[Paper Review] DocTER: Evaluating Document-based Knowledge Editing

Suhang Wu, Wang, Ante+5 more|arXiv (Cornell University)|Aug 19, 2023

Topic Modeling4 citations

TL;DR

This paper introduces Eva-KELLM, a new benchmark for evaluating knowledge editing in large language models (LLMs) using raw documents instead of factual triplets. The framework evaluates editing effectiveness, unrelated knowledge retention, reasoning with altered facts, and cross-lingual transfer, revealing that current methods struggle significantly with reasoning and multilingual generalization despite improvements in direct editing.

ABSTRACT

Knowledge editing aims to correct outdated or inaccurate knowledge in neural networks. In this paper, we explore knowledge editing using easily accessible documents instead of manually labeled factual triples employed in earlier research. To advance this field, we establish the first evaluation benchmark, extit{DocTER}, featuring Documents containing counterfactual knowledge for editing. A comprehensive four-perspective evaluation is introduced: Edit Success, Locality, Reasoning, and Cross-lingual Transfer. To adapt conventional triplet-based knowledge editing methods for this task, we develop an Extract-then-Edit pipeline that extracts triples from documents before applying existing methods. Experiments on popular knowledge editing methods demonstrate that editing with documents presents significantly greater challenges than using triples. In document-based scenarios, even the best-performing in-context editing approach still lags behind by 10 points in editing success when compared to using gold triples. This observation also holds for both reasoning and cross-lingual test sets. We further analyze key factors influencing task performance, including the quality of extracted triples, the frequency and position of edited knowledge in documents, various methods for enhancing reasoning, and performance differences across various directions in cross-lingual knowledge editing, which provide valuable insights for future research.

Motivation & Objective

• To address the limitations of existing knowledge editing benchmarks that rely on costly, narrow factual triplets.
• To develop a more practical and general evaluation framework using raw documents as editing inputs.
• To comprehensively assess LLMs not only on direct editing success but also on reasoning with altered knowledge and cross-lingual transfer.
• To identify critical gaps in current knowledge editing methods, especially in reasoning and multilingual capabilities.

Proposed method

• The benchmark uses raw documents as input for knowledge editing, enabling a more natural and scalable approach compared to triplet-based methods.
• It evaluates LLMs across four dimensions: Direct Knowledge Editing Evaluation (DKEE), Unrelated Knowledge Retention Evaluation (UKRE), Indirect Knowledge Editing Evaluation (IKEE) for reasoning, and Cross-lingual Knowledge Editing Evaluation (CKEE) for multilingual transfer.
• The framework employs probability-based scoring to measure model confidence in original vs. edited facts, ensuring quantifiable assessment.
• Fine-tuning with LoRA is applied to specific layers (e.g., MLP and self-attention) to analyze parameter-level impact on editing performance.
• The dataset is constructed to include diverse factual edits, reasoning questions, and cross-lingual query pairs to support comprehensive evaluation.
• Experiments compare methods like LoRA-finetuning on different model components (MLP, self-attention) to isolate effective editing strategies.

Experimental results

Research questions

• RQ1 Can knowledge editing using raw documents achieve higher effectiveness and broader applicability than triplet-based methods?
• RQ2 To what extent can edited LLMs reason correctly with newly inserted knowledge rather than just memorizing it?
• RQ3 How well do current editing methods generalize to cross-lingual settings where queries are in a different language than the training documents?
• RQ4 What role do specific model layers play in successful knowledge editing, and which are most effective for parameter adjustment?

Key findings

• The +LoRA(Self-attention+MLP) method improved the ES score from 23.88 to 44.33 on BLOOM-3B, showing significant progress in direct editing, but at the cost of reduced unrelated knowledge retention.
• The UKRE metric showed that improved DKEE performance often led to worse retention of unrelated knowledge, highlighting a key trade-off.
• Reasoning with altered knowledge remained weak: the highest IKEE accuracy was only 48.00 on BLOOM-7.1B, indicating poor integration of edited facts into reasoning.
• Cross-lingual knowledge transfer was severely limited, with CES scores significantly lower than ES scores, showing poor generalization to non-target languages.
• Middle layers (11–20) of the model were most effective for editing, with +LoRA(MLP 11–20) and +LoRA(Self-attention 11–20) outperforming other layer groups.
• Larger models (BLOOM-7.1B) showed greater improvement from editing methods than smaller ones (BLOOM-3B), suggesting model scale enhances editability.