@furongjun1999/dsh-memory 0.4.11 → 0.5.0

package/md_cg/semantic/en_normalizer.py

@@ -1,364 +1,364 @@
-#!/usr/bin/env python3
-# -*- coding: utf-8 -*-
-"""英文查询归一器：English query → 语义原子序列（中文语素）→ 检索键
-
-原则（设计者裁定）：
-  · 时态全部抛弃，时间由时间词承载
-  · 无联系虚词（冠词/介词/系动词）不入语义序列
-  · 音译/专有词：保留原名（不归一化），标记为 proper_noun
-  · 构词法：英文复合词按中文语素序拆解重组（beef → cow_meat → 牛 肉）
-
-用法：
-    from en_normalizer import normalize
-    terms = normalize("I eat beef yesterday")
-    # → ["我", "吃", "牛肉", "昨天"]  (用于 cg.search)
-"""
-import json, io, os, re
-
-# ---- 时态还原表：英文屈折 → 基形 ----
-IRREGULAR = {
-    "ate": "eat", "eaten": "eat", "went": "go", "gone": "go",
-    "saw": "see", "seen": "see", "wrote": "write", "written": "write",
-    "took": "take", "taken": "take", "made": "make", "ran": "run",
-    "bought": "buy", "brought": "bring", "thought": "think",
-    "taught": "teach", "caught": "catch", "sought": "seek",
-    "fought": "fight", "sold": "sell", "told": "tell", "felt": "feel",
-    "fell": "fall", "sent": "send", "spent": "spend", "built": "build",
-    "lost": "lose", "met": "meet", "paid": "pay", "led": "lead",
-    "won": "win", "sat": "sit", "stood": "stand",
-    "understood": "understand", "heard": "hear",
-    "spoke": "speak", "spoken": "speak", "broke": "break", "broken": "break",
-    "chose": "choose", "chosen": "choose", "drew": "draw", "drawn": "draw",
-    "drove": "drive", "driven": "drive", "grew": "grow", "grown": "grow",
-    "knew": "know", "known": "know", "gave": "give", "given": "give",
-    "was": "be", "were": "be", "been": "be", "had": "have", "has": "have",
-    "did": "do", "done": "do", "said": "say", "got": "get",
-    "left": "leave", "kept": "keep", "held": "hold",
-    "slept": "sleep", "swept": "sweep", "meant": "mean",
-    "dealt": "deal", "lent": "lend", "bent": "bend",
-    # e-脱落动词机械去 -ed 会得错形（inspired→inspir）；词次取证达标者逐词收编
-    "inspired": "inspire",
-}
-
-# ---- 虚词（不入语义序列）----
-STOPWORDS = {
-    "a", "an", "the", "of", "to", "in", "on", "at", "for", "with",
-    "by", "from", "up", "about", "into", "over", "after", "is", "are",
-    "was", "were", "be", "been", "being", "do", "does", "did", "have",
-    "has", "had", "will", "would", "could", "should", "may", "might",
-    "and", "or", "but", "not", "no", "so", "if", "then", "than",
-    "this", "that", "these", "those", "it", "its", "as", "also",
-    # locomo-500 词频取证补充（expand_en_zh analyze，2026-09-14）：物主/宾格
-    # 代词与疑问框架词对中文库检索零语义贡献，保留只制造噪声原子
-    # （unknown_keep 头部：his 90 / her 72 / get 20 / during 19 词次）。
-    # 主格代词不动（i/we/you/he/she/they 已映射 我/我们/你/他/她/他们）。
-    "his", "her", "hers", "him", "their", "theirs", "them", "us",
-    "when", "where", "why", "how", "what", "which", "who", "whose",
-    "during", "out", "get", "kind",
-}
-
-# ---- en→zh 语义原子映射 ----
-# 构词法：英文语素 → 中文语素（与 standard_en.json 的 morphemes 表互逆）
-# 同时覆盖：常用动词基形 / 名词 / 时间词 / 形容词
-EN_ZH = {
-    # 代词
-    "i": "我", "me": "我", "my": "我", "you": "你", "he": "他", "she": "她",
-    "we": "我们", "they": "他们",
-    # 动作
-    "eat": "吃", "go": "去", "come": "来", "see": "看", "hear": "听",
-    "read": "读", "write": "写", "walk": "走", "run": "跑", "fly": "飞",
-    "make": "做", "build": "建", "find": "找", "give": "给", "take": "拿",
-    "know": "知道", "think": "想", "say": "说", "speak": "说",
-    "love": "爱", "use": "用", "buy": "买", "sell": "卖",
-    "teach": "教", "learn": "学", "study": "学", "work": "工作",
-    "sleep": "睡", "open": "开", "close": "关", "drive": "驾驶",
-    # 名词
-    "book": "书", "water": "水", "fire": "火", "meat": "肉",
-    "milk": "奶", "egg": "蛋", "road": "路", "bridge": "桥",
-    "tree": "木", "star": "星", "moon": "月", "sun": "日",
-    "sea": "海", "mountain": "山", "city": "城", "country": "国",
-    "people": "人", "person": "人", "king": "王", "heart": "心",
-    "hand": "手", "head": "头", "eye": "眼", "blood": "血",
