Active research and competition system
Clinical Entity Recognition and Linking Pipeline
- Started
- 2026-06
- Last revised
A Vietnamese clinical NLP pipeline that separates span detection, assertion handling, candidate retrieval, and ontology linking so each source of error can be inspected independently.
- XLM-R
- Linear-chain CRF
- PyTorch
- BM25
- Dense retrieval
- Cross-encoder reranking
- RxNorm
- ICD
Research question
How can a Vietnamese clinical system preserve precise entity boundaries and assertion context while linking noisy surface forms to controlled medical vocabularies with limited annotated data?
From input to inspectable output
System architecture
- 01
Clinical text normalization
Preserve medically meaningful punctuation while standardizing whitespace and known abbreviation variants.
- in
- note text
- out
- token sequence
- 02
Span recognition
XLM-R contextual representations feed a CRF so adjacent BIO decisions are decoded as a valid sequence.
- in
- tokens
- out
- typed spans
- 03
Assertion layer
A separate contextual classifier assigns present, absent, possible, or historical status without changing span boundaries.
- in
- span + context
- out
- asserted entity
- 04A
Lexical candidate retrieval
Normalized aliases and character-aware lexical search recover terminology with strong surface overlap.
- in
- entity mention
- out
- top-k lexical IDs
- 04B
Dense candidate retrieval
Multilingual mention and concept embeddings recover semantic candidates missed by exact terminology.
- in
- entity + context
- out
- top-k semantic IDs
- 05
Candidate fusion and reranking
Deduplicated candidates are calibrated and reranked with the local sentence before an ontology ID is selected.
- in
- candidate union
- out
- RxNorm / ICD link
What the design must respect
Operating constraints
- 01
Clinical abbreviations, spelling variation, and mixed Vietnamese–English terminology make exact lexical matching brittle.
- 02
The annotation budget is limited, so the pipeline must make useful progress before a large task-specific corpus exists.
- 03
Negation, hypothetical mentions, and family-history context must remain attached to an entity after recognition.
- 04
RxNorm and ICD candidate spaces differ in granularity; a shared retrieval score is not automatically comparable.
- 05
Competition-time inference needs bounded candidate sets and reproducible preprocessing.
Evaluation ledger
Current results
Values will be added only with the evaluation set, protocol, ontology version, and relevant runtime conditions.
- R01Entity span F1
- Not reportedPlaceholder
Reserved for a held-out, versioned evaluation set; no score is claimed yet.
- R02Assertion macro F1
- Not reportedPlaceholder
Will be split by present, absent, possible, and historical classes.
- R03Linking recall@20
- Not reportedPlaceholder
The first retrieval measure to report, before reranking accuracy.
- R04End-to-end linked entity F1
- Not reportedPlaceholder
Requires the span, assertion, and ontology versions to be frozen together.
Negative evidence
Failed or insufficient approaches
Single-stage exact alias matching
Useful as a precision-oriented baseline, but it cannot resolve abbreviations, inflections, or context-dependent concepts.
Keep exact matches as a retrieval feature and add lexical plus dense candidate generation.
Linking every raw model span immediately
Boundary errors become opaque linking errors, which makes evaluation difficult to interpret.
Evaluate span detection, candidate recall, and reranking separately before measuring the end-to-end path.
One similarity threshold for every ontology
Scores are not calibrated across concept families or between RxNorm and ICD inventories.
Calibrate by ontology and entity type, and retain an explicit abstention option.
Dropping assertion context after NER
A correctly linked medication can still be clinically misleading when the note says it was discontinued or denied.
Treat assertion as part of the entity record and carry it through serialization and evaluation.
What changed in the model of the problem
Lessons
- Candidate recall is the ceiling for a reranker; retrieval diagnostics should come before model complexity.
- Span-boundary and ontology-link errors need separate labels in the error ledger.
- Character-aware lexical signals remain valuable beside multilingual embeddings.
- Abstention is a meaningful prediction when the terminology does not support a defensible link.
- Dataset, ontology, and normalization versions belong in every experiment record.
Queue, not promises
Next experiments
- 01Compare reciprocal-rank fusion with a learned lexical–dense score calibration layer.
- 02Measure candidate recall by entity type and by normalized mention frequency.
- 03Test whether assertion-aware context improves linking for medication and diagnosis mentions.
- 04Build a small adjudicated challenge slice containing abbreviations and overlapping spans.
- 05Profile the accuracy–latency trade-off for cross-encoder reranking at different candidate depths.