PaperlayerView exact revision on arXiv

Source-linked AI summary

Probing Factual Knowledge Transfer with Training Data Interventions

Romina Oji, Marc Braun, Marcel Bollmann, Marco Kuhlmann, Jenny Kunz

arXiv:2609.01341v1cs.CLcs.AI

TL;DR

The paper asks whether facts in multilingual models transfer across languages during continued pretraining or are learned directly from target-language data. It intervenes on Persian training data starting from an English-pretrained model and evaluates transfer with SIFT; transfer is very limited, especially for rare Persian-related entities, and depends strongly on evaluation difficulty and source-language frequency.

  • Problem

    Existing evaluation datasets do not support controlled training-data manipulation, leaving it difficult to distinguish cross-lingual transfer from direct target-language learning.

  • Method

    The study continues pretraining an English model on Persian data with systematically removed facts and uses SIFT’s 500 triples, native Persian templates, and stratified candidate sets for evaluation.

  • Results

    Cross-lingual fact transfer is very limited under strict fact removal and hard-distractor evaluation, with transfer largely confined to general facts and Persian-related facts near chance.

  • Takeaways & Limitations

    Strong English-side representation is necessary but insufficient for reliable transfer, while rare Persian-related entities transfer especially poorly.

  • Takeaways & Limitations

    The study is limited to the English–Persian language pair and models with 572M parameters, leaving behavior across scripts and larger models open.

Abstract

from arXiv · show

Do multilingual language models transfer factual knowledge across languages during continued pretraining, or do they mostly recall facts learned directly from the target-language data? To answer this question more reliably, we propose an intervention-based framework: starting from an English-pretrained model, we continue pretraining on Persian data from which specific facts have been systematically removed at varying levels of granularity. We construct SIFT, a resource of 500 triples across 20 relations, stratified by the cultural origin of each fact's subject into general (globally prominent) and Persian-related entities, designed for both systematic fact removal from training data and evaluation, with natively written Persian cloze templates. Our results show that fact transfer is very limited: under the strictest removal condition, a large majority of English-acquired facts fail to transfer into Persian. We further show that sentence-level co-occurrence removal is insufficient to eliminate fact signal, and that easier (randomly selected) negative candidate sets substantially inflate apparent transfer by rewarding shallow associative heuristics, while performance on a harder candidate set that allows for less reliance on heuristics is much lower. Finally, we show that source-language entity frequency has a large influence, with Persian-related facts, which are orders of magnitude rarer in the English corpus, hardly transferring.

1 Introduction

The paper tests whether factual knowledge transfers from English into Persian during continued pretraining rather than being learned directly from Persian data or inferred from shallow cues. Controlled interventions show that transfer is very limited, especially for Persian-related facts.

  • The causal framework separates Persian recall into English-to-Persian transfer, direct Persian acquisition, and shallow-cue inference.The interventions block direct acquisition and shallow cues to estimate the transfer path.
  • The study targets inconsistent evidence about whether multilingual models transfer facts across languages, a question difficult to resolve without access to training data.The intervention framework addresses this evidence gap by manipulating Persian training data.
  • SIFT provides 500 triples across 20 relations, divided into general and Persian-related entities for training-data intervention and evaluation.The resource supports controlled fact removal and distinguishes globally prominent from culturally rooted subjects.
  • Under strict subject removal and hard-distractor evaluation, the English–Persian setting shows very limited cross-lingual fact transfer.The languages share no orthographic overlap and virtually no vocabulary, requiring learned semantic associations.
  • Transfer covers only a small fraction of English-acquired knowledge, largely general facts, while Persian-related facts transfer at substantially lower rates.Persian-related subjects are orders of magnitude rarer in English; frequent source-language representation is necessary but not sufficient.

2 Related Work

Prior work establishes that multilingual factual recall is uneven and can rely on shallow cues, but existing resources do not support controlled training-data interventions. This motivates a dataset designed to isolate cross-lingual transfer from direct learning.

  • Existing multilingual fact-probing datasets are designed for evaluation rather than systematic training-data manipulation.Their design prevents researchers from cleanly separating transferred knowledge from facts learned directly in the target language.
  • Multilingual models show performance gaps across languages and often retrieve facts less reliably outside the language in which they were expressed.Prior findings also report only shallow knowledge alignment from multilingual pretraining and instruction tuning.
  • Models may use prompt biases, co-occurrence statistics, and other statistical shortcuts instead of genuinely storing facts.This makes candidate difficulty and training-data interventions important for interpreting apparent factual transfer.

