Receiver inference in relational event history data: Reconstructing communication networks with large language models

Receiver inference is central to social network analysis because it reveals interaction dynamics such as participation structures, turn-taking patterns, and directed communication ties. Recent advances in large language models (LLMs) have improved the analysis of conversational data; however, their ability to recover sender–receiver structures from multi-party conversations, evaluate uncertainty in these predictions, and assess downstream consequences for relational event modeling remains underexplored. This study addresses this gap by evaluating GPT-4.1 for receiver inference in structured multi-party conversations, prompting the model to predict intended receivers and comparing its outputs with human annotations. We examine multiple prompt designs and benchmark the proposed GPT-4.1-based approach against an open-source LLM and two heuristic baselines: a most-recently-mentioned participant rule and a last-speaker rule. Predicted and annotated sender–receiver relations are converted into relational event history sequences and used to reconstruct directed interaction networks, enabling comparison at both the turn and network levels. The results show that the GPT-4.1-based approach consistently outperforms the open-source model and heuristic baselines, and that tailored prompt design substantially improves receiver inference accuracy. We further assess whether model-based uncertainty estimates reflect prediction correctness by analyzing token-probability traces and self-assessed confidence scores. These confidence measures show moderate correspondence with accuracy, indicating partial calibration. Finally, we apply a relational event model (REM) to examine how LLM-based receiver predictions affect estimated model parameters. Overall, the findings demonstrate the potential of LLMs to enrich relational event history (REH) datasets through receiver inference.