A Machine Learning Approach to Enterprise Matchmaking Using Multilabel Text Classification Based on Semi-structured Website Content

Abstract

Finding the right business partner to drive innovation or acquire technology transfer is a labor and time-intensive process. To simplify this process, there is a need for improved methods of automated matchmaking that can quickly identify the best potential collaboration partners. This paper presents a novel approach for semi-automated business matchmaking between companies and research institutes, that is applied to a first case study. For this purpose, we compare two transformer-based text classification models and evaluate how dataset quality affects few-shot learning performance. Flair's TARS classifier performed very well in our use case, requiring only 40 examples per class to achieve an F1 score of about 90%. This is already very close to the Hugging Face standard text classifier, which achieved an F1 score of 92% with much more annotation effort. The results show that few-shot learning models like TARS can achieve accurate results even with few training samples compared to regular transformer-based language models. Our novel approach allows the time-consuming and labor-intensive task of manual partner matchmaking to be significantly reduced.