PaperlayerView exact revision on arXiv

Source-linked AI summary

AKEW: Assessing Knowledge Editing in the Wild

Xiaobao Wu, Liangming Pan, William Yang Wang, Anh Tuan Luu

arXiv:2402.18909v2cs.CLcs.AI

TL;DR

Knowledge-editing evaluations largely rely on curated structured facts and often omit practical text sources and real-world updates. AKEW benchmarks structured facts, unstructured facts, and extracted triplets using counterfactual and real-world datasets. Experiments find that methods struggle with unstructured facts, even with extracted-triplet assistance.

  • Problem

    Current evaluations focus on curated structured facts and often omit practical unstructured sources and real-world knowledge updates.

    Source: §1 · p. 1, §3.1 · p. 3

  • Method

    AKEW evaluates structured facts, unstructured texts carrying the same updates, and triplets extracted from those texts, using counterfactual and real-world datasets.

    Source: §1 · p. 2

  • Results

    Methods commonly excel on structured facts but fail significantly on unstructured facts, while extracted triplets help some methods yet still fall short of structured facts.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings highlight the challenging nature of practical knowledge editing and the need for more research in this area.

    Source: §6 · p. 9

    The work mainly uses Wikipedia articles or generated Wikipedia-style paragraphs, leaving more diverse sources such as news articles and scientific papers for further evaluation.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Knowledge editing injects knowledge updates into language models to keep them correct and up-to-date. However, its current evaluations deviate significantly from practice: their knowledge updates solely consist of structured facts derived from meticulously crafted datasets, instead of practical sources -- unstructured texts like news articles, and they often overlook practical real-world knowledge updates. To address these issues, in this paper we propose AKEW (Assessing Knowledge Editing in the Wild), a new practical benchmark for knowledge editing. AKEW fully covers three editing settings of knowledge updates: structured facts, unstructured texts as facts, and extracted triplets. It further introduces new datasets featuring both counterfactual and real-world knowledge updates. Through extensive experiments, we demonstrate the considerable gap between state-of-the-art knowledge-editing methods and practical scenarios. Our analyses further highlight key insights to motivate future research for practical knowledge editing.

1 Introduction

Knowledge-editing evaluations largely use curated structured facts, leaving practical text-based and real-world updates underexamined. AKEW evaluates three update formats across new datasets, and experiments reveal a gap between strong structured-fact performance and practical settings.

1 Introduction

  • Methods commonly excel on structured facts but fail significantly on unstructured facts; extracted triplets help some methods but still fall short of structured facts.The results disclose a considerable gap between existing methods and practical scenarios.Source: Figure 1 · p. 1, §1 · p. 2
  • The benchmark is designed to assess how knowledge-editing methods perform in practice across these update formats.Knowledge editing aims to inject target knowledge efficiently and assess whether models recall it.Source: §1 · p. 1
  • Existing evaluations rely on meticulously crafted structured facts and overlook unstructured practical sources and real-world updates.The paper frames this as leaving the diversity and complexity of practical knowledge editing unaddressed.Source: §1 · pp. 1–2
  • AKEW evaluates structured facts, unstructured texts containing the same updates, and triplets extracted from those texts.Its three datasets include counterfactual updates with generated unstructured facts and WIKIUPDATE real-world updates sourced from Wikipedia.Source: §1 · p. 2
  • The authors report extensive experiments with state-of-the-art methods and identify general limitations in practical scenarios, motivating further research.The paper presents three new datasets featuring counterfactual and real-world updates.Source: §1 · p. 2

2 Related Work

Related work spans parameter-updating and parameter-preserving editing methods, while prior evaluations largely center on curated structured triplets. AKEW distinguishes itself by testing more update formats and both counterfactual and real-world updates.

Knowledge-Editing Methods

Knowledge-Editing Evaluations

  • Memory-based methods preserve parameters, while prior evaluations mainly use curated triplets; unlike raw-document work limited to fine-tuning and counterfacts, AKEW tests state-of-the-art methods on three formats and real-world updates.The passage also describes hyper-networks and gradient transformations as parameter-preserving approaches, and memory strategies including SERAC and retrieval-augmented methods.Source: §2 · pp. 2–3

3 Assessing Knowledge Editing in the Wild

The section identifies a mismatch between curated structured-fact evaluations and practical knowledge updates, then introduces AKEW's broader assessment settings. Its formats encompass structured facts, unstructured text, and triplets extracted from that text.