-    "medicine": "医", "drug": "药", "disease": "病",
-    "knowledge": "知识", "work": "工作", "discipline": "纪律",
-    "memory": "记忆", "cognition": "认知", "graph": "图",
-    "dog": "狗", "cat": "猫", "bird": "鸟", "fish": "鱼",
-    "cow": "牛", "pig": "猪", "sheep": "羊", "chicken": "鸡",
-    # 盲测缺口修补（bench_blind_comp L2 白箱探针锁定，2026-09-14）：
-    # horse/mare/goat/duck/oil/fat 六词未映射 → 英文组合词展开缺语素。
-    # 只补探针暴露的缺口，不顺手扩表——修复面与暴露面精确对齐，净效应可归因。
-    "horse": "马", "mare": "马", "goat": "羊", "duck": "鸭",
-    "oil": "油", "fat": "油",
-    # 形容词
-    "big": "大", "small": "小", "high": "高", "low": "低",
-    "new": "新", "old": "旧", "good": "好", "bad": "坏",
-    "hot": "热", "cold": "冷", "fast": "快", "slow": "慢",
-    # 时间词
-    "yesterday": "昨天", "today": "今天", "tomorrow": "明天",
-    "morning": "早上", "evening": "晚上", "night": "夜",
-    "spring": "春天", "summer": "夏天", "autumn": "秋天", "winter": "冬天",
-    "year": "年", "month": "月", "day": "天", "week": "周",
-    "hour": "小时", "minute": "分", "second": "秒",
-    # 数
-    "one": "一", "two": "二", "three": "三", "four": "四", "five": "五",
-    "six": "六", "seven": "七", "eight": "八", "nine": "九", "ten": "十",
-    "hundred": "百", "thousand": "千",
-    # 其他
-    "water": "水", "light": "光", "sound": "音", "metal": "金",
-    "stone": "石", "wood": "木", "cloud": "云", "wind": "风",
-    "rain": "雨", "snow": "雪",
-    # locomo-500 词频取证定向补词（expand_en_zh analyze top30 实词，2026-09-14；
-    # 映射逐一核对唯一；partner 搭档/伴侣一对多跳过——宁缺勿滥）
-    "friend": "朋友", "family": "家庭", "trip": "旅行", "plan": "计划",
-    "game": "游戏", "type": "类型", "dance": "舞", "share": "分享",
-    "painting": "画", "favorite": "喜欢", "feel": "感觉",
-    "activity": "活动", "project": "项目", "photo": "照片",
-    "studio": "工作室", "long": "长", "first": "第一", "recently": "最近",
-    "start": "开始", "inspire": "激励",
-    # CEDICT 词汇分布缺口定向补（2026-09-14）：CEDICT 用英式 mum（mom 缺）、
-    # pet 冲突集全书面词在口语语料零命中、日常义项缺词条——映射唯一明确的
-    # top30 级残留词逐词补，修复面=证据面
-    "pet": "宠物", "mom": "妈妈", "tournament": "锦标赛",
-    "festival": "节日", "advice": "建议", "often": "经常",
-    "pottery": "陶艺",
-    # 第三方 LoCoMo 报告错译暴露面语境化修正（2026-09-15，修复面=暴露面对齐）：
-    # scared 错切链路 = strip_tense(scared)→scar × CEDICT scar→创痕，复合错译
-    # （应为 害怕）；shelter/tough/times 为 CEDICT 第一义项与口语语料语境失配
-    # （CEDICT 庇护/厉害/时间 → 语境 收容所/坚强/次）。低置信标记
-    # （source=cedict）对此全部捕获后，人工校对层逐一核对的落地
-    "scared": "害怕", "shelter": "收容所", "tough": "坚强", "times": "次",
-}
-
-# ---- 复合词映射（英文复合 → 中文标准概念）----
-COMPOUND_ZH = {
-    "soul hub": "灵枢",
-    "beef": "牛肉", "pork": "猪肉", "mutton": "羊肉", "chicken": "鸡肉",
-    "computer": "电脑", "telephone": "电话", "television": "电视",
-    "movie": "电影", "film": "电影", "battery": "电池",
-    "train": "火车", "car": "汽车", "airplane": "飞机", "mobile": "手机",
-    "volcano": "火山", "seawater": "海水", "moonlight": "月光",
-    "monday": "星期一", "tuesday": "星期二", "wednesday": "星期三",
-    "thursday": "星期四", "friday": "星期五", "saturday": "星期六",
-    "sunday": "星期日",
-    "mathematics": "数学", "chemistry": "化学", "physics": "物理学",
-    "geography": "地理学", "biology": "生物学", "thermodynamics": "热力学",
-    "student": "学生", "doctor": "医生", "lawyer": "律师", "worker": "工人",
-    "knowledge": "知识", "graph": "图",
-    # 短语映射
-    "electric brain": "电脑", "soul hub": "灵枢", "memory store": "记忆库",
-    "fire vehicle": "火车", "cow meat": "牛肉", "gas vehicle": "汽车",
-    "work discipline": "工作纪律", "knowledge graph": "知识图谱",
-    "cognition graph": "认知图", "memory store": "记忆库",
-    # 第三方 LoCoMo 报告错译暴露面（2026-09-15）：scavenger hunt 逐词直译
-    # → 食腐动物 猎取（应为 寻宝游戏）；three times a week → 三 时间 周
-    # （应为 每周三次）。配合内嵌短语扫描机制在长文本/问句内部生效
-    "scavenger hunt": "寻宝游戏", "three times a week": "每周三次",
-    "word": "字", "soul": "灵", "hub": "枢",
-}
-
-# 含空格多词短语预计算（内嵌短语扫描用，长键优先防前缀吞并）
-COMPOUND_ZH_PHRASES = sorted(
