简历解析加固：标签字间空格容错（姓名：X）、在读学历识别（硕士在读/大二→本科在读）、docx 表格文本按文档顺序提取（修复技能区整块丢失）、词表源补 LLM/大语言模型/Transformers/Prompt Engineering/智能体开发等真实简历高频别名（generate_dataset.py 为词表唯一源，data/ 不入库）、在读/应届年限显示，新增回归测试 2 例，real_jd_500 F1 99.83% 无回退

2026-08-31 20:44:57 +08:00
parent a4db91bf33
commit 896ac10f12
5 changed files with 80 additions and 13 deletions
@@ -18,13 +18,17 @@ logger = logging.getLogger(__name__)

 _PHONE_RE = re.compile(r"1[3-9]\d{9}")
 _EMAIL_RE = re.compile(r"[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,}")
-_NAME_RE = re.compile(r"姓名[:：]\s*([一-龥]{2,4})")
+# 标签字间可含空格/全角空格（PDF/Word 排版常见："姓 名 ：孙千越"）
+_NAME_RE = re.compile(r"姓\s*名\s*[:：]\s*([一-龥]{2,4})")
 # 简历常见写法：首行裸姓名（无"姓名："前缀），如"      王伟航      "
 _BARE_NAME_RE = re.compile(r"^\s*([一-龥]{2,4})\s*$", re.MULTILINE)
 _YEARS_RE = re.compile(r"(\d+(?:\.\d+)?)\s*年(?:以上)?[^\n，。；]{0,6}?经验")
-_GENDER_RE = re.compile(r"性别[:：]?\s*(男|女)")
+_GENDER_RE = re.compile(r"性\s*别\s*[:：]?\s*(男|女)")
 _EDU_RE = re.compile(r"(博士后|博士|专硕|学硕|硕士|研究生|MBA|本科|大专|专科)")
 _EDU_MAP = {"专硕": "硕士", "学硕": "硕士", "研究生": "硕士"}
+# 在读学历：「硕士在读」或「大二/大三」这类年级写法 → 本科（在读）
+_EDU_READING_RE = re.compile(r"(本科|硕士|博士|研究生|MBA)\s*在读")
+_GRADE_RE = re.compile(r"大[一二三四]")
 _ADDR_RE = re.compile(r"(?:家庭住址|现居(?:住)?(?:地|址)?|联系地址|通讯地址|籍贯)[:：]\s*([^\n，。；]{2,30})")

 # 常见简历分节标题（整行匹配，可带冒号）
@@ -48,6 +52,17 @@ def _extract_name(text: str) -> str:
    return ""


+def _extract_education(text: str) -> str:
+    """学历：优先识别在读（"硕士在读"/"大二"年级写法），再匹配学位词。"""
+    rm = _EDU_READING_RE.search(text)
+    if rm:
+        return f"{_EDU_MAP.get(rm.group(1), rm.group(1))}（在读）"
+    if _GRADE_RE.search(text):
+        return "本科（在读）"
+    m = _EDU_RE.search(text)
+    return _EDU_MAP.get(m.group(1), m.group(1)) if m else ""
+
+
 def mask_text(text: str) -> str:
    """脱敏：手机、邮箱、姓名（含首行裸姓名）、住址打码。"""
    text = _PHONE_RE.sub(lambda m: m.group(0)[:3] + "****" + m.group(0)[-2:], text)
@@ -110,6 +125,34 @@ def extract_photo_b64(path: str | Path) -> str:
    return ""


