Entity Alignment with Noisy Annotations from Large Language Models

Authors: Shengyuan Chen, Qinggang Zhang, Junnan Dong, Wen Hua, Qing Li, Xiao Huang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the advantages of LLM4EA on four benchmark datasets in terms of effectiveness, robustness, and efficiency.
Researcher Affiliation Academia Shengyuan Chen Department of Computing The Hong Kong Polytechnic University Hung Hom, Hong Kong SAR shengyuan.chen@connect.polyu.hk Qinggang Zhang Department of Computing The Hong Kong Polytechnic University Hung Hom, Hong Kong SAR qinggangg.zhang@connect.polyu.hk Junnan Dong Department of Computing The Hong Kong Polytechnic University Hung Hom, Hong Kong SAR hanson.dong@connect.polyu.hk Wen Hua Department of Computing The Hong Kong Polytechnic University Hung Hom, Hong Kong SAR wency.hua@polyu.edu.hk Qing Li Department of Computing The Hong Kong Polytechnic University Hung Hom, Hong Kong SAR csqli@comp.polyu.edu.hk Xiao Huang Department of Computing The Hong Kong Polytechnic University Hung Hom, Hong Kong SAR xiaohuang@comp.polyu.edu.hk
Pseudocode Yes Algorithm 1 The greedy label refinement algorithm
Open Source Code Yes We have provided the code for the framework, accessible via this URL: https://github.com/chensyCN/llm4ea_official.
Open Datasets Yes In this study, we use the widely-adopted Open EA dataset (Sun et al., 2020), including two monolingual datasets (D-W-15K and D-Y-15K) and two cross-lingual datasets (ENDE-15K and EN-FR-15K). Open EA comes in two versions: "V1" the normal version, and "V2" the dense version. We employ "V2" in the experiments in the main text.
Dataset Splits No The paper mentions training, but does not explicitly state training, validation, and test splits with percentages or counts. It refers to standard datasets but not their specific splits.
Hardware Specification Yes Our experiments were conducted on a server equipped with six NVIDIA Ge Force RTX 3090 GPUs, 48 Intel(R) Xeon(R) Silver 4214R CPUs, and 376GB of host memory.
Software Dependencies Yes The details of the software packages used in our experiments are listed in Table 4. Table 4: Package configurations of our experiments. Package tqdm numpy scipy tensorflow keras openai Version 4.66.2 1.24.4 1.10.1 2.7.0 2.7.0 1.30.1
Experiment Setup Yes Setup of LLM4EA. We employ GPT-3.5 as the default LLM due to its cost efficiency. Other parameters are n = 3, nlr, k = 20, δ0 = 0.5, δ1 = 0.9.