Chinese terminology grounding¶

OpenMed can ground Chinese CONDITION and MEDICATION spans against a dictionary that you supply. Matching is local and deterministic: exact matches run first, followed by an optional normalized fuzzy match. The loader never downloads, scrapes, or silently substitutes a Chinese ICD-10 or drug vocabulary.

Not a medical device

Chinese terminology grounding is assistive software, not a medical device. Codes are user-vocabulary-dependent and are not for clinical decisions. Verify every match against your source dictionary and its license.

Dictionary schema¶

Provide one UTF-8 CSV, TSV, or JSONL file with these fields:

Field	Required	Value
`label`	yes	`CONDITION` or `MEDICATION`
`system`	no	`ICD-10-CN` for conditions or `CN-DRUG` for medications
`code`	yes	Code from your licensed/user-managed vocabulary
`display`	yes	Preferred Chinese display term
`aliases`	no	`\|`-separated text in CSV/TSV, or a string array in JSONL

The repository's examples/data/chinese_terminology_synthetic.csv file contains twelve invented entries for tests and documentation. Its SYN-CN-* codes and terms are not a real clinical vocabulary.

Ground an analysis result¶

from openmed import analyze_text
from openmed.clinical import ChineseTerminologyGrounder

analysis = analyze_text(
    "诊断星芒性热病，予晨露安片。",
    model_name="your-chinese-clinical-ner-model",
    sentence_language="zh",
)

grounder = ChineseTerminologyGrounder.from_path(
    "/secure/local/path/chinese-terminology.csv",
    license_acknowledged=True,
)
grounded = grounder.ground_result(analysis)

for entity in grounded.entities:
    print(entity.text, entity.metadata)

The license acknowledgement is mandatory. Omitting the path raises an error that explains that OpenMed has no bundled fallback. Set fuzzy=False when only exact matches are acceptable, or adjust min_score from its conservative default of 0.8 for a vocabulary-specific validation study.

Dictionary values can be emitted in grounding metadata. Supply terminology only: do not put patient names, identifiers, or other PHI in code, display, or aliases fields.

Matched spans gain only system, code, and display metadata. Their text, start/end offsets, confidence, and any de-identification action remain unchanged. Result-level metadata emits the medical-device disclaimer and the user-vocabulary note without recording dictionary contents or source PHI.