Zero-shot cross-lingual evaluation of contextual embeddings using target-language development sets
Abstract
Abstract: Multilingual contextual embeddings have demonstrated state-of-the-art performance in zero-shot cross-lingual transfer learning, where multilingual BERT is fine-tuned on one source language and evaluated on a different target language. However, published results for mBERT zero-shot accuracy vary as much as 17 points on the MLDoc classification task across four papers. We show that the standard practice of using English dev accuracy for model selection in the zero-shot setting makes it difficult to obtain reproducible results on the MLDoc and XNLI tasks. English dev accuracy is often uncorrelate
Research Question
What is the impact of using target-language-specific development sets instead of English for model selection in zero-shot cross-lingual evaluation of contextual embeddings, as measured by F1 scores on the XNLI dataset?
Verification Level
| Paper level | L2, Source-grounded claims | |
| Source-grounded claims | 8 | |
| Claim record source | parsed source sections |
Descriptive public verification status only; aggregate claim counts are public, but individual claim records are not exposed here.
Truth-Engine Gate Verdict
| Status | Falsified | |
| Gate | Gate 2 — Verification (formal proof or sandbox reproduction) | |
| Reason | [Gate 3 RED-TEAM FALSIFIED] avg_attack_score=9.1/10. COUNTEREXAMPLE_HUNTER(9.0):Fatal Dataset Mismatch: The research goal explicitly asks about results on the *; CITATION_AUDITOR(9.2):The verification script performs a trivial identity assertion (hardcoding 15.0 a; REPLICATION_ATTACKER(9.0):The verification script merely hardcodes the claimed value (15.0) and asserts eq | |
| Evaluated | 2026-07-13T13:27:36.866690+00:00 |
A claim in this record was tested against Gate 2 and failed: a counterexample was found, a proof did not type-check, or a reproduction attempt did not match the reported results. Evidence for the failure is attached to this record. VERIFIED requires an attached reproducible artifact (Lean4 proof source, or repro script and results) before this status can be set; it is not derived from review score or claim count.
Quality Tier
| Tier | Flagship candidate | |
| Basis | Review score, verified-claim count, and public artifact coverage meet flagship-candidate thresholds. |
Descriptive public triage only; this tier does not alter current publication or DOI behavior.
Quality Dimensions
| Evidence strength | MEDIUM | |
| Citation grounding | MEDIUM | |
| Uncertainty disclosure | MEDIUM | |
| Reproducibility status | HIGH |
Automated triage signals derived from public fields; not human peer review or independent validation.
Correction Record
| Status | CURRENT |
| Correction count | 0 |
| Manifest contract | paper-manifest-v1.1 |
| Correction contract | correction-record-v1 |
Public corrections are additive records. Current status does not claim the synthesis is error-free.
Provenance
| Publisher | Assignee Research |
| Public provenance | L4, External archival record |
| Report artifact | Available |
| External record | Registered |
| Claim lineage | 8 aggregate source-grounded claims |
| Review method | Automated multi-reviewer assessment |
| Quality guide | How to read scores, claims, manifests, and evidence links |
| Provenance contract | source-provenance-v1 |
| Note | Machine-generated synthesis of existing literature. Not primary research. |