| |
| """Augment router weak-label dataset. |
| |
| The curated UpToDate titles yield only grounded single-intent examples. A robust |
| router must also learn: |
| (a) alert mode : chronicity / recurrence / acute / dangerous signs / population |
| (b) compound : parallel multi-sub-intent |
| This script synthesizes realistic patient utterances for these, and labels each by |
| running the actual rule engine (consistent weak supervision), asserting expected mode. |
| |
| Output : data/router/router_train.jsonl (appended) |
| Algorithmic note: we *generate* the utterances from templates + vocab, then trust |
| scan() as the labeler; we also self-check mode against the template intent and print |
| any mismatch so the manual vocabulary stays aligned with the rules. |
| """ |
| import itertools |
| import json |
| import os |
| import re |
| import sys |
|
|
| sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..")) |
| from scripts.red_flag_rules import scan |
|
|
| OUT = "/workspace/TASK18/data/router" |
| os.makedirs(OUT, exist_ok=True) |
|
|
| CAT_ENTITY = { |
| "heart-and-blood-vessel-disease": ["头痛", "头疼", "胸痛", "胸闷", "心慌", "血压高", "心跳快"], |
| "gastrointestinal-system": ["胃痛", "肚子痛", "腹痛", "反酸", "恶心", "拉肚子", "胃胀", "便秘"], |
| "diabetes": ["血糖高", "血糖低", "头晕没劲", "视力模糊", "手脚发麻"], |
| "cancer": ["淋巴结肿大", "持续低烧", "体重下降", "身上有肿块"], |
| "childrens-health": ["孩子发烧", "孩子咳嗽", "宝宝拉肚子", "孩子不愿意吃饭", "孩子一直哭"], |
| "lung-disease": ["咳喘", "气短", "咳嗽有痰", "喘不上气", "晚上憋醒"], |
| "brain-and-nerves": ["偏头痛", "头晕", "手麻", "癫痫发作", "记忆下降", "站不稳"], |
| "nervous-other": ["焦虑", "睡不好", "没精神", "情绪低落", "失眠", "担心"], |
| "senior-health": ["老人跌倒", "关节痛", "走几步就喘", "记性差", "尿频"], |
| "kidneys-and-urinary-system": ["尿频", "排尿痛", "尿量少", "下肢水肿"], |
| "women": ["月经不调", "痛经", "异常出血", "乳房胀痛", "白带异常"], |
| "muscles-joints": ["膝盖痛", "腰痛", "肩颈酸痛", "关节晨僵", "肌肉酸痛"], |
| "skin": ["长疹子", "皮肤痒", "湿疹", "红肿", "破皮渗液"], |
| "mental": ["心慌焦虑", "整夜失眠", "情绪崩溃", "坐立不安"], |
| "infections": ["嗓子疼", "发烧", "流鼻涕", "咳嗽", "头痛"], |
| } |
|
|
| |
| ALERT_TEMPLATES = [ |
| |
| ("{e}已经{time}了", ["chronicity"], "duration2w"), |
| ("{e}一直不好，{time}了", ["chronicity"], "duration2w"), |
| ("{e}反复发作{rec}次", ["chronicity"], "recurrent"), |
| |
| ("突然{e}得很厉害", ["acuteness"], "acute"), |
| ("一下子{e}，人一下子就", ["acuteness"], "acute"), |
| |
| ("{e}越来越严重", ["progression"], "worsening"), |
| ("{e}加重了，比前几天差", ["progression"], "worsening"), |
| ] |
| TIME_CHRONIC = ["两周", "一个月", "三个月", "半年", "一年", "两个星期", "一个月多"] |
| REC = ["三", "四", "五六"] |
| ACUTE_COMBOS = {"胸痛": "突发的压榨样胸痛", "头痛": "此生最痛的头痛", |
| "喘不上气": "突然剧烈喘不上气", "肚子痛": "骤然的剧痛"} |
|
|
| def fmt(t, e, time=None, rec=None): |
| return t.format(e=e, time=time, rec=rec) |
