Project partner extraction from research contracts

Abstract

Universities require systematic tracking of their project partners and the scope of these partnerships, with relevant information frequently contained in research contracts. These documents exhibit considerable variability in structure and combine legal language with academic context, which complicates the consistent and complete extraction of project partner information. Manual extraction of such information is slow, expensive and error-prone, while automatic approaches, although more efficient, remain limited in handling the complexity and domain-specific nuances of research contracts. Consequently, there is a mismatch between the need for accurate extraction of research project partners and the abilities of existing automatic extraction methods to effectively handle the hybrid and often complex nature of research contracts. In order to address this mismatch, we explore the capabilities of modern deep learning models for project partner extraction from research contracts. This work covers preprocessing of the documents in the form of OCR and the extraction using the open-source language models - we evaluate the performance of OCR and extraction methods and assess the influence of OCR quality on the downstream task. Specifically, we test sequence-labeling BERT-based models, generative sequence-to-sequence T5 model variations and foundational large language models on an expert-annotated corpus of 155 research contracts. Our best combination, GOTv2.0 OCR engine and T5, achieves F1-score of 0.9 on short documents and 0.29 on lengthy research contracts consisting of 40 pages on average. The T5-based combination outperforms the BERT-based baseline and achieves on-par performance with 24B Mistral LLM, surpassing it in precision. These results demonstrate the potential of leveraging language models for efficient and accurate project partner extraction, offering a foundation for further research into automated contract analysis.