Research proposal content extraction using natural language processing and semi-supervised clustering: A demonstration and comparative analysis

Funding institutions often solicit text-based research proposals to evaluate potential recipients. Leveraging the information contained in these documents could help institutions understand the supply of research within their domain. In this work, an end-to-end methodology for semi-supervised document clustering is introduced to partially automate classification of research proposals based on thematic areas of interest. The methodology consists of three stages: (1) manual annotation of a document sample; (2) semi-supervised clustering of documents; (3) evaluation of cluster results using quantitative metrics and qualitative ratings (coherence, relevance, distinctiveness) by experts. The methodology is described in detail to encourage replication and is demonstrated on a real-world data set. This demonstration sought to categorize proposals submitted to the US Army Telemedicine and Advanced Technology Research Center (TATRC) related to technological innovations in military medicine. A comparative analysis of method features was performed, including unsupervised vs. semi-supervised clustering, several document vectorization techniques, and several cluster result selection strategies. Outcomes suggest that pretrained Bidirectional Encoder Representations from Transformers (BERT) embeddings were better suited for the task than older text embedding techniques. When comparing expert ratings between algorithms, semi-supervised clustering produced coherence ratings similar to 25% better on average compared to standard unsupervised clustering with negligible differences in cluster distinctiveness. Last, it was shown that a cluster result selection strategy that balances internal and external validity produced ideal results. With further refinement, this methodological framework shows promise as a useful analytical tool for institutions to unlock hidden insights from untapped archives and similar administrative document repositories.