Research on traditional Mongolian sentiment analysis combined with prior knowledge

School of Information Engineering, Inner Mongolia University of Technology, Hohhot 010051, China

^* Corresponding author: 20201100118@imut.edu.cn

Abstract

In order to solve some problems of traditional machine learning algorithms in Mongolian sentiment analysis tasks, such as low accuracy, few sentiment corpus, and poor training effect, a Traditional Mongolian sentiment classification algorithm integrates prior knowledge is proposed. First and foremost, 1.3 million unlabeled Mongolian corpora are pre-trained and preliminarily segmented to obtain basic Mongolian vocabulary information. Secondly, the regularization method is used to segment the corpus. Next, a Mongolian text sentiment dictionary is created, which is used as prior knowledge. At the same time, an attention mechanism is integrated into the model to obtain emotional features in a dynamic form. In the last step, based on the Mongolian sentiment data set, the model is further trained and slightly modified to obtain the final traditional Mongolian sentiment analysis model. From the experimental results, the proposed model, compared with the classical sentiment classification algorithms such as Fasttext, BiLSTM, and CNN, has stronger feature extraction ability, sentiment classification performance, and faster convergence speed.