Related Knowledge Perturbation Matters

This repository includes the data and python implementation for the paper "Related Knowledge Perturbation Matters: Rethinking Multiple Pieces of Knowledge Editing in Same-Subject." (Accepted by NAACL 2025 Main Conference).

DataSet

Our dataset is located in S2RKE.json. Compared to the dataset COUTERFACT, the knowledge density surrounding the same topic is higher.

Item	S2RKE	COUNTERFACT
Records	22064	21919
Subjects	4503	20391
Relations	43	32
Maximum records per subject	13	4
Minimum records per subject	3	1
Average records per subject	4.9	1.1

Example

{
    "subjectID": 0,"class": "xxx",
        "subject": {
            "name": "Double Xposure","URL": "/resource/schema:Movie",
            "original_name": "<Double_Xposure>","description": "2012 film directed by Li Yu"
        },
        "Star_Topology": [
            {
                "requested_rewrite": {
                    "prompt": "{}, which is located in","relation": "<isLocatedIn>",
                    "target_true": "China", "target_new": "Głogów County"
                },
                "paraphrase_prompts": [
                    "Double Xposure, a 2012 film by Li Yu, is set in"
                ],
                "neighborhood_prompts": [
                    "Ping Zhang is a citizen of", "Ricky Lee has citizenship in"
                ]
            },"..." ]}

Experiment

Model preparation

Model name	Link
GPT2-XL	[Huggingface link]
GPT-J-6B	[Huggingface link]
Llama2-7B	[Huggingface link]

Results of Same-Editing