| With the vigorous development of Internet technology,more and more people have begun to use entertainment tools such as Kuaishou,Douyin,Pinduoduo,Jingdong and other entertainment tools and third-party trading platforms to entertain their lives.These application platforms provide users with the opportunity to make a voice that has important research and application value for businesses and governments because it covers the user’s emotions and needs.Text sentiment analysis is a research hotspot in recent years,and this project aims to carry out text sentiment analysis research in the field of Mongolian.Unlike widely used languages such as Chinese and English,Mongolian is only spoken by a small number of people,so the collection of Mongolian corpus is relatively difficult and the corpus is not sufficient.The current pre-trained models can alleviate the problem of insufficient affective corpus in Mongolian to some extent,but because these models are designed for multitasking,important emotional features are ignored when processing sentiment analysis tasks,resulting in reduced sentiment classification performance.To address these problems,this project will focus on the following three aspects of research:Aiming at the problem that the data set in the Mongolian text sentiment analysis task is too small,resulting in poor model performance,this paper uses data augmentation technology to augment the Mongolian data set.Specifically,Mongolian emotional dictionaries,retirement dictionaries,and synonym dictionaries are first constructed.Secondly,EDA(Easy Data Augmentation)technology is used to perform synonym replacement,position exchange,random deletion,random insertion and other operations on the existing data set to enrich the data set.Finally,experiments show that the performance of model classification can be improved by using the enhanced dataset.Aiming at the problem that the model extracts insufficient emotional features in the sentiment analysis task,an improved BERT word embeddings combination method is proposed.On the basis of the original position embeddings,segment embeddings and token embeddings,this method adds emotional word and negative word embeddings(Emotion-Negative Embeddings),and uses the combination of these four vectors as input embeddings.Experiments show that by increasing the emotional word and negative word embeddings network,the model can learn emotional features more fully and in a targeted manner,and improve the performance of sentiment analysis.Aiming at the problem that the random masking mechanism in the sentiment analysis task leads to the loss of emotional features,an improved BERT masking mechanism model is proposed.Firstly,the model assigns higher masking probabilities to emotional words and negative words to obtain more emotional features while keeping the overall masking probability unchanged.Secondly,by using the whole word masking technique,the ability of the model to extract emotional features is further improved.Experiments show that the improved model can effectively improve the performance of classification.Experiments show that the improved research proposed in this paper has obtained performance improvement in sentiment analysis tasks and has practical value.In addition,this research can better meet the needs of the Mongolian people and help promote national unity,which is of great significance. |