-    ((k, v) for k, v in COMPOUND_ZH.items() if " " in k),
-    key=lambda kv: -len(kv[0]))
-
-
-# 生效条件：w 长度 >2 且末字符与倒数第二字符相同且末字符不属于 "aeiousl" 时返回 w[:-1]，否则原样返回 w。
-def _de_double(w):
-    """双写辅音还原（CVC 动词屈折）：regrett→regret / runn→run / stopp→stop。
-
-    s / l 结尾排除（正字法惯例保留）：crossed→cross、called→call。
-    """
-    if len(w) > 2 and w[-1] == w[-2] and w[-1] not in "aeiousl":
-        return w[:-1]
-    return w
-
-
-# 生效条件：word 小写后命中 IRREGULAR 键则返回该表值；否则按序判定小写形——以 "ed" 结尾且长度 >4 返回 _de_double(w[:-2])、以 "ing" 结尾且长度 >5 返回 _de_double(w[:-3])、以 "s" 结尾且不以 "ss" 结尾返回 w[:-1]、其余返回该小写形 w。
-def strip_tense(word):
-    """英文屈折归零：不规则动词查表；规则动词去 -ed/-ing/-s + 双写辅音还原"""
-    w = word.lower()
-    if w in IRREGULAR:
-        return IRREGULAR[w]
-    if w.endswith("ed") and len(w) > 4:
-        return _de_double(w[:-2])
-    if w.endswith("ing") and len(w) > 5:
-        return _de_double(w[:-3])
-    if w.endswith("s") and not w.endswith("ss"):
-        return w[:-1]
-    return w
-
-
-# 生效条件：w.lower() 属于模块级 STOPWORDS 时返回 True，否则返回 False（w 为空串时 lower 为空串，返回 False）。
-def is_stopword(w):
-    return w.lower() in STOPWORDS
-
-
-# 生效条件：w 为真值（非空串）时返回 w[0].isupper()，w 为假值时返回 False。
-def is_proper(w):
-    """专有词：原始 query 中首字母大写（人名/品牌/地名不归一化）"""
-    return w[0].isupper() if w else False
-
-
-_CEDICT_CACHE = None
-
-
-# 生效条件：模块级 _CEDICT_CACHE 为 None 时读取 lexicon/cedict_en_zh.json 并取 json 的 "map" 键（缺该键回落 {}）写入缓存，遇 OSError/ValueError 写入 {}；缓存非 None 时直接返回 _CEDICT_CACHE。
-def cedict_map():
-    """词级 CEDICT 反查表（build_cedict_en_zh.py 产物，17700 键级）。
-
-    CC BY-SA 4.0 派生数据独立文件署名，不内联 MIT 的 EN_ZH；
-    缺文件返回空表——纯手工表兜底，零外部依赖路径保持（第7条兜底纪律）。
-    """
-    global _CEDICT_CACHE
-    if _CEDICT_CACHE is None:
-        import json
-        import os
-        path = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
-                            "lexicon", "cedict_en_zh.json")
-        try:
-            with io.open(path, encoding="utf-8") as f:
-                _CEDICT_CACHE = json.load(f).get("map", {})
-        except (OSError, ValueError):
-            _CEDICT_CACHE = {}
-    return _CEDICT_CACHE
-
-
-# 生效条件：extra_map 为真值（非 None 且非空映射）时链首加入 (extra_map, "extra")，(EN_ZH, "manual") 恒加入，cedict_map() 返回非空映射时追加 (cedict, "cedict")，最后返回该链 chain。
-def _lookup_chain(extra_map=None):
-    """置信分层查表链（第三方验证 2026-09-15）：extra（实验注入）→
-    EN_ZH（人工逐词校对）→ CEDICT（机械反查，第一义项直译）。
-
-    分层依据 = 词条来源的人工校对程度，即映射质量的确定性代理信号：
-    CEDICT 层命中即 low_confidence（语境失配风险，实证 scavenger→食腐动物，
-    见 docs/eval/第三方验证报告_LoCoMo_灵枢_.md 发现 5）。
-    """
-    chain = []
-    if extra_map:
-        chain.append((extra_map, "extra"))
-    chain.append((EN_ZH, "manual"))
-    cedict = cedict_map()
-    if cedict:
-        chain.append((cedict, "cedict"))
-    return chain
-
-
-# 生效条件：query 为字符串，query.lower().strip() 命中 COMPOUND_ZH 即返回单元素 [comp_zh] 与 phrase_mapped 记录；否则经 COMPOUND_ZH_PHRASES 内嵌替换、剥除 's 与分词后逐词映射，返回 terms 与 inline_hits + detail；
-def normalize_en_query(query, extra_map=None):
-    """英文 query → 语义原子序列（中文语素）
-
-    返回 (normalized_terms, detail)：
-      normalized_terms: 用于检索的中文/保留词序列
-      detail: 逐词归一化记录（mapped 词带 source/low_confidence 置信信号；
-              low_confidence=True = 机械反查未人工校对，语境失配风险）
-    """
-    chain = _lookup_chain(extra_map)
-    # 短语优先匹配：先尝试多词短语整体映射（复合概念名）
-    phrase_key = query.lower().strip()
-    if phrase_key in COMPOUND_ZH:
-        comp_zh = COMPOUND_ZH[phrase_key]
-        return [comp_zh], [{"orig": query, "phrase_zh": comp_zh,
-                            "action": "phrase_mapped"}]
-    # 内嵌短语扫描：多词短语在长文本/问句内部出现时整体替换为中文
-    # （整体 query 匹配只救短语独占 query 的形态；doc 侧英文原文与含修饰语
-    # 的问句靠此层，否则逐词直译拆散复合语义——第三方 LoCoMo 报告错译样本
-    # scavenger hunt → 食腐动物 猎取 即此缺口。短语表键 ≤11 个，逐键扫描
-    # 成本可忽略）。中文替换段经 findall 整体成原子，后续逐词链路不受影响
-    inline_hits = []
-    for phrase, pzh in COMPOUND_ZH_PHRASES:
