基于要点匹配的文科主观题通用评分

Key Points Matching Based Scoring Method for Liberal Arts Subjective Questions

主观题自动评分是智慧教育创新中的重要环节,逐步成为人工智能与教育行业领域交叉的热门方向之一。该文面向文科要点主观题,提出基于多任务学习的要点匹配评价模型:评估学生作答与标准答案各个要点之间的匹配等级,并抽取其中与要点相对应的具体片段,通过这两个任务的结果同时刻画学生对每个要点的掌握程度,并作为自动评分的关键特征;将要点匹配评价结果与文本相似度特征相结合,实现主观题作答自动评分,在无定标数据的通用评分场景下大幅提升了效果。对比实验证明了相比传统特征,基于要点匹配评价结果的特征在评分模型中更加重要。

Automatic scoring of subjective questions has become an important aspect of smart education innovation,gradually becoming one of the hot topics in the intersection of artificial intelligence and the education industry.This paper proposes a key point matching evaluation model based on multi-task learning for liberal arts subjective questions of key points:the model evaluates the matching level between the student's answer and each key point of the standard answer,and extract the specific fragments corresponding to the key points.Through the results of these two tasks,depict the student's mastery of each key point as a key feature of automatic scoring.The key point matching evaluation results are combined with text similarity features to achieve automated scoring of subjective questions,significantly improving the effectiveness in general scoring scenarios without calibration data.Comparative experiments have shown that compared with traditional features,features based on key point matching evaluation results are more important in the scoring model.