Integrating Prior Scenario Knowledge for Composition Review Generation

Reviews generation is an important task for the elementary and middle school composition evaluation. Existing methods only focus on reviewing the composition contents without consideration of the effects of different grades and the types of the composition. To solve this problem, we proposed a light-weight composition review generation method. It incorporates the scenario dependent prior knowledge to reflect the diverse requirements of composition evaluation across different grades and writing styles, which includes three parts, the frequent word bank, the prior token distribution and the domain prior distribution. To incorporate scenario-dependent prior knowledge, we first encode the composition and the frequent word bank. Then we use the cross-attention mechanism to integrate the frequent word bank, and use the dynamic weight to integrate the prior token distribution into the decoding state, separately. Meanwhile, the correction module is designed to calibrate the style of the generated reviews based on the domain prior distribution. Our proposed method is compared with the SOTA works on real datasets. The experimental results demonstrate that it outperforms several strong baselines and the reviews generated by our method are better than those generated by the baselines in terms of fluency, correctness and rationality. Additionally, to verify our method’s potential applicability to other review tasks, we transfer our method to the task of generating reviews for academic papers and the results show the effectiveness of our method.