Enhancing Name Disambiguation via Iterative Self-Refining with LLMs

Xiaocheng Zhang, Yang Zhou, Haoru Chen, Mengjiao Bao, Peng Yan

Keywords: Natural Language Processing, Entity Disambiguation, Entity Clus- tering, Large Language Model

TL;DR: This paper describes the exploration of large language model on the task of name disambiguation, using an iterative self-refining approach.

Abstract:

This paper presents the solution of our team BlackPearl in the WhoIsWho-IND Task of KDD Cup 2024 Open Academic Graph(OAG) Challenge.

The goal of the competition is to explore ways to discover paper assignment errors for given authors. In this paper, We present a LLM-based Name Disambiguator via Iterative Self-Refining. Our method transforms the clustering task into a comparison task, and improves the model’s confidence that the current author belongs to the main class by iteratively improving the proportion of correct authors contained in the model input during reasoning. In addi- tion, we employed Train-Time Difficulty Increase(TTDI), Test-Time Augmentation (TTA) techniques, and multi-source information model ensemble to maximizing the utilization of various informa- tion sources. Our method ranks 1st in the final leaderboard, code is publicly available at https://github.com/BlackPearl-Lab/KddCup-2024-OAG-Challenge-1st-Solutions.

Submission Number: 9