💻 computer science

Author Name Disambiguation in Portuguese Scientific Publications: The SBC-AND dataset

This paper introduces SBC-AND, a curated dataset of 442 Portuguese scientific publication records designed to address the underrepresentation of non-English languages in Author Name Disambiguation research and serve as a challenging benchmark for evaluating multilingual disambiguation models.

Original authors: Natan de S. Rodrigues, Samuel Gomes dos Santos, Vitória Maria Diniz, Sirlene A. Rodrigues Costa

Published 2026-07-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Natan de S. Rodrigues, Samuel Gomes dos Santos, Vitória Maria Diniz, Sirlene A. Rodrigues Costa

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

The Big Problem: The "Name Tag" Mix-Up

Imagine you walk into a massive library where millions of people have written books. The problem is that many people have the same name, like "John Smith" or "Maria Silva."

In the world of science, this is a huge headache. If a researcher named "Maria Silva" writes a paper about computers, and another "Maria Silva" writes a paper about history, the library might accidentally think they are the same person. This messes up their reputation, their citation counts, and the way we understand who is collaborating with whom. This task of figuring out which "Maria Silva" wrote which paper is called Author Name Disambiguation (AND).

The Gap: The "English-Only" Party

For a long time, scientists have built tools to solve this mix-up, but they mostly practiced on English names. It's like having a party planner who is great at organizing English-speaking guests but has never met anyone with a Portuguese name.

Portuguese names are tricky. They often have multiple last names (like "Silva" or "Santos") and little connecting words like "de," "da," or "dos." Plus, in scientific records, these names are often shortened or written inconsistently (e.g., "A. Silva" vs. "Ana Silva"). Because of this, the existing tools struggle when they encounter Portuguese publications.

The Solution: The SBC-AND "Training Gym"

The authors of this paper created a new tool called SBC-AND. Think of this as a specialized training gym for computer programs.

What's inside? It's a collection of 442 real scientific articles from Brazilian computer science journals.
How is it organized? The authors took these articles and sorted them into "ambiguous blocks." Imagine a box labeled "Silva." Inside, there are papers from 232 different real people who all share that last name.
The Goal: The dataset is "curated," meaning a human has already checked and labeled exactly which papers belong to which real person. This gives the computer a "correct answer key" to test itself against.

The Experiment: Testing the Computer's Brain

The researchers didn't just make the dataset; they put it to the test. They used a flexible computer system (called ADAN) that tries to sort the papers back into the correct groups.

They tested the system using two different "brain" models (AI models that understand language):

BAAI/bge-m3: A multilingual model.
IBM Granite: Another multilingual model.

They also tweaked the system's "depth" (how many layers of thinking it does) and how long it practiced (training epochs).

The Results: A Tough Challenge

Here is what happened when the computer tried to sort the papers:

The "Big Picture" vs. The "Details": The computer was actually quite good at keeping the groups separate from each other (like keeping the "Silva" box away from the "Oliveira" box). However, it struggled to sort the papers inside the "Silva" box correctly.
The Score: The computer got a decent score on the overall structure (around 90% accuracy), but when it came to matching specific pairs of papers to the right author, the score dropped significantly (around 45%).
Why? The paper explains that this is because many of the real authors in the dataset only have one or two papers listed. It's like trying to identify a person at a party when you only see them once; it's very hard to tell them apart from someone else who looks similar.
The Winner: The IBM Granite model performed slightly better than the other, especially when it used a specific amount of "thinking depth" (two layers). Making the system "deeper" (three layers) didn't help; it actually made things noisier.

The Takeaway

The paper concludes that SBC-AND is a very tough test for current technology. It highlights that while computers are getting better at organizing scientific data, they still struggle with languages like Portuguese where names are complex and authors often don't have many publications to help identify them.

The authors have made this dataset available online so other researchers can use it to build better tools, ensuring that in the future, every "Maria Silva" gets credit for their own work.

