文献类型：期刊文献

英文题名：Enhanced cross-prompt trait scoring via syntactic feature fusion and contrastive learning

作者：Sun, Jingbo[1];Peng, Weiming[2,3];Song, Tianbao[4];Liu, Haitao[1];Zhu, Shuqin[5];Song, Jihua[1]

第一作者：Sun, Jingbo

通讯作者：Sun, JB[1];Song, JH[1]

机构：[1]Beijing Normal Univ, Sch Artificial Intelligence, 19 Xinjiekouwai St, Beijing 100875, Peoples R China;[2]Beijing Normal Univ, Chinese Character Res & Applicat Lab, Beijing 100875, Peoples R China;[3]Univ Penn, Linguist Data Consortium, Philadelphia, PA 19104 USA;[4]Beijing Technol & Business Univ, Sch Comp Sci & Engn, 11,33,Fucheng Rd, Beijing 100048, Peoples R China;[5]Beijing Union Univ, Teachers Coll, 5 Waiguanxie St, Beijing 100011, Peoples R China

第一机构：Beijing Normal Univ, Sch Artificial Intelligence, 19 Xinjiekouwai St, Beijing 100875, Peoples R China

通讯机构：[1]corresponding author), Beijing Normal Univ, Sch Artificial Intelligence, 19 Xinjiekouwai St, Beijing 100875, Peoples R China.

年份：2023

外文期刊名：JOURNAL OF SUPERCOMPUTING

收录：;EI(收录号：20234014828625);Scopus(收录号：2-s2.0-85172939443);WOS:【SCI-EXPANDED(收录号:WOS:001073779700003)】；

基金：The authors would like to thank the anonymous reviewers for providing helpful comments to improve the quality of the article

语种：英文

外文关键词：Automated essay scoring; Natural language processing; Contrastive learning; Data augmentation; Information fusion

摘要：Automated essay scoring aims to evaluate the quality of an essay automatically. It is one of the main educational applications in the field of natural language processing. Recently, the research scope has been extended from prompt-special scoring to cross-prompt scoring and further concentrating on scoring different traits. However, cross-prompt trait scoring requires identifying inner-relations, domain knowledge, and trait representation as well as dealing with insufficient training data for the specific traits. To address these problems, we propose a RDCTS model that employs contrastive learning and utilizes Kullback-Leibler divergence to measure the similarity of positive and negative samples, and we design a feature fusion algorithm that combines POS and syntactic features instead of using single text attribute features as input for the neural AES system. We incorporate implicit data augmentation by adding the dropout layer to the word level and sentence level of the hierarchical model to mitigate the effects of limited data. Experimental results show that our RDCTS achieves state-of-the-art performance and greater consistency.