BERT-Based Automatic Scoring Model for Speech-Oriented Text Modality

Oral proficiency reflects L2 learners' ability to use a second language in the actual language environment and can reflect the learners' language proficiency in a more realistic way. In China, English speech and English oral test are the main ways to assess English speaking ability. In this scenario, the automatic assessment system for English speech and oral English test are of value. However, the current studies lack the consideration of language usage. With the continuous development of natural language processing(NLP), the research of automatic essay scoring(AES) is increasing. The direct application of AES to the scenarios mentioned above cannot achieve the best results yet. Therefore, in this paper, we analyze the demand the demand for text modality in this field in the context of the current development of automatic assessment of spoken English and English speech. First, we implement Base Model in light of the development of AES, which is slightly better than the baseline metrics on the ASAP dataset in AES domain. Second, on the basis of the characteristics of the demand of speech-oriented text modal automatic scoring, the feature engineering approach and the neural network approach are combined on the Base Model to implement an automatic scoring model for speech-oriented text modality based on Bert. Finally, the overall experiment is analyzed and summarized.