基于预训练模型的专家匹配任务分类重构OA

In research project evaluation scenarios requiring"small-circle peer review,"traditional expert retrieval/matching methods rely on semantic alignment between project descriptions and expert resumes(dual-text),facing practical challenges such as scarcity of expert resume data and semantic drift of cross-disciplinary terms.To address these issues,this paper proposes a classification-based reframing model for matching tasks using pretrained models.Based on paper text-expert label data pairs,the model trains a neural network that maps paper text to expert labels,transforming the project-expert matching task into a"text-to-label"classification problem.The model requires only the project description as input and directly outputs expert identity labels,thereby bypassing the dependency on resumes and reducing the difficulty of term alignment.To mitigate the scarcity of annotated Chinese expert data,sentiment classification is introduced as a proxy task to systematical-ly evaluate the performance differences among five models:BERT-base-Chinese,CNN,LSTM,RNN,and SVM/TF-IDF.Experiments show that BERT-base-Chinese achieves an accuracy rate above 97%and an F1_weighted-score above 97%on a self-constructed cross-disciplin-ary dataset,significantly outperforming other models.Its Whole Word Masking(WWM)technique and layered fine-tuning strategy effectively resolve term semantic drift.