Abstract:

Motivation:
Massive open online course (MOOC) broke into the mainstream vision with the outbreak of the COVID-19. But the MOOC also exposed many problems, such as the monotonous form of expression and the delay in the design of interactive activities so that learners would not feel the real presence sense. With the emergence of bullet screen video, it has powerful interactive functions that can make up for the deficiencies in MOOC learning and make MOOC learning highly interactive. However, there are two problems in the direct application of Bullet screen in MOOC instructional videos. On the one hand, the speeches made by the learners have nothing to do with the teaching content and are only for entertainment. On the other hand, a large number of Bullet screen interferes with the learning effect, causing learners to only see the comments but not the teaching content clearly.

Method:
This paper proposes a BERT_DPCNN method combining Bidirectional Encoder Representations from Transformers (BERT) and Deep Pyramid Convolutional Neural Networks (DPCNN) to identify the Bullet screen that are meaningful for education. Firstly, we use crawler technology to obtain the teaching Bullet screen data on website of Bilibili, which the current most trafficked Bullet screen video website. Secondly, we invite teachers of education technology and bullet screen professionals to perform several rounds of tagging meaningful barrage in education to obtain the barrage text database for model training. Thirdly, we use the BERT model to vectorize the Bullet screen text in order to consider the impact of each word in the sentence on words in other contexts, as well as the different meanings of the same word in different contexts. Finally, we use the DPCNN model to extract local semantic features of the obtained vector matrix to improve the accuracy of screening meaningful bulletins for education.

Results:
When testing, we input the bullet screen text to be tested for the trained model. If the test result is a bullet screen that is meaningful to education, it will be displayed on the screen. If it is a bullet screen that has no meaning for education, it will be filtered directly and the next bullet screen will be screened. The process is iterative in turn. Finally, the experimental results show that the F1 score is 96.21%.

Conclusion:
This method can not only control the number of bullet screens, but also find out the screen bullet screens for teaching, create an open, shared, flexible and convenient online learning space for learners, improve the interaction between students and teachers, and let students feel the sense of the scene of teaching.

Keywords:

MOOC Educational bullet screen recognition, Deep learning, BERT_DPCNN model, Bullet screen data annotation.