-        if phrase in query.lower():
-            query = re.sub(re.escape(phrase), pzh, query, flags=re.IGNORECASE)
-            inline_hits.append({"orig": phrase, "phrase_zh": pzh,
-                                "action": "phrase_inline"})
-    # 屈折还原 + 查表。所有格剥离：'s 是正字法黏着成分非独立词
-    # （Melanie's → Melanie）——不剥离则分词残留 "s" 成为伪 OOV
-    # （locomo-500 实测 85 词次）。
-    words = re.findall(r"[A-Za-z\u4e00-\u9fff]+",
-                       re.sub(r"'s\b", "", query))
-    terms = []
-    detail = []
-    for w in words:
-        wl = w.lower()
-        base0 = strip_tense(wl)
-        # 原形与还原形都查虚词表：getting/doing 屈折形原表漏网（还原后
-        # get/do 是虚词，保留只产生 OOV 噪声原子）
-        if wl in STOPWORDS or base0 in STOPWORDS:
-            detail.append({"orig": w, "action": "stopword_drop"})
-            continue
-        base = base0
-        zh = src = None
-        for m, s in chain:
-            z = m.get(base) or m.get(wl)
-            if z:
-                zh, src = z, s
-                break
-            if not base.endswith("e"):
-                # e-脱落动词词表感知还原：loved→lov(错形)→love。词表小时收益≈0
-                # （2026-09-14 早前取证 102 词次判不修）；词级表 17700 键后
-                # motivated→motivate / visited 类命中真实存在，条件已变
-                z = m.get(base + "e")
-                if z:
-                    base = base + "e"
-                    zh, src = z, s
-                    break
-        if zh and is_proper(w) and wl not in EN_ZH:
-            # 专名词表命中双原子：doc 侧音译/原文两形态并存（corpus567 实测
-            # 地名 巴黎13/Paris4 音译主导，人名 Caroline127/卡罗琳0 原文主导）
-            # ——原文+译文都进匹配面；未命中侧在 doc 侧零出现，组合共现需
-            # 双方在场，不构成假匹配。手工表已有词（i/we/you 等句首大写代词、
-            # 基础词）形态唯一，不具双形态不确定性，排除
-            terms.append(w)
-            terms.append(zh)
-            detail.append({"orig": w, "base": base, "zh": zh, "source": src,
-                           "low_confidence": src == "cedict",
-                           "action": "proper_mapped_both"})
-        elif zh:
-            terms.append(zh)
-            detail.append({"orig": w, "base": base, "zh": zh, "source": src,
-                           "low_confidence": src == "cedict",
-                           "action": "mapped"})
-        elif is_proper(w):
-            terms.append(w)  # 专有词保留原名
-            detail.append({"orig": w, "action": "proper_noun_keep"})
-        else:
-            # 未知词：尝试复合拆解（compound→zh morphemes）
-            comp = COMPOUND_ZH.get(wl) or COMPOUND_ZH.get(base)
-            if comp:
-                terms.append(comp)
-                detail.append({"orig": w, "compound_zh": comp, "action": "compound_mapped"})
-            else:
-                terms.append(w)
-                detail.append({"orig": w, "action": "unknown_keep"})
-    return terms, inline_hits + detail
-
-
-# 生效条件：detail 中某项 d 的 d.get("low_confidence") 为真值时，取 (d.get("orig") or "").lower()，仅当结果非空且尚未出现在 out 中才按序追加，最终返回去重保序的 out。
-def low_confidence_terms(detail):
-    """从归一化 detail 提取低置信映射词清单（去重保序）。
-
-    低置信 = mapped 词来自 CEDICT 机械反查（第一义项直译，未经人工
-    语境校对）。用途：错译审计靶子 / 词表人工校对优先级（警告不拒绝，
-    与 semantic_oov 同哲学——标记是观测依据，不改检索行为）。
-    """
-    out = []
-    for d in detail:
-        if d.get("low_confidence"):
-            w = (d.get("orig") or "").lower()
-            if w and w not in out:
-                out.append(w)
-    return out
-
-
-# 生效条件：len(sys.argv) > 1 时 query 为 sys.argv[1:] 以空格连接，否则 query 为字面默认 "I eat beef yesterday"；随后打印归一化 terms 与逐条 detail，无返回值。
-def main():
-    import sys
-    query = " ".join(sys.argv[1:]) if len(sys.argv) > 1 else "I eat beef yesterday"
-    terms, detail = normalize_en_query(query)
-    print("query:", query)
-    print("normalized terms:", terms)
-    print("search string:", " ".join(terms))
-    for d in detail:
-        print(" ", json.dumps(d, ensure_ascii=False))