3 SIFT

SIFT is a curated, bilingual resource designed to make fact removal meaningful and evaluation challenging. It combines provenance-stratified triples, native Persian templates, and hard or random distractors.

  • SIFT resource: SIFT contains 500 factual triples across 20 relations, annotated as 114 Persian-related and 386 general triples.The resource is designed for both training-data intervention and evaluation.
  • SIFT resource: The dataset fills gaps in prior resources by selecting facts with sufficient Persian-corpus frequency for meaningful removal and adding Persian-related entities.It avoids entities so frequent that removing them would distort the corpus.
  • Fact Triples: Triples serve both as targets for removing training evidence and as items for evaluating factual recall.Each entry also records English and Persian subject frequencies and subject–object co-occurrence counts.
  • Evaluation Templates: Persian templates are written natively rather than machine-translated, allowing natural word order and lexical variation.Five templates are created per relation, and Persian forms include orthographic and lexical variants.
  • Candidate sets: Hard distractors are semantically close to the correct object, whereas random distractors are sampled without regard to semantic proximity.Hard candidates reduce the usefulness of surface cues, while random candidates can often be solved using shallow associations.

4 Experimental Setup

The experiments continue pretraining an English base model on Persian corpora modified by controlled fact-removal interventions, then evaluate recall and transfer with SIFT. Multiple model conditions and candidate sets separate direct learning, corpus-size effects, and shallow heuristics.

  • Training and intervention: The experimental design applies three Persian-corpus fact-removal levels, a random-removal control, continued pretraining, and SIFT-based evaluation.The interventions target direct Persian factual acquisition while the evaluation tests recall and cross-lingual transfer.
  • Training and intervention: Co-occurrence removal deletes sentences where subject and object variants appear together, while full subject removal deletes every sentence containing a subject variant.Full template replacement preserves the subject surface form in neutral sentences without exposing target factual associations.
  • Training and intervention: The random-removal control removes a comparable amount of text while retaining sentences containing SIFT subjects or objects.This separates fact-specific removal from the effect of reducing corpus size.
  • Models: All models use the same 572M-parameter architecture, while continued-pretraining models include approximately 12% replay data from the English and code corpora.Replay is included to help preserve source-language factual knowledge during Persian adaptation.
  • Evaluation: Transfer is Persian accuracy conditioned on correct English answers, while Non-Transfer conditions on incorrect English answers; all metrics are computed for hard and random candidate sets.RankC measures the Spearman correlation between English and Persian candidate orderings.

5 Results and Discussion

The results show that rigorous subject removal is necessary to isolate cross-lingual transfer, which remains limited and is especially vulnerable to shallow heuristics and sparse English representations. Transfer rises with English subject frequency, but Persian-related facts remain substantially less transferable than general facts.

  • How Much Removal Is Enough?: 25.8 vs. 20.6 Overall and 48.9 vs. 39.9 Transfer show that co-occurrence removal leaves more fact signal than full subject removal on hard candidates.Under CPT-CoocRM, the model also achieves 30.1% Non-Transfer accuracy, suggesting Persian evidence remains even when English retrieval fails.
  • How Much Removal Is Enough?: 47.2% Persian accuracy after random removal is statistically indistinguishable from CPT-Full’s 46.6%, whereas CPT-SubjRM reaches only 32.4% on hard candidates.Comparable token removal therefore does not explain the decline; the authors attribute it to removing factual evidence.
  • Limited Cross-Lingual Transfer: CPT-SubjRM achieves only 20.6% Overall and 39.9% Transfer on hard candidates, indicating that most English-acquired facts do not transfer successfully into Persian.The 15.5-point Transfer–Non-Transfer gap, absent in EN-Base, is consistent with transfer rather than only inherent fact difficulty.
  • Precise Knowledge or Shallow Heuristics?: 10 and 15.8 percentage-point Persian and English accuracy gaps between random and hard candidates show that easier candidates inflate performance through shallow associative cues.Hard candidates use semantically and culturally close distractors to reduce reliance on these heuristics.
  • Entity Frequency and Provenance Affect Transferability: 46–51% Transfer for subjects in the 100K–10M English-frequency range contrasts with near-floor transfer below 1,000 occurrences, while the highest-frequency bin remains below CPT-Full at 51% vs. 82%.English-side representation strength is therefore associated with greater Persian retrieval, but English exposure alone is insufficient for reliable transfer.
  • Entity Frequency and Provenance Affect Transferability: 41.5% of general facts transfer on hard candidates versus 25.0% of Persian-related facts, whose median English frequency is 284 compared with 196K for general subjects.EN-Base shows no comparable transfer gap, while CPT-SubjRM English accuracy is 21.1% for Persian-related facts versus 60.6% for general facts.