Technical Summary: Author Name Disambiguation in Portuguese Scientific Publications: The SBC-AND Dataset

Problem Statement
Author Name Disambiguation (AND) is critical for maintaining the accuracy of bibliographic data, influencing citation counts, collaboration networks, and research evaluation metrics. While significant advances have been made in AND over the last two decades, particularly using text embeddings and graph models, the majority of public datasets and benchmarks focus on English-language publications. This leaves Portuguese-language scientific records underrepresented. Portuguese names present specific challenges, including multiple surnames, particles (e.g., "de," "da," "dos"), and inconsistent metadata formatting (e.g., missing accents, truncations, lack of ORCID identifiers). Existing Portuguese datasets, such as BDBComp, lack the standardized structure required for integration with modern, configurable AND frameworks, hindering reproducibility and comparative analysis.

Methodology
To address these gaps, the authors developed SBC-AND, a curated and standardized dataset specifically for AND in Portuguese scientific publications.

Data Curation: The dataset aggregates 442 publication records from open-access journals maintained by the Brazilian Computer Society (SBC), including the Journal of the Brazilian Computer Society and others hosted on the SOL platform. The collection focuses on Brazilian and Portuguese researchers to capture Lusophone linguistic patterns.
Structure: The data is organized into 21 ambiguous author blocks (grouped by abbreviated signatures, e.g., "A. Silva") containing a total of 232 real author clusters. Each block is stored as an individual JSON file containing metadata such as title, authors, coauthors, abstract, keywords, venue, and publication year.
Curation Process: A Python script collected initial metadata, followed by manual curation to verify real authors and resolve ambiguities using evidence from titles, affiliations, coauthors, and research areas.
Experimental Framework: The dataset was evaluated using ADAN, a configurable AND framework. The experiments utilized:
- Embedding Models: Two multilingual models were tested: BAAI/bge-m3 and IBM Granite.
- Graph Architecture: Heterogeneous graphs were constructed where publication records are nodes and bibliographic evidence (coauthorship, venues) are relations. Graph Convolutional Networks (GCN) were applied to learn graph-aware representations.
- Configurations: The authors tested 18 configurations by varying the number of GCN layers (1, 2, or 3) and training epochs (1000, 2000, or 3000).
Evaluation Metrics: Performance was assessed using Pairwise Precision ($pP$), Pairwise Recall ($pR$), Pairwise F1 ($pF1$), Average Cluster Purity (ACP), Average Author Purity (AAP), and the K-metric ( $K$ ).

Key Contributions
The paper makes three primary contributions:

SBC-AND Dataset: The introduction of a new, standardized dataset for Portuguese AND research, featuring complete metadata and a JSON structure compatible with modern frameworks.
Configurable Evaluation: The application of a flexible, configurable framework (ADAN) to evaluate the dataset under various experimental settings, demonstrating how different embedding models and graph depths perform on Portuguese data.
Characterization of Challenges: A quantitative and qualitative analysis identifying specific difficulties in Portuguese AND, particularly regarding frequent surnames and low record recurrence per author.

Results
The evaluation of SBC-AND using the ADAN framework yielded the following findings:

Benchmark Difficulty: The dataset proves to be a demanding benchmark. While global clustering metrics (ACP, AAP, and $K$ ) remained high (with $K$ reaching 0.9168), Pairwise F1 scores were significantly lower (best $pF1$ of 0.4600). This indicates that while the system maintains a coherent global cluster structure, it struggles to recover all individual record associations correctly.
Model Performance: The IBM Granite embedding model outperformed BAAI/bge-m3 across all configurations, achieving the best $pF1$ score (0.4600) with two GCN layers and 1000 epochs.
Graph Depth: Increasing the GCN layers beyond two did not improve performance; in some cases, deeper propagation (3 layers) appeared to introduce noise or excessive smoothing.
Block-Level Variance: Performance varied significantly across ambiguous blocks. Blocks with common surnames (e.g., "silva," "santos") and low record recurrence (authors with only one or few records) resulted in $pF1$ scores near 0.00, whereas blocks with higher author recurrence showed better performance.

Significance
The authors position SBC-AND as a necessary resource to support reproducibility and future research in multilingual and low-resource bibliographic data. By providing a standardized dataset for Portuguese publications, the work enables the testing of current AND methods in a linguistic context previously dominated by English data. The results highlight that while global clustering structures can be preserved, the specific challenge of distinguishing authors with common surnames and sparse publication records remains a significant hurdle for current state-of-the-art methods. The dataset is made available in an online repository to facilitate further study and the development of more robust disambiguation tools for the Lusophone scientific community.