-
-
-if __name__ == "__main__":
-    main()
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""英文查询归一器：English query → 语义原子序列（中文语素）→ 检索键
+
+原则（设计者裁定）：
+  · 时态全部抛弃，时间由时间词承载
+  · 无联系虚词（冠词/介词/系动词）不入语义序列
+  · 音译/专有词：保留原名（不归一化），标记为 proper_noun
+  · 构词法：英文复合词按中文语素序拆解重组（beef → cow_meat → 牛 肉）
+
+用法：
+    from en_normalizer import normalize
+    terms = normalize("I eat beef yesterday")
+    # → ["我", "吃", "牛肉", "昨天"]  (用于 cg.search)
+"""
+import json, io, os, re
+
+# ---- 时态还原表：英文屈折 → 基形 ----
+IRREGULAR = {
+    "ate": "eat", "eaten": "eat", "went": "go", "gone": "go",
+    "saw": "see", "seen": "see", "wrote": "write", "written": "write",
+    "took": "take", "taken": "take", "made": "make", "ran": "run",
+    "bought": "buy", "brought": "bring", "thought": "think",
+    "taught": "teach", "caught": "catch", "sought": "seek",
+    "fought": "fight", "sold": "sell", "told": "tell", "felt": "feel",
+    "fell": "fall", "sent": "send", "spent": "spend", "built": "build",
+    "lost": "lose", "met": "meet", "paid": "pay", "led": "lead",
+    "won": "win", "sat": "sit", "stood": "stand",
+    "understood": "understand", "heard": "hear",
+    "spoke": "speak", "spoken": "speak", "broke": "break", "broken": "break",
+    "chose": "choose", "chosen": "choose", "drew": "draw", "drawn": "draw",
+    "drove": "drive", "driven": "drive", "grew": "grow", "grown": "grow",
+    "knew": "know", "known": "know", "gave": "give", "given": "give",
+    "was": "be", "were": "be", "been": "be", "had": "have", "has": "have",
+    "did": "do", "done": "do", "said": "say", "got": "get",
+    "left": "leave", "kept": "keep", "held": "hold",
+    "slept": "sleep", "swept": "sweep", "meant": "mean",
+    "dealt": "deal", "lent": "lend", "bent": "bend",
+    # e-脱落动词机械去 -ed 会得错形（inspired→inspir）；词次取证达标者逐词收编
+    "inspired": "inspire",
+}
+
+# ---- 虚词（不入语义序列）----
+STOPWORDS = {
+    "a", "an", "the", "of", "to", "in", "on", "at", "for", "with",
+    "by", "from", "up", "about", "into", "over", "after", "is", "are",
+    "was", "were", "be", "been", "being", "do", "does", "did", "have",
+    "has", "had", "will", "would", "could", "should", "may", "might",
+    "and", "or", "but", "not", "no", "so", "if", "then", "than",
+    "this", "that", "these", "those", "it", "its", "as", "also",
+    # locomo-500 词频取证补充（expand_en_zh analyze，2026-09-14）：物主/宾格
+    # 代词与疑问框架词对中文库检索零语义贡献，保留只制造噪声原子
+    # （unknown_keep 头部：his 90 / her 72 / get 20 / during 19 词次）。
+    # 主格代词不动（i/we/you/he/she/they 已映射 我/我们/你/他/她/他们）。
+    "his", "her", "hers", "him", "their", "theirs", "them", "us",
+    "when", "where", "why", "how", "what", "which", "who", "whose",
+    "during", "out", "get", "kind",
+}
+
+# ---- en→zh 语义原子映射 ----
+# 构词法：英文语素 → 中文语素（与 standard_en.json 的 morphemes 表互逆）
+# 同时覆盖：常用动词基形 / 名词 / 时间词 / 形容词
+EN_ZH = {
+    # 代词
+    "i": "我", "me": "我", "my": "我", "you": "你", "he": "他", "she": "她",
+    "we": "我们", "they": "他们",
+    # 动作
+    "eat": "吃", "go": "去", "come": "来", "see": "看", "hear": "听",
+    "read": "读", "write": "写", "walk": "走", "run": "跑", "fly": "飞",
+    "make": "做", "build": "建", "find": "找", "give": "给", "take": "拿",
+    "know": "知道", "think": "想", "say": "说", "speak": "说",
+    "love": "爱", "use": "用", "buy": "买", "sell": "卖",
+    "teach": "教", "learn": "学", "study": "学", "work": "工作",
+    "sleep": "睡", "open": "开", "close": "关", "drive": "驾驶",
+    # 名词
+    "book": "书", "water": "水", "fire": "火", "meat": "肉",
+    "milk": "奶", "egg": "蛋", "road": "路", "bridge": "桥",
+    "tree": "木", "star": "星", "moon": "月", "sun": "日",
+    "sea": "海", "mountain": "山", "city": "城", "country": "国",
+    "people": "人", "person": "人", "king": "王", "heart": "心",
+    "hand": "手", "head": "头", "eye": "眼", "blood": "血",
+    "medicine": "医", "drug": "药", "disease": "病",
