Will_go at SemEval-2020 Task 9: An Accurate Approach for Sentiment Analysis on Hindi-English Tweets Based on Bert and Pesudo Label Strategy

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1348–1353 Barcelona, Spain (Online), December 12, 2020. 1348 Will go at SemEval-2020 Task 9: An Accurate Approach for Sentiment Analysis on Hindi-English Tweets Based on Bert and pseudo Label Strategy Wei Bao1,Weilong Chen2, Wei Bai3, Yan Zhuang4,Mingyuan Cheng5, Xiangyu Ma6 1,6Southeast University 2,3,4University of Electronic Science and Technology of China 5Renmin University of China {willinseu,chenweilong921,cellurbw,delecisz}@gmail.com {mundane0827,xma2182}@gmail.com Abstract Mixing languages are widely used in social media, especially in multilingual societies like India. Detecting the emotions contained in these languages, which is of great significance to the development of society and political trends. In this paper, we propose an ensemble of pseudo-label based Bert model and TFIDF based SGDClassifier model to identify the sentiments of Hindi- English (Hi-En) code-mixed data. The ensemble model combines the strengths of rich semantic information from the Bert model and word frequency information from the probabilistic ngram model to predict the sentiment of a given code-mixed tweet. Finally, our team got an average F1 score of 0.686 on the final leaderboard, and our codalab username is will go. 1 Introduction The rapid development of modern social media makes it possible for people to express their opinions almost at any time. Detecting the sentiments of these views can roughly analyze the social and economic development of the user’s current region, the psychological characteristics, interests and hobbies of people of different ages, so as to facilitate the company to better deliver information. It is also possible to judge the emotion and psychological characteristics of a certain type of users according to their statements, so as to eliminate some adverse factors to the society. Many companies put efforts to concern customer feedback to achieve better product optimization, and then win the trust of consumers. However, in social media, speech is not required to conform to certain norms, so there are differences in language and format. In some multilingual societies, when different languages are mixed, it is even more difficult to detect the sentiments. In India and Spain, a number of bi-lingual hybrids produced many texts. The process of switching text between two or more languages is called code-mixing and a significant part is the mixing of the native language and English. The difference of language habits makes the problem even harder. A quite number of people prefer to use nonstandard words like coooool rather than cool, thx instead of thanks, which causes obstacles. Code mixing has always been one of the most important directions of natural language processing, and there is a lot of valuable research and a lot of good results in language identifying, POS tagging and Named Entity Recognition of code-mixed from a lot of researchers (Bali et al., 2014; Kumar et al., 2018). LSTM model shows great results in the sentiment analysis of code-mixing (Prabhu et al., 2016). However, the noise of the code mixed data exists. The model we propose in this paper combines multi-sample-dropout and pseudo label based on the BERT. Besides, TDIDF is also …

Content truncated to 500 words