+def _docx_text(path: Path) -> str:
+    """按文档顺序提取 docx 文本：段落 + 表格（含嵌套表格单元格）。
+
+    简历模板普遍用无边框表格排版个人信息/技能区，只读 paragraphs 会整块丢失。
+    """
+    from docx import Document  # python-docx，requirements 已声明
+    from docx.table import Table
+    from docx.text.paragraph import Paragraph
+
+    doc = Document(str(path))
+    parts: list[str] = []
+
+    def walk_table(tbl) -> None:
+        for row in tbl.rows:
+            for cell in row.cells:
+                if cell.text.strip():
+                    parts.append(cell.text)
+                for nested in cell.tables:
+                    walk_table(nested)
+
+    for child in doc.element.body.iterchildren():
+        if child.tag.endswith("}p"):
+            parts.append(Paragraph(child, doc).text)
+        elif child.tag.endswith("}tbl"):
+            walk_table(Table(child, doc))
+    return "\n".join(parts)
+
+
 def read_resume_file(path: str | Path) -> str:
    """按扩展名读取简历文件文本。"""
    path = Path(path)
@@ -117,9 +160,7 @@ def read_resume_file(path: str | Path) -> str:
    if suffix == ".txt":
        return path.read_text(encoding="utf-8")
    if suffix == ".docx":
-        from docx import Document  # python-docx，requirements 已声明
-        doc = Document(str(path))
-        return "\n".join(p.text for p in doc.paragraphs)
+        return _docx_text(path)
    if suffix == ".pdf":
        try:
            from pypdf import PdfReader
@@ -147,7 +188,6 @@ def parse_resume(text: str, matcher: SkillMatcher | None = None,
    email_m = _EMAIL_RE.search(text)
    years_m = _YEARS_RE.search(text)
    gender_m = _GENDER_RE.search(text)
-    edu_m = _EDU_RE.search(text)
    addr_m = _ADDR_RE.search(text)
    masked = mask_text(text)
    return ResumeProfile(
@@ -159,7 +199,7 @@ def parse_resume(text: str, matcher: SkillMatcher | None = None,
        scenarios=matcher.find_scenarios(text),
        raw_masked_text=masked,
        gender=gender_m.group(1) if gender_m else "",
-        education=_EDU_MAP.get(edu_m.group(1), edu_m.group(1)) if edu_m else "",
+        education=_extract_education(text),
        address=_mask_addr(addr_m.group(1).strip()) if addr_m else "",
        photo_b64=photo_b64,
        sections=split_sections(masked),
@@ -84,7 +84,7 @@ SKILLS = [
    ("数据挖掘", "大数据", [], "L2", ["Python", "SQL"]),
    # AI 与算法
    ("机器学习", "AI算法", ["机器学习算法"], "L2", ["Python"]),
-    ("深度学习", "AI算法", [], "L3", ["机器学习"]),
+    ("深度学习", "AI算法", ["Transformers", "Transformer"], "L3", ["机器学习"]),
    ("PyTorch", "AI算法", [], "L2", ["Python", "深度学习"]),
    ("TensorFlow", "AI算法", [], "L2", ["Python", "深度学习"]),
    ("Keras", "AI算法", [], "L2", ["TensorFlow"]),
@@ -95,9 +95,12 @@ SKILLS = [
    ("OpenCV", "AI算法", [], "L2", ["Python"]), ("OCR", "AI算法", ["文字识别"], "L2", ["计算机视觉"]),
    ("语音识别", "AI算法", [], "L3", ["深度学习"]), ("推荐算法", "AI算法", ["推荐系统"], "L3", ["机器学习"]),
    ("强化学习", "AI算法", [], "L3", ["深度学习"]), ("知识图谱", "AI算法", [], "L3", ["NLP"]),
-    ("大模型应用开发", "AI算法", ["LLM应用开发", "大模型应用", "大语言模型应用"], "L3", ["Python", "NLP"]),
+    ("大模型应用开发", "AI算法",
+     ["LLM应用开发", "大模型应用", "大语言模型应用", "LLM", "大语言模型", "大模型", "LLM推理", "LLM评测", "大模型推理", "大语言模型应用开发"],
+     "L3", ["Python", "NLP"]),
    ("RAG", "AI算法", ["检索增强生成"], "L3", ["大模型应用开发", "向量数据库"]),
-    ("提示词工程", "AI算法", ["Prompt工程"], "L2", ["大模型应用开发"]),
+    ("提示词工程", "AI算法", ["Prompt工程", "Prompt Engineering", "提示词设计"], "L2", ["大模型应用开发"]),
+    ("智能体开发", "AI算法", ["Agent开发", "多智能体", "Multi-Agent", "智能体", "Agent", "AI Agent"], "L3", ["大模型应用开发"]),
    ("向量数据库", "AI算法", ["向量库"], "L2", ["数据库基础"]),
    ("LangChain", "AI算法", [], "L2", ["大模型应用开发"]),
    ("模型微调", "AI算法", ["Fine-tuning", "SFT"], "L3", ["深度学习", "PyTorch"]),
@@ -107,6 +107,27 @@ def test_resume_bare_name_and_edu_alias():
    assert "Python" in p.skills


