Overview

In this paper, we propose HLMEA, a novel hybrid language model-based unsupervised EA method that synergistically integrates small language models (SLMs) and LLMs. HLMEA formulates the EA task into a filtering and single-choice problem. SLMs filter top candidate entities based on textual representations generated from KG triples. Then, LLMs refine this selection to identify the most semantically aligned entities. An iterative self-training mechanism allows SLMs to distill knowledge from LLM outputs, enhancing the ability of hybrid language models in subsequent rounds cooperatively. The overall architecture is shown as follows:

Environment

To run the script, you need to install the conda first and then the following environment:

conda env create -f environment.yml
conda activate al4ea

The expected structure of files is:

HLMEA
 |-- scripts
 |-- dataset
 |    |-- DBP15K_FR_EN
 |    |-- DBP15K_JA_EN
 |    |-- DBP15K_ZH_EN
 |    |-- DBP15K_DE_EN_V1
 |    |-- DBP15K_FR_EN_V1
 |    |-- DBP100K_DE_EN_V1
 |    |-- DBP100K_FR_EN_V1
 |    |-- DW15K_V1
 |    |-- DY15K_V1

How to use

Firstly, modify llm_call.py with your API key (line 13), service URL (line 32), and access token (line 45).

Then, generate TRE triples based on the datasets, e.g., DBP15K_ZH_EN:

cd scripts
python ent_triple_generation.py -d=dze

Finally, execute HLMEA:

python overall_process.py -d=dze -p=labse -l=gpt3.5 -m=train -r=0

More details about arguments can be found using the -h option or referring to the code.