Research map / Large language models / Retrieval and knowledge
HLMEA: Unsupervised Entity Alignment Based on Hybrid Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HLMEA, an unsupervised entity alignment method where SLMs (BERT-style encoders) filter top-k candidates from textual entity representations and LLMs choose the aligned entity via majority-voted single-choice prompts, with iterative SLM self-training on LLM outputs.
Problem
Entity alignment across knowledge graphs without human labels is hard for conventional methods, and naively prompting LLMs for EA is infeasible due to long prompts, cost, and hallucination; the background knowledge of LLMs must be exploited without supervision.
Approach
Formulate EA as a filtering plus single-choice problem: small language models pre-filter candidates via embedding similarity over pruned textual entity representations, LLMs pick the match by majority voting over repeated queries, and the SLM is self-trained on LLM vote distributions with a negative sampling rule, iterating over rounds.
Key results
- On DBP15K unimodal average Hit@1, HLMEA reaches 0.951 vs. SLOTAlign 0.937 and UPLR 0.927, and exceeds multimodal MEAformer (0.944) on average.
- On DBP15K DE-EN/FR-EN V1, DW15K V1, DY15K V1 HLMEA averages 0.874 Hit@1, 11.6% above RDGCN (0.758) which uses 20% seeds.
- On DBP100K V1 HLMEA averages 0.912 Hit@1, 23.1% above RDGCN (0.681).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PF-IKF rather than random triple selection for TRE generation (ablation) | HLMEA · DBP15KZH−EN | Hit@1 | Helped |
| Use PF-IKF rather than random triple selection for TRE generation (ablation) | HLMEA · DW15K V1 | Hit@1 | Mixed |
| Reduce the number of triples per entity to m=5 (ablation) | HLMEA · DY15K V1 | Hit@1 | Helped |
| Increase the number of SLM self-training rounds (ablation) | HLMEA · DBP15K and DBP100K datasets | Hit@1 | Mixed |
| Replace ChatGPT with Qwen-7B as the LLM (ablation) | HLMEA · DBP15K and DW15K V1 datasets | Hit@1 | Mixed |
| Vary the SLM model (ablation) | HLMEA · DBP15K and DW15K V1 datasets | Hit@1 | Mixed |
| Increase candidate count from top-10 to top-20 (ablation) | HLMEA · DBP15KZH−EN and DW15K V1 | Hit@1 | No clear effect |
| Compare HLMEA with the strongest unimodal baseline | HLMEA · DBP15KZH−EN | Hit@1: .902 -> .930 | Helped |
| Compare HLMEA with the strongest unimodal baseline | HLMEA · DBP15KJA−EN | Hit@1: .930 -> .938 | Helped |
| Compare HLMEA with the strongest unimodal baseline | HLMEA · DBP15KFR−EN | Hit@1: .992 -> .986 | Hurt |
| Use HLMEA rather than the strongest supervised baseline | HLMEA · DBP15KDE−EN V1 | Hit@1: .830 -> .955 | Helped |
| Use HLMEA rather than the strongest supervised baseline | HLMEA · DBP15KFR−EN V1 | Hit@1: .755 -> .957 | Helped |
Surprising findings
- HLMEA outperforms supervised baselines trained with 20% seeds on DBP15K, DW15K and DY15K, despite using no labels.
- Qwen-7B scored higher than ChatGPT on DW15K V1 in the reported LLM comparison.
Limitations
- Relies on closed-source LLMs (ChatGPT, ERNIE) for the best results, with substantial API cost and dependence on external services.
- Performance degrades in later rounds may occur due to noisy candidates or training data, and the method is sensitive to top-k and LLM/SLM choice.
- Cost estimate relies on a single pricing assumption and an external human annotation cost reference.
Experimental setup
Datasets: DBP15K ZH-EN DBP15K JA-EN DBP15K FR-EN DBP15K DE-EN DW15K V1 DY15K V1 DBP100K DE-EN DBP100K FR-EN
Benchmarks: DBP15K ZH-EN DBP15K JA-EN DBP15K FR-EN DBP15K DE-EN DW15K V1 DY15K V1 DBP100K DE-EN DBP100K FR-EN
Metrics: Hit@1 Hit@5 MRR LLM accuracy (LA)
Baselines: MultiKE SelfKG UPLR SLOTAlign EVA MCLEA ICLEA XGEA MEAformer BootEA RDGCN GAEA
Builds on: ChatEA SimpleHHEA BERT CMD distance
Closest papers by meaning
- Entity Alignment with Noisy Annotations from Large Language Models (NeurIPS 2024 · similarity 0.92)
- A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation (AAAI 2025 · similarity 0.88)
- Tackling Uncertain Correspondences for Multi-Modal Entity Alignment (NeurIPS 2024 · similarity 0.88)
- MCL-NER: Cross-Lingual Named Entity Recognition via Multi-View Contrastive Learning (AAAI 2024 · similarity 0.87)
- Lambda: Learning Matchable Prior For Entity Alignment with Unlabeled Dangling Cases (NeurIPS 2024 · similarity 0.87)
- NeuSymEA: Neuro-symbolic Entity Alignment via Variational Inference (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.