基于多任务学习的同行评审论文接收预测

Prediction of Peer-Reviewed Paper Acceptance Based on Multi-Task Learning

摘要: 同行评审论文接收预测是一项具有重要意义的任务，其有效提升了同行评审的效率和质量。以往的同行评审论文接收预测方法大多以单任务的形式研究，并未充分利用论文评分等其它辅助信息，同时也未有效提取同行评审文本的语义特征。针对上述问题，文中提出了一种多任务同行评审文本分析模型BCLJ(BERT-CNN-LSTM-Joint Model, BCLJ)。首先，使用BERT作为词向量获得文本的矩阵表示；然后，引入卷积神经网络(Convolutional Neural Network, CNN)和长短期记忆网络(Long Short-Term Memory Network, LSTM)进行语义特征的提取，并运用注意力机制增强对文本信息的理解能力；最后，利用不同的全连接层进行多任务学习，获得论文接受预测和评分预测两种输出，通过评分预测任务来优化主分类任务。实验结果表明，多任务模型在论文接收预测任务和评分预测任务中表现出色高于其他的基线模型，在论文接收预测任务中准确率达到了0.7117，F1值达到了0.7101，在论文评分预测任务中MSE，RMSE和MAE分别为1.3690，1.1700和0.9324。

Abstract: Peer-reviewed paper acceptance prediction is a task of great significance, which effectively improves the efficiency and quality of peer review. Most of the previous peer-reviewed paper acceptance prediction methods are in the form of single-task research, which do not make full use of other auxiliary information such as paper ratings, and do not effectively extract the semantic features of peer-reviewed text. To address the above problems, a multi-task peer-review text analysis model BCLJ (BERT-CNN-LSTM-Joint Model, BCLJ) is proposed in the paper. First, BERT is used as the word vector to obtain the matrix representation of the text; then, Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) are introduced for semantic feature extraction, and the attention mechanism is applied to enhance the comprehension of the textual information; finally, multi-task learning is performed by utilizing different fully-connected layers, and two outputs of paper acceptance prediction and scoring prediction are obtained to optimize the main classification task through the rating prediction task to optimize the main classification task. The experimental results show that the multi-task model performs better than other baseline models in the paper acceptance prediction task and the rating prediction task, with an accuracy of 0.7117 and an F1 value of 0.7101 in the paper acceptance prediction task, and MSE, RMSE, and MAE of 1.3690, 1.1700 and 0.9324 in the paper rating prediction task, respectively.