💻 computer science

A Cooperative Multi-Agent Framework for Author Name Disambiguation

This paper introduces AIDA-AND, a cooperative multi-agent framework that integrates semantic representations with interpretable metadata evidence through specialized agents to achieve competitive, modular, and efficient author name disambiguation, significantly reducing execution time while improving performance on benchmark datasets.

Original authors: Natan de Souza Rodrigues, Samuel Gomes dos Santos, Vitória Maria Diniz Pereira, Célia Ghedini Ralha

Published 2026-08-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Natan de Souza Rodrigues, Samuel Gomes dos Santos, Vitória Maria Diniz Pereira, Célia Ghedini Ralha

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast digital archives of modern science, where millions of research papers are stored and searched, a quiet but persistent problem threatens to distort our understanding of who did what. When a researcher publishes a paper, their name is the primary key used to find their work, track their career, and measure their impact. Yet names are often ambiguous. Two different scientists might share the exact same name, or a single scientist might publish under slightly different variations of their name, such as a full first name in one paper and an initial in another. Without a way to sort these out, the works of different people can get lumped together into a single, confused profile, while the work of one person can be scattered across several different profiles. This confusion skews the statistics used to evaluate scientific progress and makes it difficult to see the true landscape of collaboration. The task of untangling these names and correctly assigning each paper to its true author is known as author name disambiguation, a critical challenge for anyone trying to make sense of the world's scientific literature.

To solve this, a team of researchers from universities in Brazil has developed a new system called AIDA-AND. Instead of relying on a single, massive algorithm that tries to process all the information at once, they built a cooperative team of specialized digital agents. Imagine a group of experts in a room, each with a specific job: one looks at the spelling of names, another checks the topics of the papers, a third examines who the authors worked with, and others look at when the papers were published and where the authors were employed. In this framework, each agent acts independently, gathering evidence and forming an opinion on whether two papers were written by the same person. These opinions are then brought together by a decision agent that weighs the evidence, considering how strong and reliable each piece of information is, to make a final call. This approach allows the system to be transparent and adjustable, letting researchers see exactly which clues led to a decision and how much trust to place in each type of evidence.

The researchers tested this system on two large collections of real-world scientific data, known as AMiner12 and DBLP, which contain thousands of ambiguous author names. They compared their new multi-agent team against other existing methods, including complex systems that use graph networks and artificial intelligence models. The results showed that the cooperative approach was highly effective, particularly on the AMiner12 dataset, where it outperformed all other methods in correctly identifying pairs of papers written by the same author. The system achieved a success rate that was significantly higher than the second-best method, suggesting that combining different types of evidence through a team of specialists is a powerful strategy. On the DBLP dataset, while another method performed slightly better overall, the new system still produced competitive results and demonstrated that it could handle the complexity of real-world data without needing to build massive, intricate networks of connections.

A key finding of the study was that the system's performance depended heavily on the quality and type of information available. The researchers found that the agents responsible for checking name spellings, semantic meaning, and co-author relationships were the most influential in making correct decisions. When these specific agents were removed from the team, the system's accuracy dropped sharply, proving that these clues are essential. However, the system also showed that having access to institutional information, such as the university or organization an author belongs to, made a significant difference in datasets where that information was available. This highlights that the system is flexible; it can adapt to different types of data, though it performs best when it has a rich variety of clues to work with.

Beyond accuracy, the researchers discovered that their team-based approach was remarkably fast. By running the agents in parallel and avoiding the need to construct complex mathematical graphs or train heavy artificial intelligence models for every decision, the system completed its tasks much quicker than the competing methods. In one test, it finished the job in half the time, and in another, it was nearly eight times faster. This speed, combined with the ability to explain its reasoning, suggests that the framework offers a practical and efficient alternative for organizing the world's scientific records. The study concludes that while no single method is perfect for every situation, a cooperative team of specialized agents provides a robust, interpretable, and fast way to solve the enduring puzzle of author name disambiguation.

Technical Summary: A Cooperative Multi-Agent Framework for Author Name Disambiguation

Problem Statement
Author Name Disambiguation (AND) is a critical challenge in digital libraries and bibliographic data management. Ambiguous author names—arising from homonyms (different authors with similar names) and synonyms (the same author using name variants)—lead to incorrect publication attribution, fragmented author profiles, and distorted bibliometric indicators. While AND is often formulated as a clustering task, the decision process requires synthesizing heterogeneous and often noisy evidence sources, including nominal data, coauthorship, publication venues, institutions, and temporal information. Existing approaches range from rule-based heuristics to complex graph-based models and Large Language Model (LLM) pipelines, which may suffer from high computational costs, lack of interpretability, or dependence on repeated inference.

