Multi-angle Prediction Based on Prompt Learning for Text Classification.

The assessment of Chinese essays with respect to text coherence using deep learning has been relatively understudied due to the lack of large-scale, high-quality discourse coherence evaluation data resources. Existing research predominantly focuses on characters, words, and sentences, neglecting automatic evaluation of Chinese essays based on articles’ coherence. This paper aims to research automatic evaluation of Chinese essays based on articles’ coherence by leveraging some data from LEssay, a Chinese essay coherence evaluation dataset jointly constructed by the CubeNLP laboratory of East China Normal University and Microsoft. The coherence of Chinese essays is primarily evaluated based on two big aspects: 1. The smoothness of logic (the appropriateness of using related words, and the appropriateness of logical relationship between contexts) 2. The reasonableness of sentence breaks (how well punctuation is used and how well the sentence is structured). Therefore, in this paper, we adopt prompt learning and cleverly design a multi-angle prediction prompt template that can realize the assessment of the coherence of Chinese essay from four angles. During the inference stage, the prediction of the coherence of Chinese essays is obtained through the multi-angle prediction template and voting mechanism. Notably, the proposed method demonstrates excellent results in the NLPCC2023 SharedTask7 Track1.