+    "knowledge": "知识", "work": "工作", "discipline": "纪律",
+    "memory": "记忆", "cognition": "认知", "graph": "图",
+    "dog": "狗", "cat": "猫", "bird": "鸟", "fish": "鱼",
+    "cow": "牛", "pig": "猪", "sheep": "羊", "chicken": "鸡",
+    # 盲测缺口修补（bench_blind_comp L2 白箱探针锁定，2026-09-14）：
+    # horse/mare/goat/duck/oil/fat 六词未映射 → 英文组合词展开缺语素。
+    # 只补探针暴露的缺口，不顺手扩表——修复面与暴露面精确对齐，净效应可归因。
+    "horse": "马", "mare": "马", "goat": "羊", "duck": "鸭",
+    "oil": "油", "fat": "油",
+    # 形容词
+    "big": "大", "small": "小", "high": "高", "low": "低",
+    "new": "新", "old": "旧", "good": "好", "bad": "坏",
+    "hot": "热", "cold": "冷", "fast": "快", "slow": "慢",
+    # 时间词
+    "yesterday": "昨天", "today": "今天", "tomorrow": "明天",
+    "morning": "早上", "evening": "晚上", "night": "夜",
+    "spring": "春天", "summer": "夏天", "autumn": "秋天", "winter": "冬天",
+    "year": "年", "month": "月", "day": "天", "week": "周",
+    "hour": "小时", "minute": "分", "second": "秒",
+    # 数
+    "one": "一", "two": "二", "three": "三", "four": "四", "five": "五",
+    "six": "六", "seven": "七", "eight": "八", "nine": "九", "ten": "十",
+    "hundred": "百", "thousand": "千",
+    # 其他
+    "water": "水", "light": "光", "sound": "音", "metal": "金",
+    "stone": "石", "wood": "木", "cloud": "云", "wind": "风",
+    "rain": "雨", "snow": "雪",
+    # locomo-500 词频取证定向补词（expand_en_zh analyze top30 实词，2026-09-14；
+    # 映射逐一核对唯一；partner 搭档/伴侣一对多跳过——宁缺勿滥）
+    "friend": "朋友", "family": "家庭", "trip": "旅行", "plan": "计划",
+    "game": "游戏", "type": "类型", "dance": "舞", "share": "分享",
+    "painting": "画", "favorite": "喜欢", "feel": "感觉",
+    "activity": "活动", "project": "项目", "photo": "照片",
+    "studio": "工作室", "long": "长", "first": "第一", "recently": "最近",
+    "start": "开始", "inspire": "激励",
+    # CEDICT 词汇分布缺口定向补（2026-09-14）：CEDICT 用英式 mum（mom 缺）、
+    # pet 冲突集全书面词在口语语料零命中、日常义项缺词条——映射唯一明确的
+    # top30 级残留词逐词补，修复面=证据面
+    "pet": "宠物", "mom": "妈妈", "tournament": "锦标赛",
+    "festival": "节日", "advice": "建议", "often": "经常",
+    "pottery": "陶艺",
+    # 第三方 LoCoMo 报告错译暴露面语境化修正（2026-09-15，修复面=暴露面对齐）：
+    # scared 错切链路 = strip_tense(scared)→scar × CEDICT scar→创痕，复合错译
+    # （应为 害怕）；shelter/tough/times 为 CEDICT 第一义项与口语语料语境失配
+    # （CEDICT 庇护/厉害/时间 → 语境 收容所/坚强/次）。低置信标记
+    # （source=cedict）对此全部捕获后，人工校对层逐一核对的落地
+    "scared": "害怕", "shelter": "收容所", "tough": "坚强", "times": "次",
+}
+
+# ---- 复合词映射（英文复合 → 中文标准概念）----
+COMPOUND_ZH = {
+    "soul hub": "灵枢",
+    "beef": "牛肉", "pork": "猪肉", "mutton": "羊肉", "chicken": "鸡肉",
+    "computer": "电脑", "telephone": "电话", "television": "电视",
+    "movie": "电影", "film": "电影", "battery": "电池",
+    "train": "火车", "car": "汽车", "airplane": "飞机", "mobile": "手机",
+    "volcano": "火山", "seawater": "海水", "moonlight": "月光",
+    "monday": "星期一", "tuesday": "星期二", "wednesday": "星期三",
+    "thursday": "星期四", "friday": "星期五", "saturday": "星期六",
+    "sunday": "星期日",
+    "mathematics": "数学", "chemistry": "化学", "physics": "物理学",
+    "geography": "地理学", "biology": "生物学", "thermodynamics": "热力学",
+    "student": "学生", "doctor": "医生", "lawyer": "律师", "worker": "工人",
+    "knowledge": "知识", "graph": "图",
+    # 短语映射
+    "electric brain": "电脑", "soul hub": "灵枢", "memory store": "记忆库",
+    "fire vehicle": "火车", "cow meat": "牛肉", "gas vehicle": "汽车",
+    "work discipline": "工作纪律", "knowledge graph": "知识图谱",
+    "cognition graph": "认知图", "memory store": "记忆库",
+    # 第三方 LoCoMo 报告错译暴露面（2026-09-15）：scavenger hunt 逐词直译
+    # → 食腐动物 猎取（应为 寻宝游戏）；three times a week → 三 时间 周
+    # （应为 每周三次）。配合内嵌短语扫描机制在长文本/问句内部生效
+    "scavenger hunt": "寻宝游戏", "three times a week": "每周三次",
+    "word": "字", "soul": "灵", "hub": "枢",
+}
+
+# 含空格多词短语预计算（内嵌短语扫描用，长键优先防前缀吞并）
+COMPOUND_ZH_PHRASES = sorted(
+    ((k, v) for k, v in COMPOUND_ZH.items() if " " in k),
