💬 NLP

Cross-Source Reasoning-based Correction for Author Name Disambiguation

This paper introduces CrossND, a full-stack framework that leverages cross-source reasoning to automatically correct author name disambiguation errors by identifying and resolving inconsistent paper-author assignments across different sources without requiring expert annotation.

Original authors: Fanjin Zhang, Yunhe Pang, Bo Chen, Zhiyu Shen, Yanghui Rao, Evgeny Kharlamov, Jie Tang

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fanjin Zhang, Yunhe Pang, Bo Chen, Zhiyu Shen, Yanghui Rao, Evgeny Kharlamov, Jie Tang

Original paper licensed under CC BY 4.0 (http://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 Twin" Mix-up

Imagine you are looking for a specific author named "Quanquan Gu" in a massive digital library. The problem is that there are two different people with that exact name.

Person A is a computer scientist at UCLA who writes about algorithms.
Person B is a doctor at Zhejiang University who writes about Parkinson's disease.

In many academic databases, the computer system gets confused. It might accidentally give Person A's computer science papers to Person B, or vice versa. This is called Author Name Disambiguation. When these mistakes pile up, it messes up author rankings, award decisions, and research records.

The Old Way: Guessing Alone

Previously, computers tried to fix these mistakes by looking at one single library (like AMiner) and trying to figure out which papers belong to which person. They would look at the paper titles and keywords.

The Flaw: If the library itself has a mistake, the computer keeps making the same mistake over and over. It's like trying to find a typo in a book by only reading that one book; if the typo is there, you might think it's correct because you have no other reference.

The New Solution: The "Cross-Check" Detective (CrossND)

This paper introduces a new system called CrossND. Instead of looking at just one library, it acts like a detective who checks two different libraries (e.g., AMiner and MAG) against each other to find the truth.

Here is how CrossND works, broken down into three simple steps:

1. The "Clean-Up Crew" (Chain-of-Refinement)

Before the detective starts solving the case, they need to clean up their evidence.

The Analogy: Imagine a messy desk full of papers. Some are the right ones, but some are junk or belong to someone else.
What CrossND does: It uses a very smart AI (a Large Language Model) to look at an author's list of papers and say, "This paper definitely belongs here," and "This one looks weird, let's throw it out for now." This creates a clean, reliable "core" list of papers for that author.

2. The "Cross-Examination" (Cross-Source Reasoning)

Now the detective compares the two libraries.

The Analogy: Imagine you have two witnesses (Library A and Library B) telling you about a crime.
- If both witnesses say, "The suspect was wearing a red hat," you are very confident that is true.
- If Witness A says "Red hat" but Witness B says "Blue hat," you know something is wrong.
What CrossND does: It looks at the papers in Library A and compares them to Library B.
- If the authors in both libraries look very similar (same research topics), the system trusts the match.
- If they look totally different (one writes about AI, the other about medicine), the system flags the paper as a likely mistake.
- The Magic Trick: It uses a special logic tool (called Probabilistic Soft Logic) that helps the AI reason through the confusion. It doesn't just say "Yes" or "No"; it weighs the evidence like a judge, asking, "If Author A matches Author B, and Paper X matches Author B, does Paper X really match Author A?"

3. The "Second Opinion" (Test-Time Scaling)

Sometimes, even smart detectives get tired or make a quick, wrong guess.

The Analogy: If you are unsure about a math problem, you might solve it three times in a row to see if you get the same answer.
What CrossND does: It runs the check multiple times, rearranging the order of the papers it looks at. By doing this, it "bootstraps" its own confidence. If it keeps getting the same result, it becomes very sure. If the results wobble, it knows to be more careful.

Why Is This Better?

The authors tested this system on real-world data (thousands of papers and authors).

The Result: CrossND beat 17 other existing methods. It found more errors and fixed them without needing humans to manually check every single paper.
The Cost: It is surprisingly cheap to run. The system costs less than $0.002 per paper to check, which is very low for such a complex task.
Real-World Use: The system is already being used in a prototype tool that helps researchers check their profiles across different databases (AMiner, MAG, and Google Scholar).

Summary

Think of CrossND as a super-smart librarian who doesn't just rely on one shelf of books. Instead, they walk over to a second library, compare the lists, and use logic to figure out which books were put on the wrong shelves. By cross-checking sources and cleaning up the data first, they can fix the library's catalog much faster and more accurately than anyone else.

Technical Summary: Cross-Source Reasoning-based Correction for Author Name Disambiguation (CrossND)

1. Problem Definition

Author Name Disambiguation (AND) is a fundamental challenge in academic search systems, aiming to distinguish between different individuals sharing the same name. While existing approaches primarily focus on "From-Scratch" disambiguation (clustering papers by name) or real-time correction within a single data source, they suffer from two critical limitations:

Cumulative Errors: Imperfect algorithms lead to inevitable accumulation of errors in paper-author assignments.
Source Inconsistency: Different academic databases (e.g., AMiner, MAG, Google Scholar) often contain inconsistent paper-author assignments. Current methods typically ignore these cross-source discrepancies or rely on resource-intensive expert annotation to rectify them.

The paper proposes a new paradigm: Cross-Source Correction. Instead of treating external data as ground truth, the goal is to leverage inconsistencies between an internal source ( $G_{in}$ ) and an external source ( $G_{out}$ ) to identify and correct erroneous paper-author assignments without human intervention.