|
|
| def gen_redflag(): |
| out = [] |
| n_syn = 0 |
| for cat, ents in CAT_ENTITY.items(): |
| for e in ents: |
| if not re.search(r"孩子|宝宝", e): |
| for t, exp_types, tag in ALERT_TEMPLATES[0:3]: |
| time = TIME_CHRONIC[n_syn % len(TIME_CHRONIC)] |
| rec = REC[n_syn % len(REC)] |
| u = fmt(t, e, time=time, rec=rec) |
| red, subs, mode = scan(u) |
| if mode != "alert": |
| print(f"[mismatch] expected alert: {u} -> {mode} red={red}") |
| continue |
| out.append(build(u)) |
| n_syn += 1 |
| return out |
|
|
| ACUTE_ENT = ["胸痛", "头痛", "喘不上气", "肚子痛", "昏迷", "抽搐"] |
| def gen_acute(): |
| out = [] |
| for e in ACUTE_ENT: |
| u = ACUTE_COMBOS.get(e, f"突然{e}") |
| red, subs, mode = scan(u) |
| if mode == "alert": |
| out.append(build(u)) |
| return out |
|
|
| POP_UTT = [ |
| "我怀孕三个月了，最近{rem}", |
| "我妈妈78岁，{rem}", |
| "我孩子1岁，{rem}", |
| "我做完化疗，现在{rem}", |
| "我在吃华法林，{rem}", |
| ] |
| POP_REM = ["有点出血", "发烧", "头晕", "脸色很差"] |
| def gen_pop(): |
| out = [] |
| for u0 in POP_UTT: |
| for rem in POP_REM: |
| u = u0.format(rem=rem) |
| red, subs, mode = scan(u) |
| |
| out.append(build(u)) |
| return out |
|
|
| def gen_compound(): |
| """Parallel multi-intent: two symptom/entity asks joined.""" |
| pairs = [ |
| ("头疼", "失眠", "有没有什么办法缓解，还有睡不好"), |
| ("胃痛", "恶心", "是啥病，怎么办"), |
| ("腿肿", "走不了路", "要紧吗，怎么处理"), |
| ("脖子酸", "手麻", "正常吗，要不要看医生"), |
| ("咳嗽", "低烧", "该吃什么药"), |
| ("膝盖疼", "上下楼难受", "是关节炎吗"), |
| ] |
| out = [] |
| for a, b, c in pairs: |
| u = f"{a}和{b}，{c}" |
| _, subs, mode = scan(u) |
| out.append(build(u)) |
| return out |
|
|
| def build(u): |
| red, subs, mode = scan(u) |
| plan = {"mode": mode, "red_flags": red if mode == "alert" else [], "sub_intents": subs} |
| return { |
| "messages": [ |
| {"role": "user", "content": f"请把下面这句话拆成结构化的患者意图计划。只输出 JSON。\n患者：{u}"}, |
| {"role": "assistant", "content": json.dumps(plan, ensure_ascii=False)}, |
| ], |
| "utterance": u, |
| "augmented": True, |
| } |
|
|
|
|
| def main(): |
| recs = [] |
| recs += gen_redflag() |
| recs += gen_acute() |
| recs += gen_pop() |
| recs += gen_compound() |
| path = os.path.join(OUT, "router_augmented.jsonl") |
| with open(path, "w", encoding="utf-8") as f: |
| for r in recs: |
| f.write(json.dumps(r, ensure_ascii=False) + "\n") |
| from collections import Counter |
| modes = Counter(json.loads(r["messages"][1]["content"])["mode"] for r in recs) |
| print("augmented records:", len(recs)) |
| print("mode dist:", dict(modes)) |
| |
| import random |
| random.seed(1) |
| for r in random.sample(recs, 6): |
| print("\nU:", r["utterance"]) |
| print("P:", r["messages"][1]["content"]) |
| print("saved ->", path) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
