Multilingual Clinical Relations¶
OpenMed can extract deterministic Chinese and Hindi clinical relations from already-extracted NER spans. It does not run or train a second NER model. The candidate layer uses character offsets rather than whitespace tokens, then passes typed edges through the same constrained span-graph decoder used by the other relation backends.
from openmed.clinical import extract_relations
text = "肺炎使用阿莫西林治疗。"
spans = [
{"label": "CONDITION", "start": 0, "end": 2},
{"label": "MEDICATION", "start": 4, "end": 8},
]
relations = extract_relations(text, spans, language="zh")
print(relations[0].to_dict())
The output retains both relation arguments and their source character offsets. Relation extraction is assistive support and should be validated before clinical use.
Registry Version 1¶
MULTILINGUAL_RELATION_REGISTRY_VERSION is 1. The Chinese registry preserves all 44 distinct predicates in the official CMeIE schema and maps them onto stable lower-snake-case labels. CMeIE defines 53 subject-predicate-object schemas because some predicates allow more than one entity-type pairing. The registry includes labels and type compatibility only; OpenMed does not bundle the CMeIE corpus. See the official CBLUE benchmark for the dataset boundary and evaluation format.
| CMeIE predicate | Canonical label |
|---|---|
| 预防 | prevention |
| 内窥镜检查 | endoscopic_examination |
| 病理分型 | pathology_type |
| 病史 | history |
| 阶段 | stage |
| 筛查 | screening |
| 相关（导致） | causes |
| 遗传因素 | genetic_factor |
| 就诊科室 | care_department |
| 多发群体 | prevalent_population |
| 相关（转化） | transforms_to |
| 发病机制 | pathogenesis |
| 辅助治疗 | adjuvant_treatment |
| 发病率 | incidence |
| 相关（症状） | associated_symptom |
| 病理生理 | pathophysiology |
| 化疗 | chemotherapy |
| 发病年龄 | onset_age |
| 鉴别诊断 | differential_diagnosis |
| 药物治疗 | drug_treatment |
| 放射治疗 | radiation_treatment |
| 多发地区 | prevalent_region |
| 临床表现 | clinical_manifestation |
| 发病部位 | affected_body_site |
| 手术治疗 | surgical_treatment |
| 发病性别倾向 | sex_predisposition |
| 治疗后症状 | post_treatment_symptom |
| 转移部位 | metastasis_site |
| 实验室检查 | laboratory_test |
| 死亡率 | mortality |
| 侵及周围组织转移的症状 | tissue_invasion_symptom |
| 外侵部位 | external_invasion_site |
| 影像学检查 | imaging_test |
| 传播途径 | transmission_route |
| 病因 | etiology |
| 预后状况 | prognosis |
| 辅助检查 | auxiliary_examination |
| 多发季节 | prevalent_season |
| 高危因素 | high_risk_factor |
| 预后生存率 | survival_rate |
| 组织学检查 | histological_examination |
| 并发症 | complication |
| 风险评估因素 | risk_assessment_factor |
| 同义词 | synonym |
The Hindi Indic subset covers the relation families exercised by the shipped synthetic clinical fixtures:
| Hindi relation | Canonical label |
|---|---|
| दवा उपचार | drug_treatment |
| शल्य उपचार | surgical_treatment |
| रोकथाम | prevention |
| नैदानिक अभिव्यक्ति | clinical_manifestation |
| कारण | etiology |
| जटिलता | complication |
| प्रयोगशाला जांच | laboratory_test |
| इमेजिंग जांच | imaging_test |
Assertion And Evaluation Behavior¶
Chinese extraction uses the Chinese ConText cue pack and Hindi extraction uses the Hindi pack. Missing cues default to recent, certain, and affirmed; an English-looking record marker does not negate a Chinese or Hindi relation. Refuted and conditional relations are withheld by default, while uncertain relations remain available as possible for review.
The synthetic fixtures live at:
openmed/eval/golden/fixtures/i18n/relations_zh.jsonlopenmed/eval/golden/fixtures/i18n/relations_indic.jsonl
run_relation_benchmark() and ModelScorecard report strict and relaxed F1 under relation_extraction.per_language, so Chinese and Hindi performance cannot be hidden inside an aggregate relation score.