The influence of stop words on the sentence classifier based on BERT

Abstract

Machine learning algorithms advance rapidly and show good perfor- mance. However, interpreting the results calculated by algorithms or train- ing is unclear. BERT, which exhibits excellent performance in natural lan- guage processing, is pre-trained with large-scale data and then used to im- prove performance through fine-tuning according to the user’s intention. BERT is also difficult to analyze how the output is calculated. In this paper, we try to find out which word BERT pays attention to and outputs the re- sults when classifying a given sentence. To interpret the results of BERT’s output, we analyzed the influence of masked words in pre-trained and fine- tuned BERT. Additionally, through the LIME technique, we mainly exam- ined which words have a great influence on the embedding vector results of sentences. Finally, we found that the high-frequency words in a sentence significantly affect the results, but not all the high-frequency words in the sentence impact the result.