A frustratingly easy way of extracting political networks from text.

This study demonstrates the use of GPT-4 and variants, advanced language models readily accessible to many social scientists, in extracting political networks from text. This approach showcases the novel integration of GPT-4's capabilities in entity recognition, relation extraction, entity linking, and sentiment analysis into a single cohesive process. Based on a corpus of 1009 Chilean political news articles, the study validates the graph extraction method using 'legislative agreement', i.e., the proportion of times two politicians vote the same way. It finds that sentiments identified by GPT-4 align with how frequently parliamentarians vote together in roll calls. Comprising two parts, the first involves a linear regression analysis indicating that negative relationships predicted by GPT-4 correspond with reduced legislative agreement between two parliamentarians. The second part employs node embeddings to analyze the impact of network distance, considering both with and without sentiment, on legislative agreements. This analysis reveals a notably stronger predictive power when sentiments are included. The findings underscore GPT-4's versatility in political network analysis.