Methodology: The AIDA-AND Framework
The paper proposes AIDA-AND (Agent-based Intelligent Disambiguation of Authors), a cooperative multi-agent system (MAS) designed to modularize the AND process. The framework decomposes the disambiguation task into specialist agents that evaluate specific evidence sources, coordinated by an orchestrator and a decision agent.

Architecture and Agents: The system utilizes seven specialist agents:
- Nominal Agent: Evaluates name strings, surnames, initials, and variants using token overlap, string similarity, and Jaccard similarity.
- Semantic Agent: Computes semantic similarity using cached embeddings from the IBM Granite model (granite-embedding-311m-multilingual-r2) applied to metadata (titles, abstracts, keywords, etc.).
- Social Agent: Analyzes coauthor overlap, awarding bonuses for repeated shared coauthors.
- Temporal Agent: Assesses publication year compatibility, penalizing large temporal gaps.
- Institutional Agent: Evaluates affiliations, organizations, and email domains.
- Decision Agent: Aggregates opinions from the specialist agents.
- Orchestrator: Coordinates communication among agents using the SPADE/XMPP protocol and updates predicted clusters.
Evidence Scoring and Decision Strategy:
Each specialist agent produces a score and a confidence value based on heuristic, configurable rules (not learned from labeled data). The Decision Agent calculates a weighted aggregated score ( $S$ ) and global confidence ( $C$ ) using the formula:
$S = \frac{\sum w_g c_g s_g}{\sum w_g c_g}$
where $w_g$ , $c_g$ , and $s_g$ represent the weight, confidence, and score of agent $g$ , respectively. The final decision (merge, non-merge, or uncertain) is determined by comparing the aggregated score against a configurable merge threshold and applying conflict rules (e.g., handling weak nominal compatibility or institutional conflicts).
Execution: The framework processes ambiguous author-name blocks in parallel groups. It avoids graph construction, Graph Convolutional Networks (GCN), and Graph-Enhanced Hierarchical Agglomerative Clustering (GHAC), relying instead on lightweight heuristic signals combined with semantic embeddings.

Key Contributions

Cooperative Multi-Agent Architecture: A modular system separating evidence evaluation into specialist agents for nominal, semantic, social, temporal, and institutional data.
Configurable Decision Strategy: A mechanism to combine weighted evidence, merge thresholds, and conflict rules without requiring end-to-end learning from labeled data.
SPADE/XMPP Implementation: A practical implementation of agent communication and coordination using the SPADE platform and XMPP messaging.
Comprehensive Evaluation: An experimental assessment on AMiner12 and DBLP datasets, including comparisons with baselines (ATGEP, MD Similarity, Higher-order SA, and ComMAND), threshold sensitivity analysis, ablation studies, and runtime evaluation.

Experimental Results

AMiner12 Performance: AIDA-AND achieved the highest pairwise F1 (pF1) score (0.834) among all compared methods, outperforming the second-best (MD Similarity, 0.784) by a relative improvement of 6.4%. It also achieved the highest pairwise recall (0.875). However, ComMAND outperformed AIDA-AND in cluster-level metrics (ACP, AAP, and K).
DBLP Performance: On the DBLP dataset, ComMAND achieved the highest values across all metrics. AIDA-AND achieved a pF1 of 0.605 and a K value of 0.805. The lower pairwise precision (0.536) indicated a higher rate of false positive merges compared to baselines, likely due to the absence of affiliation data in the DBLP version used, which limited the Institutional Agent's contribution.
Ablation Study: Removing the Nominal, Semantic, or Social agents resulted in the most significant drops in pF1, confirming these as the most influential components. Temporal and Institutional agents provided complementary but less dominant signals.
Runtime Efficiency: Under the adopted parallel execution settings (5 groups for AMiner12, 20 for DBLP), AIDA-AND reduced execution time by 50.1% on AMiner12 and 87.4% on DBLP compared to ComMAND. This efficiency is attributed to the avoidance of graph construction and GCN training, combined with parallel processing.

Significance and Claims
The authors position AIDA-AND as a competitive, modular, and interpretable alternative for combining semantic and heuristic evidence in AND. The framework addresses the need for transparency in decision-making, contrasting with "black-box" graph-based or LLM-heavy approaches. The results suggest that while graph-based methods (like ComMAND) may yield superior cluster consistency in certain contexts, the cooperative multi-agent approach offers a viable path for recovering same-author pairs, particularly when multiple metadata sources are available. The authors emphasize that the framework's performance is sensitive to dataset characteristics (e.g., availability of affiliation data) and that the merge threshold must be tuned per dataset. The work concludes that AIDA-AND successfully demonstrates that a cooperative, agent-based decision process can effectively balance heterogeneous evidence without relying on complex graph learning or repeated LLM inference.

Technical Summary: A Cooperative Multi-Agent Framework for Author Name Disambiguation

More like this