3 Assessing Knowledge Editing in the Wild

  • Knowledge-editing evaluations commonly use isolated structured triplets, queried after editing to test whether models recall the facts.Knowledge editing targets low-cost knowledge updates without expensive retraining.Source: §3.1 · p. 3

3.1 Problems of Knowledge-Editing Evaluations

  • Evaluations overlook practical unstructured sources and real-world updates, leaving the performance of current methods in practical scenarios uncertain.Updates may arise in sources such as news, academic publications, and Wikipedia; existing edits mainly involve outdated information or fabricated counterfacts.Source: §3.1 · p. 3

3.2 AKEW

  • AKEW assesses structured triplets, unstructured texts carrying the corresponding update, and automatically extracted triplets to test whether extraction helps methods handle unstructured facts.The unstructured texts can contain more knowledge than their structured counterparts; together, the settings broaden assessment beyond curated edits alone.Source: §3.2 · pp. 3–4

4 Dataset Construction for AKEW

AKEW constructs datasets covering counterfactual and real-world knowledge updates in structured, unstructured, and extracted-triplet forms. The construction pairs generated or retrieved text with checks for alignment and retains multiple extracted triplets per edit to reflect practical updates.

4.1 Counterfactual Update Datasets

  • COUNTERFACT and MQUAKE-CF supply counterfactual edits; ChatGPT generates corresponding Wikipedia-style text, which is checked against the structured fact and manually filtered for alignment.The generated paragraphs describe the structured facts while ignoring their information accuracy.Source: §4.1 · pp. 4–5

4.2 Real-World Update Dataset

  • WIKIUPDATE adds real-world updates derived from Wikidata, with structured edits paired to aligned Wikipedia summaries as unstructured facts.The dataset construction identifies changes over time using start-time qualifiers and filters summaries that do not align with their structured counterparts.Source: §4.2 · p. 5

4.3 Triplets Extraction

  • Automatic extraction produces multiple related triplets per edit, and the authors leave them unfiltered to test methods' ability to edit all related updates at once.The authors argue filtering could remove valuable updates and prefer methods that handle related triplets together.Source: §4.3 · pp. 5–6

4.4 Dataset Summary

  • WIKIUPDATE contains longer unstructured facts and more extracted triplets per edit than the other two datasets.Source: §4.4 · p. 6

5 Experiment

The experiments compare editing methods across three datasets and editing settings, using GPT2-XL, GPT-J, and Mistral as base models. Methods generally struggle more with unstructured and real-world updates; extracted triplets help some methods but can impair retrieval-based editing.

5.1 Experiment Setup

  • The experiments use GPT2-XL, GPT-J, and Mistral, evaluate editing accuracy across three datasets, and also measure multi-hop accuracy on MQUAKE-CF.ROME and MEMIT cannot take unstructured facts as input; FT and LoRA use them as training inputs, while in-context methods can handle them directly.Source: §5.1 · pp. 6–7, Table 3 · p. 7

5.2 Main Results

  • Unstructured facts cause substantial accuracy declines: on COUNTERFACT, LoRA drops 79% with GPT-J and 94% with Mistral, as methods must parse and integrate more complex natural-language updates.IKE and MeLLo fare relatively better by retrieving unstructured facts as auxiliary context rather than editing model parameters.Source: §5.2 · p. 7
  • IKE and MeLLo retain more performance on unstructured facts by retrieving them as auxiliary context, effectively treating editing as retrieval-augmented generation.With Mistral, IKE drops 21% on COUNTERFACT, while MeLLo drops 14% on MQUAKE-CF; the respective other methods drop 94% and 38%.Source: §5.2 · pp. 7–8
  • Extracted triplets help FT and LoRA and let ROME and MEMIT handle unstructured facts, but multiple related triplets remain difficult to edit precisely and can disrupt IKE and MeLLo retrieval.FT improves from 0.07% to 11.49% on COUNTERFACT with GPT2-XL; related triplets can cause retrieval of an irrelevant fact instead of the needed one.Source: §5.2 · p. 8
  • Editing performance is lower on real-world WIKIUPDATE than on counterfactual updates; IKE's unstructured-fact decline with GPT2-XL grows from 9% on COUNTERFACT to 52% on WIKIUPDATE.The paper attributes greater difficulty partly to WIKIUPDATE's longer unstructured facts and more intricate extracted triplets.Source: §5.2 · p. 8

5.3 Error Analysis

  • Among 100 sampled COUNTERFACT cases where MEMIT fails with extracted-triplet edits, 22% of errors are extraction errors and 78% are editing errors; on WIKIUPDATE, IKE's retrieval accuracy drops sharply for unstructured facts.The analysis links the editing errors to multiple extracted triplets and attributes WIKIUPDATE's retrieval difficulty to more complex real-world updates.Source: §5.3 · pp. 8–9, Figure 4 · p. 8

6 Conclusion and Future Work

AKEW evaluates structured facts, unstructured facts, and extracted triplets across datasets with counterfactual and real-world updates. Experiments find that methods struggle with unstructured facts, while future work should address complex multi-fact edits and limitations of in-context editing.

