Improving Large-Scale Classification in Technology Management: Making Full Use of Label Information for Professional Technical Documents

Professional technical documents (PTDs) offer a wealth of information for R&D personnel and innovation management scholars. Recently, the increase in the categories and volume of PTDs has introduced new challenges for their automatic and accurate classification. Existing studies have focused on leveraging the semantic information of documents (i.e., titles and abstracts) for classification tasks. However, the standard label hierarchy of classification systems and the rich label semantic information have been generally ignored. In this paper, we propose a supervised learning-based classification model, designed to Make Full Use of Label Information (MFULI) for hierarchical multi-label PTD classification. Firstly, we deploy a Label-aware Supervised Contrastive Learning Module (LSCLM), which introduces the definition of label set similarity with the aim of improving document representation. Then, we propose a Hierarchy-aware Label Embedding Attentive Module (HLEAM) that dynamically incorporates label hierarchy information into the classification model. We evaluate our proposed model on two public patent datasets, namely USPTO-1 and WIPO-alpha. Experimental results show that our model outperforms other state-of-the-art classification models. Furthermore, we perform a series of ablation studies and analyses to demonstrate the necessity of each component of our model. This paper provides important theoretical contributions and practical implications for innovation and technology management.