應用類別對照之資料融合方法於機器學習研究分類

Exploring Class Mapping as Data Fusion Technique in Machine Learning for Research Classification

訓練機器學習分類模型須充足且高品質的資料，本研究探討類別對照作為資料融合策略，發展研究分類之機器學習模型，以2008年版與2020年版之澳洲與紐西蘭標準研究分類表為研究標的，從8家機構典藏系統蒐集179,431筆已分類文件，對二版本分類表分別建立原始資料集，及以類別對照方式擴增之資料集。結果顯示49%的2008年版文件可明確對應至2020年版，反之則為63%。進一步以SVM、SciBERT、ModernBERT-base與ModernBERT-large建立分類模型，相較僅採用原始資料集，各模型經擴增資料集訓練後，分類效能均獲改進；以ModernBERT-large表現最為顯著，其大類層級提升1.0%或2.5%，中類層級增益4.4%或2.2%，小類層級改善9.9%或11.5%，未擴增之類別亦提高32.0%或15.5%。整體而言，類別對照可用於擴展訓練資料，提升自動研究分類效能。

Access to sufficient, high-quality data is essential for effectively training and validating machine learning classifiers. This study investigates class mapping as a data fusion strategy to enhance training data for research classification. Two versions of the Australian and New Zealand Standard Research Classification, ANZSRC 2008 FoR and ANZSRC 2020 FoR, are used to organize 179,431 documents from eight institutional repositories into plain and mapped datasets. Each dataset is divided into subsets corresponding to the division, group, and field levels of the classification schemes. Results show that 49% to 63% of documents are successfully mapped between schemes. Classifiers by Support Vector Machines (SVM), SciBERT, ModernBERT-base, and ModernBERT-large are trained to assess the effectiveness of this data fusion approach on classification performance. All models show improved performance at the three levels. ModernBERT-large achieved the greatest performance gains, with the improvements in validation F1 scores of 1.0% and 2.5% at the division level, 4.4% and 2.2% at the group level, and 9.9% and 11.5% at the field level. An emergent ability was observed, as performance in non-augmented classes improved with ModernBERT-large but not with ModernBERT-base. Overall, this study demonstrates that class mapping effectively enriches training datasets, enhances classification performance, and underscores the importance of model size and architecture. These findings offer a practical and scalable strategy for improving machine learning performance in research classification tasks.