6 Conclusion and Future Work

  • Existing methods commonly struggle with unstructured facts, even when assisted by extracted triplets, underscoring the challenge of practical knowledge editing.AKEW includes structured facts, unstructured facts, and extracted triplets, with datasets featuring counterfactual and real-world updates.Source: §6 · p. 9
  • Locate-then-edit methods excel on single isolated edits but often fail on complex facts, motivating stronger multi-fact editing.The authors identify editing multiple related facts at once as a future-work direction.Source: §6 · p. 9
  • In-context editing is limited by retrieval success, especially for complicated real-world updates, and by the need for regular, laborious memory maintenance as facts evolve.The authors identify addressing both limitations as a future-work direction.Source: §6 · p. 9

Limitations

The benchmark's unstructured facts mainly come from Wikipedia articles or generated Wikipedia-style paragraphs, and the experiments focus on editing success rather than paraphrased or irrelevant facts.

Limitations

  • Unstructured facts mainly come from Wikipedia articles or generated Wikipedia-style paragraphs; news articles and scientific papers are suggested as additional sources.Evaluating those sources could extend assessment to various practical scenarios.Source: Limitations · p. 9
  • Experiments focus on whether editing succeeds and may be extended to assess performance on paraphrased and irrelevant facts.Source: Limitations · p. 9

Ethics Statement

WIKIUPDATE is built mainly from Wikidata and Wikipedia, whose information may occasionally be inaccurate. The authors describe screening incomplete or incorrectly time-qualified samples and reviewing the dataset for toxic or offensive data.

Ethics Statement

  • WIKIUPDATE is built mainly from Wikidata and Wikipedia, which may contain inaccurate information in a few cases.Source: Ethics Statement · p. 9
  • During construction, the authors removed incomplete samples and samples with incorrect time qualifiers, then reviewed WIKIUPDATE to remove toxic and offensive data.Source: Ethics Statement · p. 9

A Dataset Construction Details

WIKIUPDATE construction selects Wikidata relations, identifies changed objects using time qualifiers, and forms unstructured facts from Wikipedia summaries. The implementation follows established code and hyperparameters for the editing methods.

A Dataset Construction Details

  • The dataset uses Wikidata relations of type WI, mainly physical entities, while excluding virtual-entity types such as ExternalId and GlobeCoordinate.Source: A Dataset Construction Details · p. 13
  • WIKIUPDATE identifies updates by comparing current and pre-discovery objects for subject-relation groups using start and end time qualifiers.The qualifiers are retrieved by relation type, with start time P580 and end time P582.Source: A Dataset Construction Details · p. 13
  • Potential unstructured facts are formed by combining the first 10 sentences of Wikipedia summaries for the subject and object.Source: A Dataset Construction Details · p. 13
  • WIKIUPDATE has longer unstructured facts and more extracted triplets; its relation types span topics including sports, entertainment, business, and politics.Source: A Dataset Construction Details · p. 13, Figure 5 · p. 13

B Model Implementation

  • Implementation follows the source code of MEMIT (Meng et al., 2022b) and EasyEdit (Wang et al., 2023a), uses original IKE and MeLLo code, and retains each method's original hyperparameters.MEMIT and ROME use mom2_adjustment=true for higher editing performance.Source: B Model Implementation · pp. 13–14

C Results with Vicuna

Tests with Vicuna show that the decline on unstructured facts persists, though IKE’s decline on WIKIUPDATE is lower than with GPT-J.

C Results with Vicuna

  • The performance decline on unstructured facts also appears with Vicuna, for both IKE and MeLLo.The authors report that IKE’s decline on WIKIUPDATE is lower than with GPT-J, probably because Vicuna was fine-tuned on data after WIKIUPDATE’s 01/04/2021 update timestamp.Source: C Results with Vicuna · p. 14
  • The experiments use Vicuna (7B), a model fine-tuned from LLaMa with user-shared conversations from ShareGPT.Source: C Results with Vicuna · p. 14
  • IKE and MeLLo are the state-of-the-art editing methods tested with Vicuna.Source: C Results with Vicuna · p. 14

D Error Analysis for Editing with Extracted Triplets

The error examples show several ways extracted-triplet edits can produce incorrect answers, while Appendix E describes prompts for generating, verifying, and extracting dataset facts.

D Error Analysis for Editing with Extracted Triplets

  • Four MEMIT error examples include an omitted Tamil-language fact, an unclear network, an old answer, and an irrelevant answer.For the Eli Maor example, the triplets say he was born and raised in Portsmouth, but the predicted answer is the old object.Source: D Error Analysis for Editing with… · p. 14, Figure 6 · p. 15

E Prompts for Dataset Construction

E.1 Generating Wikipedia-style paragraphs

  • The paragraph-generation prompt asks for a meaning-preserving paraphrase that retains the explicit fact’s subject and object and does not contradict it.Source: §E.1 · p. 16

E.2 Verification

  • The verification prompt asks whether a paragraph conveys a sentence while ignoring accuracy; its example returns False for a UK-location mismatch.Source: §E.2 · p. 16

E.3 Extracting Triplets

  • The extraction prompt requests subject–relation–object triplets; its Taloga example includes the claim that Taloga is the capital of India.Source: §E.3 · p. 16

Related papers

Loading 2402.18909v2…