Using LLM to Reconstruct Communication Networks: Recipient Inference in Relational Event History Data

This project explores the use of large language models (LLMs)’ context understanding capabilities to address the core problem of missing message recipients in Relational Event History (REH) data—i.e., the lack of information on "who is responding to whom". By automatically inferring recipients, it converts traditionally unanalyzable dynamic interactions into analyzable communication network structures, and validates the method’s effectiveness using Dutch parliamentary debate data as a case study. The project compares with traditional methods and designs a two-layer evaluation system, providing innovative tools and methodological references for computational social science and social network analysis.

In the fields of social network analysis and computational social science, Relational Event History (REH) data records dynamic interaction sequences of "who did what to whom at what time" (e.g., parliamentary debates, online forums). However, a long-standing challenge is: only the speaker is known, but the recipient information of "who the speaker is responding to" is missing, making it impossible to construct accurate communication graphs, calculate centrality metrics, or track opinion propagation paths. Traditional methods rely on manual annotation (high cost) or rule-based heuristic inference (limited effectiveness in complex scenarios).

The core innovation lies in using LLMs’ context understanding capabilities to automatically infer recipients, assuming that the dialogue structure and semantic knowledge acquired by LLMs during pre-training can identify implicit response relationships. The technical implementation framework includes:

1. Recipient Inference Engine: Organize the speech and its context into prompts, and use few-shot learning to guide the model to predict recipients;
2. Multi-Baseline Comparison: Compare with rule-based heuristic methods, traditional machine learning models, and different LLM configurations;
3. Two-Layer Evaluation System: Turn-level (accuracy, F1 score, etc. for classification tasks) and network-level (structural similarity between inferred and real networks, degree of metric recovery);
4. Confidence Analysis: Explore the correlation between the model’s self-assessed confidence and error rate;
5. Relational Event Analysis: Study temporal dynamic patterns, topic shifts, and their association with response chains.

The experiment uses public debate records of the Dutch Parliament (Tweede Kamer) as the main object. This data has:

• Time span: Multi-year debate records;
• Participation scale: Dozens to hundreds of members of parliament;
• Topic diversity: Various types such as budget, legislation, inquiries;
• Structural features: Clear speech order and agenda framework. The data characteristics provide sufficient samples but also pose challenges: the large number of MPs leads to a large classification space, topic jumps increase the difficulty of context understanding, and political language differs from daily dialogue.

Research Significance:

1. Lower research threshold: Activate a large amount of historical REH data that does not require manual annotation;
2. Improve analysis accuracy: Accurate recipient inference enhances the validity of network analysis;
3. Expand research scope: Handle larger-scale and longer-time-span datasets. Methodological Implications: Provide references for adapting LLMs to social science tasks, designing micro and macro evaluation frameworks, and quantifying uncertainty. Potential Applications: Online community analysis, organizational communication research, historical document mining, multilingual expansion (parliamentary data in other languages).

The project code has been open-sourced on GitHub, providing:

• Complete Python implementation;
• Few-shot prompt templates (prompts/ directory);
• Evaluation scripts and metric calculation tools;
• Confidence analysis tools. It supports other researchers to reproduce results, extend methods, or apply to new datasets.

Limitations:

1. Dependence on annotated data: Training/validation requires some annotated data; fully unsupervised inference is still challenging;
2. Computational cost: LLM API call costs limit large-scale historical data processing;
3. Domain generalization: Does the method for parliamentary debate scenarios apply to informal dialogues?;
4. Multi-recipient problem: A message may have multiple recipients in reality, and the current classification framework simplifies this. Future Work: Explore more efficient prompt strategies, multi-recipient modeling, cross-language transfer, and integration with other network inference methods.

This project is an innovative application of LLMs in computational social science. By solving the recipient missing problem in REH data, it provides a key tool for network analysis. The detailed two-layer evaluation and baseline comparison validate the method’s effectiveness, offering valuable tools and methodological references for researchers in social network analysis, political text analysis, or dialogue mining.