+def test_resume_spaced_labels_and_reading_edu():
+    """PDF/Word 排版常见"姓 名 ：孙千越"标签字间带空格；年级写法"大二"推断本科在读。"""
+    from app.match.resume import parse_resume
+    text = ("姓 名 ：孙千越\n学 校：东北财经大学  出生年月：2007 年 12 月 23 日\n"
+            "年 级：大二  电子邮箱：qianyuesun@outlook.com\n"
+            "技术能力：Python、Transformers、LLM 推理与评测、RAG、Prompt Engineering、"
+            "Multi-Agent Workflow。")
+    p = parse_resume(text)
+    assert p.name == "孙**" and "孙千越" not in p.raw_masked_text   # 空格分隔标签 + 脱敏
+    assert p.education == "本科（在读）"                             # 年级写法 → 在读学历
+    for s in ["大模型应用开发", "智能体开发", "深度学习", "提示词工程", "RAG", "Python"]:
+        assert s in p.skills, s                                    # 新词表别名命中
+
+
+def test_resume_reading_degree_explicit():
+    """明确"硕士在读"写法：保留在读标记，不降级为已毕业学位。"""
+    from app.match.resume import parse_resume
+    p = parse_resume("姓名：张三\n学历：硕士在读\n技能：Python")
+    assert p.education == "硕士（在读）"
+
+
 def test_learning_path_stage_chunking():
    """同层技能超 3 个时拆成多个连续阶段；拓扑序不变。"""
    stages = topological_stages(["Linux", "Docker", "Kubernetes", "Python", "Java", "Go", "MySQL"],
@@ -1388,9 +1388,12 @@ function applyParsedProfile(p, sourceDesc) {
  matchState.skills = [...p.skills];
  matchState.years = p.years || 0;
  $("#resume-text").value = p.raw_masked_text || "";
+  const yearsHtml = p.years
+    ? `工作年限约 <b>${p.years}</b> 年`
+    : (/应届|在读|大[一二三四]/.test(p.raw_masked_text || "") ? "工作年限 <b>在读 / 应届</b>" : "工作年限 <b>未识别</b>");
  $("#parse-summary").innerHTML =
    `解析完成（来源：${esc(sourceDesc || "文本")}，已自动脱敏）：识别技能 <b>${p.skills.length}</b> 项 · ` +
-    `工作年限约 <b>${p.years || 0}</b> 年 · 场景经验 <b>${(p.scenarios || []).join("、") || "未识别"}</b>`;
+    `${yearsHtml} · 场景经验 <b>${(p.scenarios || []).join("、") || "未识别"}</b>`;
  $("#parse-years").value = matchState.years;
  renderSkillChips();
  setMatchStep(2);
@@ -1577,7 +1580,7 @@ function renderResumeHighlight(m) {
        <td class="rd-photo-cell" rowspan="3">${photo}</td>
        ${kv("姓名", p.name)}${kv("性别", p.gender)}
      </tr>
-      <tr>${kv("学历", p.education)}${kv("工作年限", p.years ? p.years + " 年" : "")}</tr>
+      <tr>${kv("学历", p.education)}${kv("工作年限", p.years ? p.years + " 年" : (/应届|在读|大[一二三四]/.test(p.raw_masked_text || "") ? "在读 / 应届" : ""))}</tr>
      <tr>${kv("手机", p.phone)}${kv("邮箱", p.email)}</tr>
      <tr><td class="rd-k">家庭住址</td><td class="rd-v" colspan="4">${esc(p.address || "—")}</td></tr>
      ${secs.map((sec) => `
@@ -705,7 +705,7 @@
    <i class="tour-track"><i id="tour-fill"></i></i>
  </div>

-  <script src="/static/app.js?v=20260831"></script>
+  <script src="/static/app.js?v=20260831-2"></script>
  <script src="/static/graph.js?v=20260831"></script>
 </body>
 </html>