6 Conclusion and Future Work

The paper establishes that cross-lingual factual transfer from English to Persian exists but is limited, especially when interventions remove Persian fact evidence and suppress shallow retrieval cues. Transfer is concentrated among general facts with frequent English representations, while script differences may further constrain alignment.

  • Conclusion: Controlled interventions show that English-to-Persian factual transfer exists but transfers only a small fraction of English-acquired knowledge.The conclusion attributes this evidence to removing facts from Persian data and reducing access to shallow heuristics.
  • Conclusion: Low-English-exposure entities transfer at near-chance levels, while even highly frequent entities show limited performance.English-side representation strength is necessary but insufficient for reliable transfer.
  • Conclusion: Transfer is largely confined to general facts whose subjects are frequent in the English corpus.The paper links stronger transfer to frequent English representations rather than reliable transfer across all fact types.
  • Future Work: Persian’s Arabic-derived script has no orthographic or token overlap with English, potentially hindering entity-representation alignment.The paper proposes transliterated pretraining data as a future test of whether bridging the script gap improves transfer.

Limitations

The study’s scope is constrained to a single English–Persian setting and models with 572M parameters. Extending the framework to other languages and larger models is challenging and remains necessary for testing generality and scaling effects.

  • Language scope: The study evaluates only Persian as the target language, limiting conclusions about script similarity and transfer beyond the English–Persian pair.Comparisons with Latin-script and additional non-Latin-script languages are identified as needed to test generality.
  • Language scope: Expanding to new languages requires native templates, manually curated country-specific triples, and expert-designed hard candidate sets.These requirements make extension a substantial annotation effort rather than a simple translation task.
  • Model scale: 572M-parameter models limit conclusions about whether controlled fact removal yields the same transfer pattern at larger scales.Testing this question would require training larger models.

A Causal Effect Estimation of Factual Knowledge Transfer

The study estimates cross-lingual factual transfer by removing Persian evidence and minimizing shallow cues, while retaining English-pretraining exposure. Above-chance Persian prediction under these interventions is treated as evidence of transferred English factual knowledge.

  • Causal setup: The causal design targets the contribution of English-acquired factual knowledge to correct Persian predictions.Xe represents English access, Xp Persian continued-pretraining access, M shallow cues, and Y correct Persian prediction.
  • Causal setup: Direct intervention on English factual exposure is infeasible because all models start from an English-pretrained checkpoint.The counterfactual model trained identically without the relevant English facts is never observed.
  • Interventions: The controlled effect differs from unconditional factual recall because it isolates cross-lingual transfer when Persian evidence and shallow heuristics cannot support prediction.If a fact remains available in Persian, English knowledge may add little because Persian evidence already suffices.
  • Interventions: Persian factual exposure is approximated by systematically removing facts from the Persian corpus, while hard candidates are designed to minimize shallow-heuristic utility.The hard setting is assumed to approximate removal of shallow cues, or do(M = 0).
  • Interpretation: Successful prediction above chance is interpreted as evidence that the model relies on factual knowledge transferred from English.The chance baseline is 0.2 because candidate sets contain five options.

B Triple Creation Criteria

The triple-construction criteria are designed to make interventions comprehensive and evaluations nontrivial across diverse entities, relations, surface forms, and answer spaces. The dataset also controls for subject frequency and demographic or temporal concentration.

