment semantic features and contextual inline structural relations contained in short texts.MAMN model uses n-gram features and ON-LSTM network to improve the multi-head self-attention mechanism,so as to extract the inline relationship of text at a deeper level and obtain more abundant text feature information.In addition,the multi-head attention mechanism is used to optimize the structure of multi-hop memory network,so as to effectively model the internal semantic structure of short text context and fully explore the high-level sentiment semantic features.Finally,the model is evaluated on three public English user review datasets.The results verify the effectiveness of MAMN in sentiment classification,and its performance is better than other baseline task models of CNN,LSTM and capsule architecture.3.A transfer capsule network with multi-hop attention(MHA-TCAP)is proposed to solve the scarcity of labeled training data in specific fields.In the domain-oriented fine-grained sentiment classification task,the MHA-TCAP model explores the method of performance improvement on small datasets.The model uses deep memory network and capsule network to construct the transfer learning framework.With the help of the transfer learning characteristics of capsule network,the knowledge contained in the large-scale annotated data in similar fields is transferred to the target field effectively,so as to improve the classification performance on small datasets.MHA-TCAP uses multi-dimensional combined features to make up for the deficiency of one-dimensional feature attention mechanism,while multiple attention computing layers based on domain information can be superposed to obtain deeper domain-specific sentiment feature information.The experiment is conducted on a public Chinese review dataset(including six domains).The results not only show that this model has good classification performance,but also verify its transfer learning ability.Finally,it is proved that MHA-TCAP also has good generalization ability for more fine-grained aspect-level sentiment classification.4.An attention-based aspect-level sentiment capsule network(ABASCap)is proposed to solve the problem of aspect-level fine-grained sentiment classification.By studying the internal association between the aspect and the context,a more reasonable modeling method can be used to mine the sentiment semantic features related to the aspect more effectively.An improved multi-head attention mechanism is used to process n-gram features to capture the internal structure of text and the semantic association between aspect and context.The local context window(LCW)is defined to delimit the local context area related to the aspect,and the local context mask mechanism(LCM)is proposed,which combines multi-head attention mechanism to model the strong association between aspect and its local context.The capsule network is used to generate the final text representation,and the routing algorithm and activation function are optimized according to the task.Finally,the model is evaluated on three fine-grained sentiment classification datasets.The experimental results show that ABAScap is better than other baseline models in the task,and the performance improvement is more significant after combining with BERT,which fully proves its effectiveness in aspect-level sentiment classification.