+    key=lambda kv: -len(kv[0]))
+
+
+# 生效条件：w 长度 >2 且末字符与倒数第二字符相同且末字符不属于 "aeiousl" 时返回 w[:-1]，否则原样返回 w。
+def _de_double(w):
+    """双写辅音还原（CVC 动词屈折）：regrett→regret / runn→run / stopp→stop。
+
+    s / l 结尾排除（正字法惯例保留）：crossed→cross、called→call。
+    """
+    if len(w) > 2 and w[-1] == w[-2] and w[-1] not in "aeiousl":
+        return w[:-1]
+    return w
+
+
+# 生效条件：word 小写后命中 IRREGULAR 键则返回该表值；否则按序判定小写形——以 "ed" 结尾且长度 >4 返回 _de_double(w[:-2])、以 "ing" 结尾且长度 >5 返回 _de_double(w[:-3])、以 "s" 结尾且不以 "ss" 结尾返回 w[:-1]、其余返回该小写形 w。
+def strip_tense(word):
+    """英文屈折归零：不规则动词查表；规则动词去 -ed/-ing/-s + 双写辅音还原"""
+    w = word.lower()
+    if w in IRREGULAR:
+        return IRREGULAR[w]
+    if w.endswith("ed") and len(w) > 4:
+        return _de_double(w[:-2])
+    if w.endswith("ing") and len(w) > 5:
+        return _de_double(w[:-3])
+    if w.endswith("s") and not w.endswith("ss"):
+        return w[:-1]
+    return w
+
+
+# 生效条件：w.lower() 属于模块级 STOPWORDS 时返回 True，否则返回 False（w 为空串时 lower 为空串，返回 False）。
+def is_stopword(w):
+    return w.lower() in STOPWORDS
+
+
+# 生效条件：w 为真值（非空串）时返回 w[0].isupper()，w 为假值时返回 False。
+def is_proper(w):
+    """专有词：原始 query 中首字母大写（人名/品牌/地名不归一化）"""
+    return w[0].isupper() if w else False
+
+
+_CEDICT_CACHE = None
+
+
+# 生效条件：模块级 _CEDICT_CACHE 为 None 时读取 lexicon/cedict_en_zh.json 并取 json 的 "map" 键（缺该键回落 {}）写入缓存，遇 OSError/ValueError 写入 {}；缓存非 None 时直接返回 _CEDICT_CACHE。
+def cedict_map():
+    """词级 CEDICT 反查表（build_cedict_en_zh.py 产物，17700 键级）。
+
+    CC BY-SA 4.0 派生数据独立文件署名，不内联 MIT 的 EN_ZH；
+    缺文件返回空表——纯手工表兜底，零外部依赖路径保持（第7条兜底纪律）。
+    """
+    global _CEDICT_CACHE
+    if _CEDICT_CACHE is None:
+        import json
+        import os
+        path = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
+                            "lexicon", "cedict_en_zh.json")
+        try:
+            with io.open(path, encoding="utf-8") as f:
+                _CEDICT_CACHE = json.load(f).get("map", {})
+        except (OSError, ValueError):
+            _CEDICT_CACHE = {}
+    return _CEDICT_CACHE
+
+
+# 生效条件：extra_map 为真值（非 None 且非空映射）时链首加入 (extra_map, "extra")，(EN_ZH, "manual") 恒加入，cedict_map() 返回非空映射时追加 (cedict, "cedict")，最后返回该链 chain。
+def _lookup_chain(extra_map=None):
+    """置信分层查表链（第三方验证 2026-09-15）：extra（实验注入）→
+    EN_ZH（人工逐词校对）→ CEDICT（机械反查，第一义项直译）。
+
+    分层依据 = 词条来源的人工校对程度，即映射质量的确定性代理信号：
+    CEDICT 层命中即 low_confidence（语境失配风险，实证 scavenger→食腐动物，
+    见 docs/eval/第三方验证报告_LoCoMo_灵枢_.md 发现 5）。
+    """
+    chain = []
+    if extra_map:
+        chain.append((extra_map, "extra"))
+    chain.append((EN_ZH, "manual"))
+    cedict = cedict_map()
+    if cedict:
+        chain.append((cedict, "cedict"))
+    return chain
+
+
+# 生效条件：query 为字符串，query.lower().strip() 命中 COMPOUND_ZH 即返回单元素 [comp_zh] 与 phrase_mapped 记录；否则经 COMPOUND_ZH_PHRASES 内嵌替换、剥除 's 与分词后逐词映射，返回 terms 与 inline_hits + detail；
+def normalize_en_query(query, extra_map=None):
+    """英文 query → 语义原子序列（中文语素）
+
+    返回 (normalized_terms, detail)：
+      normalized_terms: 用于检索的中文/保留词序列
+      detail: 逐词归一化记录（mapped 词带 source/low_confidence 置信信号；
+              low_confidence=True = 机械反查未人工校对，语境失配风险）
+    """
+    chain = _lookup_chain(extra_map)
+    # 短语优先匹配：先尝试多词短语整体映射（复合概念名）
+    phrase_key = query.lower().strip()
+    if phrase_key in COMPOUND_ZH:
+        comp_zh = COMPOUND_ZH[phrase_key]
+        return [comp_zh], [{"orig": query, "phrase_zh": comp_zh,
+                            "action": "phrase_mapped"}]
+    # 内嵌短语扫描：多词短语在长文本/问句内部出现时整体替换为中文
+    # （整体 query 匹配只救短语独占 query 的形态；doc 侧英文原文与含修饰语
+    # 的问句靠此层，否则逐词直译拆散复合语义——第三方 LoCoMo 报告错译样本
+    # scavenger hunt → 食腐动物 猎取 即此缺口。短语表键 ≤11 个，逐键扫描
+    # 成本可忽略）。中文替换段经 findall 整体成原子，后续逐词链路不受影响
+    inline_hits = []
+    for phrase, pzh in COMPOUND_ZH_PHRASES:
+        if phrase in query.lower():
