Transfer Learning Approach to Interdisciplinary Classification for Research Management

科研管理中的遷移學習交叉學科分類方法

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date5 Sep 2022

Abstract

Current research is interdisciplinary in nature, it creates a major challenge for government funders and research institutions to manage the interdisciplinary research effectively. AI and machine learning methods have been proposed to classify research text into disciplines, e.g., SVM method learns classification with handcrafted features; TextCNN classification method leverages efficient convolutional operation to extract features; Bi-LSTM classification method captures the sequential information of academic text; SciBERT has a huge pre-trained model using self-supervised learning. However, current research projects are managed under a single discipline; these machine learning methods can hardly be used in research management due to the scarcity of labeled interdisciplinary training data. Furthermore, current machine learning methods are not explainable and they can hardly meet the auditing requirements in research management.

A transfer-learning method is proposed to classify a research text into interdisciplinary areas in this thesis. It effectively utilizes the semantic information learned from massive single-disciplinary texts and the discipline co-occurrence information extracted from a small number of interdisciplinary texts. First, a single-discipline TextCNN classification model is trained using existing labeled single-discipline texts. The knowledge learned from the single-discipline classification is transferred to interdisciplinary classification to address the scarcity of labeled interdisciplinary data. Moreover, discipline co-occurrence information is added to the proposed model by parameter initialization. The final classifier is then obtained by fine-tuning the transferred model with interdisciplinary data. In addition, based on the proposed interdisciplinary classifier, an explainable AI method is developed to provide keyword explanations for classifying texts into specific disciplines.

The interdisciplinary classification and explainable AI methods are first proposed to effectively support interdisciplinary research management. From the theoretical perspective, the proposed methods combine transfer learning with the TextCNN model to address the interdisciplinary data scarcity problem and introduce layer-wise relevance propagation into the interdisciplinary classifier to explain classification outcomes. From the practical perspective, experiments using real-life data in the National Natural Science Foundation of China (NSFC) demonstrate that the proposed method enables efficient interdisciplinary classification with the outcome explanation. The proposed methods can be further used in interdisciplinary project recommendation, reviewer assignment, and project evaluation to promote the efficiency of interdisciplinary research management.

    Research areas

  • Research management, Interdisciplinary research, Interdisciplinary classification, Transfer learning, Explainable artificial intelligence