  • Diversity and coverage: Subject diversity avoids broad removal of unrelated text caused by entities appearing across many relations.Prominent subjects such as Iran, China, Russia, Samsung, and Microsoft accounted for approximately 6% of tokens under full subject removal.
  • Diversity and coverage: Temporal diversity adds contemporary public figures to reduce bias toward historically prominent entities in encyclopedic sources.The generated candidates were heavily skewed toward historical figures before this correction.
  • Evaluation validity: Triples with near-identical subject and object strings are replaced to prevent unintended removals and string-overlap shortcuts.For example, Germany–OFFICIAL_LANGUAGE–German was replaced by Austria–OFFICIAL_LANGUAGE–German.
  • Intervention coverage: Person-name variant lists support comprehensive matching across full and shortened forms during intervention filtering.Variants such as “Pablo Picasso” and “Picasso” were verified against English and Persian Wikipedia pages.
  • Intervention coverage: All valid objects and Persian surface-form variants are enumerated so unlisted facts cannot remain in the corpus and mimic transfer.This prevents the model from learning an omitted object directly from unfiltered Persian data.
  • Diversity and coverage: Object diversity broadens each relation beyond its most frequent values, including both prominent and less prominent capitals.RELIGION includes eight traditions, while CAPITAL includes examples such as Paris and Tokyo alongside less prominent capitals.
  • Corpus control: Corpus frequency control prevents removing highly frequent triples when doing so would distort the Persian training distribution.

C Training Data Replacement Templates

The template-replacement intervention substitutes neutral Persian sentences for sentences containing targeted subjects, preserving subject surface forms without exposing factual associations.

  • Replacement design: Neutral replacement sentences preserve each subject’s surface form while removing target factual associations from the training signal.The method uses 100 neutral Persian sentences per entity type, with a shared “{subject}” placeholder across relations of that type.

D How Does Fact Recall Vary By Relation?

Fact recall varies substantially by relation: frequent, formulaic relations are easiest and show the largest intervention drops, whereas some person-attribute relations remain above the empirical floor. Other relations remain weak or even degrade after continued pretraining, and template phrasing does not explain the main patterns.

  • Relation-specific recall: CAPITAL_OF and CONTINENT accuracy drops 32–48 points from CPT-Full to CPT-SubjRM, including 96%→48% and 92%→52%, respectively.These frequent, formulaic relations are easiest across models; CONTINENT also has only six possible objects.
  • Relation-specific recall: OCCUPATION, DEVELOPER, and LANGUAGES_SPOKEN drop only 4–20 points under subject removal and remain at 40–48% in CPT-SubjRM.For relations with few possible answers, such as RELIGION with eight unique objects, residual performance may partly reflect answer-space constraints.
  • Relation-specific recall: PLACE_OF_DEATH stays near or below the empirical floor across languages and models, with no CPT-Full-to-CPT-SubjRM drop.This fact type was not acquired even from full Persian data.
  • Relation-specific recall: LANGUAGE_OF_WORK_OR_NAME declines from 36% in EN-Base to 16% in CPT-SubjRM for Persian accuracy.
  • Template robustness: Per-template standard deviation is at most 5.2 percentage points, and model ordering is consistent across hard and random candidate settings.This supports robustness to variation across the five evaluation templates.
  • Candidate construction: Hard distractors are generated as four incorrect, semantically and culturally proximate candidates per triple and verified by a native Persian speaker.Geographic distractors use neighboring or commonly confused locations, while person-attribute distractors share cultural or professional domains.

G Mean Results on Different Templates

Results remain broadly consistent across descriptive and question-answering templates, although the relative ordering of CPT-SubjRP and CPT-SubjRM reverses under question answering. Hard and random candidate settings are evaluated using distinct distractor constructions, with template-level distributions summarized across five templates.

  • Results are broadly consistent across descriptive and question-answering formats, with the same model ordering and overall patterns.This consistency indicates that the findings are not sensitive to the specific evaluation format.
  • 2.9 points on hard and 5.7 points on random candidates separate CPT-SubjRP from CPT-SubjRM under question answering.CPT-SubjRP slightly outperforms CPT-SubjRM in question answering, reversing the pattern observed with descriptive templates.
  • Each model’s accuracy distribution aggregates five evaluation templates, with individual template scores and mean values shown.The figure reports narrow template-level variation, with standard deviations no larger than 5.2 percentage points.
  • Hard distractors are generated as four incorrect candidates per triple using relation-specific rules designed to suppress shallow associative cues.The construction includes semantically related alternatives such as neighboring capitals, similar spellings, and commonly mistaken entities.
  • Tables 4–7 report mean accuracy and standard deviation across five templates for both candidate settings and both template formats.The tables separately cover descriptive and question-answering templates with hard and random candidates.
Loading 2609.01341v1…