Research On Mongolian Emotion Analysis By Integrating The Prior Knowledge Of Emotion Wheel And Emotion Distribution

With the progress of information technology,the network platform has become the main medium for people to express their emotions and opinions,and the text emotion analysis has also emerged.Among them,the related research on low-resource languages such as Mongolian has gradually become a research hotspot.However,due to the complexity of Mongolian word types,the high cost of data annotation and the lack of relevant research basis,the development of Mongolian sentiment analysis is difficult.Text sentiment analysis is the process of analyzing,processing and extracting the subjective text.Text emotion analysis can be regarded as the fine-grained text emotion analysis.Emotion analysis is mainly to identify the emotional states contained in the text.At present,Mongolian sentiment analysis is mostly classified into two or three categories,and the task of multi-emotion recognition has not been applied in Mongolian.In order to further study the multi-emotion recognition task,emotion distribution learning can be used to solve the problem of emotion fuzziness,so as to achieve the goal of quantitatively expressing the degree of each emotion expression.However,the lack of annotated text emotion distribution datasets has become an obstacle to emotion distribution learning research,and the existing label enhancement methods are still not comprehensive in capturing emotional information.And there is no method to realize data enhancement by combining emotion distribution learning and prior knowledge.Based on the above problems,the main research contents are as follows:(1)Mongolian multi-emotion recognition method based on improved Fastformer.In order to improve the model’s ability of processing long text,an improved Fastformer model is developed by replacing Sinusoidal position embedding in the original Fastformer model with Rotary Position Embedding(RoPE).Then the improved model is used as Mongolian emotion recognition model to realize eight kinds of Mongolian emotion recognition tasks,so as to enrich the Mongolian emotion analysis research.At the same time,more Mongolian emotional texts can be labeled with emotional labels,so as to alleviate the lack of single label Mongolian emotional data set.The experiment shows that the improved model can effectively improve the effect of emotion recognition.(2)A method of Mongolian label enhancement based on semantic rules.In order to improve the capture of emotional information by label enhancement method based on prior knowledge,a Mongolian emotion distribution label enhancement method is proposed.Firstly,on the basis of Plutchik’s wheel of emotions and emotion dictionary,degree dictionary and negative dictionary are introduced to formulate the corresponding semantic rules for the emotion word combination and calculate the weight of emotion word.Then,the single label Mongolian emotion data set label obtained from emotion recognition task is enhanced into Mongolian emotion distribution data set by using the calculation formula of normal distribution.Experiments show that this marker enhancement method is superior to other contrast methods in emotion recognition tasks.(3)Mongolian emotion distribution learning based on polar coordinates of emotional wheel.In order to promote the process of text emotion distribution learning in a gradual way,an emotion distribution learning training method integrating polar coordinates is proposed.Firstly,the polar coordinates were combined with Plutchik’s wheel of emotions to form an emotion wheel polar coordinate system,and then the emotion distribution was transformed into a complex emotion vector by vector addition.Finally,the attention structure of the emotion wheel was integrated into the model for Mongolian emotion distribution learning.The progressive cyclic loss function is composed of relative entropy and polar emotion vector loss function.Experiments show that the performance of emotion distribution learning based on polar coordinate emotion representation is better than that of traditional methods.(4)Data enhancement based on Mongolian emotion distribution learning.In order to further expand the single label emotion data set of Mongolian,a data enhancement method based on emotion distribution learning is proposed.Firstly,by taking advantage of the feature that any two emotions in the Plutchik’s wheel of emotions can be mixed into binary emotions,the emotion wheel polar coordinate system is divided more accurately,and an emotion wheel with alternating basic emotions and binary emotions is obtained.Then,the predicted complex emotion vector is obtained through the above emotion distribution learning training,and the polar angle in the vector is matched with the divided emotion wheel interval.If it corresponds to 8 basic emotions,the text is assigned with the basic emotion label;otherwise,the text is assigned with the binary emotion label other than the 8basic emotions.In this case,the purpose of data enhancement is achieved.