Team Peter-Parker at SemEval-2019 Task 4: BERT-Based Method in Hyperpartisan News Detection

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 1037–1040 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 1037 Team Peter-Parker at SemEval-2019 Task 4: BERT-Based Method in Hyperpartisan News Detection Zhiyuan Ning, Yuanzhen Lin, Ruichao Zhong Department of Computer Science and Technology Beijing Normal University - Hong Kong Baptist University United International College Zhuhai, P.R.China {chineseperson5, jeremy0077jj, answer980810}@gmail.com Abstract This paper describes the team peter-parker’s participation in Hyperpartisan News Detection task (SemEval-2019 Task 4), which requires to classify whether a given news article is bias or not. We decided to use JAVA to do the arti- cle parser and the BERT model to do the bias analysis and prediction. Furthermore, we will show experiment results with analysis. 1 Introduction As the Hyperpartisan News Detection is getting more and more popular in NLP area in recen- t years, our team decided to focus on such kind of topic and choose task 4 in 2019 SemEval com- petition, which requires to decide whether a given news article is showing an unreasoning or blind al- legiance to some specific groups or persons(Kiesel et al., 2019). For the task, it also requires the com- petitor’s model to classify the news article in one of the two classes, bias or not. In previous Se- mEval competition, the classification tasks were mostly regarded as sentiment analysis on Twitter, news or scientific paper and so on. For SemEval 2019, the task focus on bias detection, which gives great help for people in daily life to acquire the news and articles in a more objective way. So far, the machine learning approaches to do the bias detection were mostly using the RNN(Iyyer et al., 2014) or the Word Vec- tors(Anil Patankar and Bose, 2017), which can get the accuracy for more than 70%. To reach a greater performance, we decided to adopt a state-of-the-art language model, BERT (Devlin et al., 2018), which set new records on many NLP tasks recently, into our political bias task analysis. For dealing with the given large dataset, JAVA was used as a parser tool to help us make those training articles more readable. 2 Model Description Our task is to predict whether an article or news is bias or not, which is entirely a binary classifi- cation task. Recently, Google released an essay about BERT and its code. They also provided the performance of BERT model on different tasks in the essay. One of them is similar to our task, which is called SST-2. It is a binary classification task, which is The Stanford Sentiment Treebank, a bina- ry single-sentence classification task consisting of sentences extracted from movie reviews with hu- man annotations of their sentiment. According to the result (about 95% accuracy), BERT perform pretty well on SST-2 task. We chose to use BERT to do the task since it had better result on binary classification. We will introduce the BERT model and its detailed imple- mentation in this section. In the following part, we …

Content truncated to 500 words