2. Methodology: The CrossND Framework

The authors propose CrossND, a full-stack framework designed to handle noisy labels and leverage cross-source reasoning. The framework consists of three sequential stages:

Stage 1: Chain-of-Refinement Pipeline

To mitigate label noise before training, the authors employ a three-step pipeline using Large Language Models (LLMs):

Profile Cleaning: An LLM analyzes an author's original profile (a list of papers) to identify and remove anomalous papers that do not align with the author's research identity, constructing a "coreset" of high-confidence papers.
Assignment Prediction: Using the cleaned profiles from both internal and external sources, the LLM predicts soft matching scores (probabilities) for paper-author pairs in batches.
Batch Score Refinement: To address calibration inconsistencies across different API call batches, the LLM is prompted again to refine the scores globally, ensuring consistent probability ranges across the entire dataset.

Stage 2: Cross-Correction Supervised Fine-Tuning (SFT)

The refined soft labels serve as supervision for fine-tuning a model (Qwen3-8B with LoRA). The core innovation here is the integration of Probabilistic Soft Logic (PSL) to model logical dependencies within the triple $(p, a_{in}, a_{out})$ :

Data Partitioning: Training data is split into a Consistent Set ( $D_{cs}$ ), where internal and external authors are highly similar, and an Inconsistent Set ( $D_{ics}$ ), where they differ.
PSL-Based Loss: The framework formulates a loss function that enforces logical constraints.
- For consistent sets, if the external paper-author match is high and the authors are similar, the internal match should also be high.
- For inconsistent sets, the model infers which source is incorrect by maximizing the likelihood of the correct assignment while minimizing the probability of the conflicting one.
- The total loss combines this PSL loss with standard Cross-Entropy loss, weighted by a hyperparameter $\lambda$ .

Stage 3: Test-Time Scaling (TTS)

To enhance prediction robustness, the authors employ a bootstrapping strategy during inference:

Confidence Estimation: Papers are sorted by the model's prediction confidence.
Snake Dealing & Interleaving: Papers are distributed into batches using a "snake" pattern (zigzag) and reordered within batches (high-low-high-low). This ensures that high-confidence samples act as "few-shot anchors" to calibrate the model's predictions for low-confidence samples within the same context window.

3. Key Contributions

New Paradigm: The paper introduces Cross-Source Correction, a perspective that utilizes inconsistent assignments across sources as a signal for error detection rather than treating them as noise or ground truth.
Full-Stack Framework: CrossND integrates data refinement (Chain-of-Refinement), logical reasoning (PSL-based SFT), and inference optimization (TTS) to address the challenges of noisy labels and complex relational dependencies.
Noise Mitigation: The Chain-of-Refinement pipeline effectively denoises author profiles and generates reliable soft labels, reducing the reliance on human annotation.
Logical Constraints: The PSL-based cross-correction module explicitly models the interdependencies between paper-author pairs across sources, improving the model's ability to reason about which source contains the error.

4. Experimental Results

The framework was evaluated on two real-world datasets: WhoIsWho (comparing AMiner and MAG) and KDD Cup (comparing AMiner and user-edited data).

Performance: CrossND consistently outperformed 17 baselines, including traditional machine learning models (GBDT, LOF), graph-based methods (DCI), noise-robust self-correction methods (Co-teaching, CORES), and direct LLM API calls (GPT-5, Claude, Gemini).
- On WhoIsWho, CrossND achieved an AUC of 82.19% and MAP of 61.79%, significantly surpassing the best baseline (GuARD: 77.14% AUC).
- On KDD Cup, CrossND achieved an AUC of 87.34% and MAP of 78.13%, outperforming the best baseline (GuARD: 85.27% AUC).
Ablation Studies:
- Removing the external source caused the largest performance drop, confirming the necessity of cross-source signals.
- Removing Profile Cleaning or Batch Score Refinement degraded performance, validating the Chain-of-Refinement pipeline.
- Removing PSL showed mixed but generally negative effects on ranking metrics (MAP), highlighting the value of logical constraints.
Cost Efficiency: The average API cost for the refinement pipeline was approximately $0.0019 per paper on WhoIsWho. The deployed SFT model processes ~26 papers/second on a single H100 GPU.
Real-World Deployment: The system was deployed to assist human verification. In a pilot, experienced annotators found 1,061 assignment errors among 2,154 pairs, a 47% increase in detected errors compared to previous systems.

5. Significance and Claims

The paper claims that CrossND offers a scalable, automated solution for correcting author name disambiguation errors by leveraging the "wisdom of crowds" across different data sources.

Robustness: The framework demonstrates that cross-source reasoning can effectively identify incorrect assignments even when both sources contain noise, provided their inconsistencies are analyzed logically.
Practicality: Unlike methods requiring extensive human annotation, CrossND operates with minimal human intervention and has been successfully deployed to support large-scale paper-author assignment correction in practice (e.g., on the AMiner platform).
Generalizability: The approach is not limited to specific datasets; the use of PSL and chain-of-refinement provides a generalizable strategy for handling noisy relational data in academic search systems.

The authors emphasize that their work does not claim to replace human expertise entirely but rather to augment it by providing a highly accurate, automated pre-screening mechanism that significantly reduces the burden of manual verification.