Abstract

This study investigates the effectiveness of two distinct computational approaches for mapping the technological landscape by extracting company and product relations from news articles. The first approach leverages Large Language Models (LLMs), specifically employing GPT-3 in a few-shot prompt setting. The second approach utilizes Pre-trained Language Models (PLM) and enhances them with Named Entity Recognition and Natural Language Inference models. To assess the performance of these methodologies, a manually annotated dataset comprising more than 200 entities and 250 relationships was used. The evaluation revealed that the PLM-based method outperformed the LLM-based method in terms of accuracy and efficiency. However, the GPT-3 based approach showed a unique strength in detecting longer-range implicit relationships between entities, highlighting its potential in comprehensive relationship mapping in the technological domain. This study contributes to the understanding of how different language model paradigms can be optimized for specific tasks in technological analysis and knowledge extraction.

Podcast

Full podcast available here: TechMining: Knowledge Models