Peer Review Research Based On Deep Text Mining

With the development of science and technology,the volume of academic papers submitted has increased significantly,and the review efficiency and quality of papers has become a great challenge in the evaluation process of academic papers nowadays.The establishment of a scientific evaluation system and an intelligent evaluation method for academic papers has become a hot topic in the field of academic paper evaluation nowadays.Due to the monolithic form of data,previous research on the evaluation of papers could only be done from a quantitative perspective,and it was impossible to explore the rich subjective information of individual reviewers on the papers they review.In this thesis,we explore the feasibility of extracting fine-grained paper review information and building an intelligent review based on the data of Open Review peer review platform and mining expert review text data from the perspective of natural language processing.Based on this,the following work is carried out in this thesis:First,the peer review text is interspersed with opinion sentences with emotional features,which contain rich subjective information of experts.In order to accurately extract these sentiment character sentences,this thesis proposes a paper review opinion sentence recognition model.The model first establishes the initial label data through the normalized fields of Open Review platform,and trains the opinion sentence recognition model based on the text classification task on this basis;then uses the label correction model to segment the interleaved distributed opinion sentences and non-opinion sentences to realize the correction of misclassified labels under the initial label system.The experimental results show that the opinion sentence recognition model has excellent performance in terms of recognition accuracy and quality through the two processes of initial label model construction and post-label correction in the scenario without manual annotation.Secondly,in order to build an intelligent paper acceptance prediction model,this thesis proposes a hierarchical attention BERT model HAB that incorporates the information of opinion sentences in paper review;the input layer of this model incorporates the opinion sentence recognition model based on the whole review text of experts to filter the influence of non-opinion sentence noise on paper review prediction;the feature extraction layer mines the key information affecting paper The feature extraction layer mines the key information affecting the paper acceptance decision in the long text of the review through the hierarchical attention mechanism;finally,the prediction of the thesis acceptance by combining the word attention layer and the sentence attention layer is achieved.The experimental results show that both the opinion sentences and the hierarchical attention model have a positive impact on the paper acceptance prediction task by incorporating the subjective sentiment information of experts,with Acc and F1 reaching 70.73% and 69.97%,respectively.Finally,to address the situation that single sentences in review texts contain multi-attribute sentiment distributions,this thesis introduces a aspect-based sentiment analysis model into peer review text research and proposes a multi-task learning-based fine-grained sentiment analysis model BLBC for peer review of papers.The model is based on a multi-task learning framework and adds the Bi LSTM-CRF module to the BERT-LCF model,so that it can The model is based on a multi-task learning framework,and the Bi LSTM-CRF module is added on top of the BERT-LCF model,so that it has the ability to accomplish both attribute word extraction and fine-grained sentiment analysis.The experimental results show that the proposed fine-grained sentiment analysis model for peer review can complete the tasks of attribute word extraction and sentiment prediction for expert review end-to-end,and the introduced Bi LSTM-CRF module can effectively improve the model’s ability to identify attribute words.Among them,the F1 of attribute extraction task and sentiment analysis task reached 89.01% and 90.71%,respectively.