+            query = re.sub(re.escape(phrase), pzh, query, flags=re.IGNORECASE)
+            inline_hits.append({"orig": phrase, "phrase_zh": pzh,
+                                "action": "phrase_inline"})
+    # 屈折还原 + 查表。所有格剥离：'s 是正字法黏着成分非独立词
+    # （Melanie's → Melanie）——不剥离则分词残留 "s" 成为伪 OOV
+    # （locomo-500 实测 85 词次）。
+    words = re.findall(r"[A-Za-z\u4e00-\u9fff]+",
+                       re.sub(r"'s\b", "", query))
+    terms = []
+    detail = []
+    for w in words:
+        wl = w.lower()
+        base0 = strip_tense(wl)
+        # 原形与还原形都查虚词表：getting/doing 屈折形原表漏网（还原后
+        # get/do 是虚词，保留只产生 OOV 噪声原子）
+        if wl in STOPWORDS or base0 in STOPWORDS:
+            detail.append({"orig": w, "action": "stopword_drop"})
+            continue
+        base = base0
+        zh = src = None
+        for m, s in chain:
+            z = m.get(base) or m.get(wl)
+            if z:
+                zh, src = z, s
+                break
+            if not base.endswith("e"):
+                # e-脱落动词词表感知还原：loved→lov(错形)→love。词表小时收益≈0
+                # （2026-09-14 早前取证 102 词次判不修）；词级表 17700 键后
+                # motivated→motivate / visited 类命中真实存在，条件已变
+                z = m.get(base + "e")
+                if z:
+                    base = base + "e"
+                    zh, src = z, s
+                    break
+        if zh and is_proper(w) and wl not in EN_ZH:
+            # 专名词表命中双原子：doc 侧音译/原文两形态并存（corpus567 实测
+            # 地名 巴黎13/Paris4 音译主导，人名 Caroline127/卡罗琳0 原文主导）
+            # ——原文+译文都进匹配面；未命中侧在 doc 侧零出现，组合共现需
+            # 双方在场，不构成假匹配。手工表已有词（i/we/you 等句首大写代词、
+            # 基础词）形态唯一，不具双形态不确定性，排除
+            terms.append(w)
+            terms.append(zh)
+            detail.append({"orig": w, "base": base, "zh": zh, "source": src,
+                           "low_confidence": src == "cedict",
+                           "action": "proper_mapped_both"})
+        elif zh:
+            terms.append(zh)
+            detail.append({"orig": w, "base": base, "zh": zh, "source": src,
+                           "low_confidence": src == "cedict",
+                           "action": "mapped"})
+        elif is_proper(w):
+            terms.append(w)  # 专有词保留原名
+            detail.append({"orig": w, "action": "proper_noun_keep"})
+        else:
+            # 未知词：尝试复合拆解（compound→zh morphemes）
+            comp = COMPOUND_ZH.get(wl) or COMPOUND_ZH.get(base)
+            if comp:
+                terms.append(comp)
+                detail.append({"orig": w, "compound_zh": comp, "action": "compound_mapped"})
+            else:
+                terms.append(w)
+                detail.append({"orig": w, "action": "unknown_keep"})
+    return terms, inline_hits + detail
+
+
+# 生效条件：detail 中某项 d 的 d.get("low_confidence") 为真值时，取 (d.get("orig") or "").lower()，仅当结果非空且尚未出现在 out 中才按序追加，最终返回去重保序的 out。
+def low_confidence_terms(detail):
+    """从归一化 detail 提取低置信映射词清单（去重保序）。
+
+    低置信 = mapped 词来自 CEDICT 机械反查（第一义项直译，未经人工
+    语境校对）。用途：错译审计靶子 / 词表人工校对优先级（警告不拒绝，
+    与 semantic_oov 同哲学——标记是观测依据，不改检索行为）。
+    """
+    out = []
+    for d in detail:
+        if d.get("low_confidence"):
+            w = (d.get("orig") or "").lower()
+            if w and w not in out:
+                out.append(w)
+    return out
+
+
+# 生效条件：len(sys.argv) > 1 时 query 为 sys.argv[1:] 以空格连接，否则 query 为字面默认 "I eat beef yesterday"；随后打印归一化 terms 与逐条 detail，无返回值。
+def main():
+    import sys
+    query = " ".join(sys.argv[1:]) if len(sys.argv) > 1 else "I eat beef yesterday"
+    terms, detail = normalize_en_query(query)
+    print("query:", query)
+    print("normalized terms:", terms)
+    print("search string:", " ".join(terms))
+    for d in detail:
+        print(" ", json.dumps(d, ensure_ascii=False))
+
+
+if __name__ == "__main__":
+    main()