ClinicalAligner26AM: A Cross-Lingual Aligner for Dataset Translation; Evidences from the MultiClinCorpus Shared Task

Technical Summary: ClinicalAligner26AM

Problem Statement
The paper addresses the challenge of cross-lingual annotation projection in clinical Natural Language Processing (NLP). In clinical domains, annotated data is scarce, and entity boundaries are critical. While source-language annotations (e.g., in Spanish) may be available, transferring them to target languages (e.g., English, Italian, Dutch) is difficult because translations often alter surface forms while preserving content. Direct span classification is often too rigid to handle the expansion, contraction, reordering, or splitting of entities that occur during translation. The paper posits that word-level cross-lingual alignment is a more reliable primitive for this task than training new extractors from scratch, yet existing neural aligners are rarely adapted to specialized biomedical domains.

Methodology
The authors introduce ClinicalAligner26AM, a large-context multilingual token aligner initialized from ClinicalEncoder26AM. The approach follows a teacher-student distillation framework inspired by AWESoME Align, adapted specifically for clinical text.

1. Teacher Construction (Frozen Target):
   Instead of relying on a single bilingual embedding space, the authors construct a "frozen" source-target cost matrix by fusing three complementary semantic views and a positional prior:

   • Sentence-level: Derived from ClinicalEncoder26AM-L2.
   • Phrase-level: Derived from ClinicalEncoder26AM-L1.
   • Token-level: Derived from XLM-Align, utilizing a 5-token contextualization heuristic (concatenating two preceding and two following tokens with a discount factor) to reduce noise from short, low-weight tokens.
   • Positional Prior: A weak bias assigning maximum mass within a 50-token window around the diagonal, decaying up to 100 tokens, to reflect the preservation of discourse order in translations.

   These signals are summed into an initial cost matrix and sharpened using the Sinkhorn–Knopp algorithm (100 iterations) to produce a doubly stochastic, sparse soft alignment target. This step converts heterogeneous similarities into a globally consistent many-to-many distribution.

2. Student Training:
   The student model, ClinicalAligner26AM, is trained to match this frozen target. It encodes source and target texts independently and computes a direct cosine-based alignment matrix. The loss function minimizes the difference between the student's naive cosine similarities and the optimized teacher target, using row-wise and column-wise cosine similarity losses rather than KL divergence to maintain efficiency and consistency.

3. Competition Variant (MCAI):
   A variant, ClinicalAligner26AM-MCAI, injects additional supervision by adding a positive bias to token pairs falling within manually mapped spans from the MultiClinCorpus training split. This nudges the teacher target toward alignments that better reflect the specific annotation-projection regime of the shared task.

4. Inference Pipeline:
   At inference, the model projects source entity membership scores onto target tokens using the learned alignment matrix. The system then decodes the longest valid high-scoring span in the target text, allowing for at most one gap token (for punctuation or mis-attributed tokens). This process can be optionally supported by predictions from a MultiClinNER system, though the alignment remains the primary transfer mechanism.

Key Contributions

• Domain-Specialized Aligner: The introduction of ClinicalAligner26AM, a token-level aligner specifically initialized and trained for biomedical and clinical text, addressing the gap in specialized domain alignment.
• Multi-View Training Recipe: A novel method for constructing alignment targets by fusing sentence, phrase, and token-level signals with optimal transport sharpening, distilled into a lightweight student model.
• Competition Variant: The development of ClinicalAligner26AM-MCAI, which demonstrates how manual span supervision can refine boundary decisions without altering the core training architecture.

Results
The systems were evaluated on the MultiClinCorpus shared task, which involves projecting Spanish clinical entity annotations (Procedure, Disease, Symptom) into six target languages.

• Ranking: The two submitted systems (CA26AM and CA26AM-MCAI) ranked first and second across all languages and entity types.
• Performance: Character-weighted F1 scores exceeded 0.95 in nearly all settings, approaching 0.97 for several language-entity combinations (particularly in English, Italian, and Romanian).
• Comparison: The aligner models significantly outperformed a baseline using only CE26AM (without source alignment), which struggled with precision and recall trade-offs typical of mention discovery. The results indicate that when source entities are known, the task simplifies to localization, where token-level alignment is highly effective.
• Variant Analysis: The MCAI variant showed modest but consistent improvements over the base aligner, suggesting that while the multi-view frozen target drives most performance, annotation-aware biasing helps refine difficult boundary decisions.

Significance and Claims
The paper claims that token-level alignment is a sufficient and effective mechanism for transferring exact clinical spans across translated documents when source annotations are trusted. By framing cross-lingual transfer as a localization problem rather than a full mention-discovery problem, the method achieves near-perfect multilingual entity transfer. The authors suggest that this formulation naturally extends to richer annotation schemas (nested spans, metadata transfer) and downstream applications such as translation auditing, error localization, and cross-lingual faithfulness assessment, without requiring complex inference-time cross-attention mechanisms.