feat: refactor _extract_from_markdown with LLM-enhanced table/formula/code extraction

examples/multi_extractor_compare.py

@@ -8,14 +8,15 @@ def all_extractor_comparison():
     print("\n=== 多抽取器对比演示 ===\n")
 
     # 创建数据集
-    dataset_path = Path("../data/WebMainBench_llm-webkit_v1_WebMainBench_7887_within_formula.jsonl")
+    dataset_path = Path("../data/sample_dataset.jsonl")
     dataset = DataLoader.load_jsonl(dataset_path)
-    
+
     # 创建webkit抽取器
     config = {
         "use_preprocessed_html": True,          # 🔑 关键配置：启用预处理HTML模式
         "preprocessed_html_field": "llm_webkit_html"  # 指定预处理HTML字段名
     }
+
     webkit_extractor = ExtractorFactory.create("llm-webkit", config=config)
     # 创建magic-extractor抽取器
     magic_extractor = ExtractorFactory.create("magic-html")

requirements.txt

@@ -12,4 +12,5 @@ https://github.com/opendatalab/magic-html/releases/download/magic_html-0.1.5-rel
 streamlit
 markdown
 jieba
-apted
+apted
+openai

webmainbench/metrics/__init__.py

@@ -11,6 +11,10 @@
 from .teds_metrics import TEDSMetric, StructureTEDSMetric
 from .calculator import MetricCalculator
 from .mainhtml_calculator import MainHTMLMetricCalculator
+from .base_extractor import ContentExtractor
+from .formula_extractor import FormulaExtractor
+from .code_extractor import CodeExtractor
+from .table_extractor import TableExtractor
 
 __all__ = [
     "BaseMetric",
@@ -27,4 +31,8 @@
     "TextEditMetric",
     "MetricCalculator",
     "MainHTMLMetricCalculator",
+    'ContentExtractor',
+    'FormulaExtractor',
+    'CodeExtractor',
+    'TableExtractor',
 ] 

webmainbench/metrics/base.py

@@ -4,12 +4,7 @@
 
 from abc import ABC, abstractmethod
 from dataclasses import dataclass
-from typing import Dict, Any, List, Optional, Union
-import traceback
-import re
-from bs4 import BeautifulSoup
-import os
-import hashlib
+from typing import Dict, Any, List, Optional
 
 @dataclass
 class MetricResult:
@@ -144,7 +139,7 @@ def split_content(text: str, content_list: List[Dict[str, Any]] = None, field_na
 
         # 从markdown文本中提取，传递字段名称
         return BaseMetric._extract_from_markdown(text or "", field_name=field_name)
-    
+
     @staticmethod
     def _extract_from_content_list(content_list: List[Dict[str, Any]]) -> Dict[str, str]:
         """从content_list中递归提取各种类型的内容"""
@@ -194,233 +189,42 @@ def _recursive_extract(items):
             'table': '\n'.join(extracted['table']),
             'text': '\n'.join(extracted['text'])
         }
-    
+
     @staticmethod
     def _extract_from_markdown(text: str, field_name: str = None) -> Dict[str, str]:
         """从markdown文本中提取各种类型的内容"""
         if not text:
             return {'code': '', 'formula': '', 'table': '', 'text': ''}
 
-        # 收集所有需要移除的内容片段
-        extracted_segments = []
-        code_parts = []
-        # # 同匹配行间代码块 ```...```
-        # pattern = r'(```[\s\S]*?```)'
-        # for match in re.finditer(pattern, text):
-        #     code_segment = match.group(0)
-        #     extracted_segments.append(code_segment)
-        #
-        #     if code_segment.startswith('```'):
-        #         # 处理代码块（保留内部缩进）
-        #         lines = code_segment.split('\n')
-        #         # 移除首尾的```标记
-        #         content_lines = lines[1:-1]
-        #         # 保留原始缩进，只拼接内容
-        #         code_content = '\n'.join(content_lines)
-        #     else:
-        #         # 处理行内代码（只去除外层`和前后空格）
-        #         code_content = code_segment[1:-1].strip()
-        #
-        #     if code_content:  # 只添加非空内容
-        #         code_parts.append(code_content)
-
-        # 1. 首先处理三个反引号包裹的代码块（优先级最高）
-        backtick_pattern = r'(```[\s\S]*?```)'
-        for match in re.finditer(backtick_pattern, text):
-            code_segment = match.group(0)
-
-            if code_segment.startswith('```'):
-                # 处理代码块
-                lines = code_segment.split('\n')
-                # 移除首尾的```标记
-                content_lines = lines[1:-1]
-                code_content = '\n'.join(content_lines)
-            else:
-                # 处理行内代码
-                code_content = code_segment[1:-1].strip()
-
-            if code_content:
-                code_parts.append(code_content)
-
-        # 2. 处理缩进代码块 - 使用更精确的匹配
-        # 匹配模式：前面有空行 + 连续的多行缩进内容 + 后面有空行
-        # 关键：要求所有匹配的行都是缩进的
-        indent_pattern = r'(?:\n\s*\n)((?:(?: {4,}|\t+)[^\n]*(?:\n|$)){2,})(?=\n\s*\n|$)'
-
-        for match in re.finditer(indent_pattern, text, re.MULTILINE):
-            code_segment = match.group(1)
-
-            # 验证：确保所有行都是缩进的（避免混合缩进和非缩进行）
-            lines = code_segment.split('\n')
-            all_indented = all(
-                line.startswith('    ') or line.startswith('\t') or not line.strip()
-                for line in lines
-                if line.strip()  # 空行不算
-            )
-
-            if not all_indented:
-                continue  # 跳过包含非缩进行的块
-
-            # 进一步验证代码特征
-            non_empty_lines = [line.strip() for line in lines if line.strip()]
-            if len(non_empty_lines) < 2:  # 至少2行非空内容
-                continue
-
-            # 检查是否有明显的非代码特征
-            has_list_features = any(
-                re.match(r'^[-•*]\s', line) or
-                re.match(r'^\d+\.\s', line) or
-                re.search(r'\$[\d,]', line) or
-                re.search(r'\b(million|billion|thousand)\b', line, re.IGNORECASE)
-                for line in non_empty_lines
-            )
-
-            if has_list_features:
-                continue  # 跳过列表内容
-
-            # 清理代码段
-            cleaned_lines = []
-            for line in code_segment.split('\n'):
-                if line.strip():
-                    if line.startswith('    '):
-                        cleaned_lines.append(line[4:])
-                    elif line.startswith('\t'):
-                        cleaned_lines.append(line[1:])
-                    else:
-                        cleaned_lines.append(line)
-
-            code_content = '\n'.join(cleaned_lines)
-            if code_content.strip():
-                code_parts.append(code_content)
-
-        # 提取公式 - 新的两步处理逻辑
-        formula_parts = []
-
-        # 第一步：先用正则提取公式
-        regex_formulas = []
-        latex_patterns = [
-            r'(?<!\\)\$\$(.*?)(?<!\\)\$\$',  # 行间 $$...$$
-            r'(?<!\\)\\\[(.*?)(?<!\\)\\\]',  # 行间 \[...\]
-            r'(?<!\\)\$(.*?)(?<!\\)\$',  # 行内 $...$
-            r'(?<!\\)\\\((.*?)(?<!\\)\\\)',  # 行内 \(...\)
-        ]
-
-        for pattern in latex_patterns:
-            for match in re.finditer(pattern, text, re.DOTALL):
-                formula_full = match.group(0)
-                formula_content = match.group(1)
-                extracted_segments.append(formula_full)
-                if formula_content.strip():
-                    regex_formulas.append(formula_content.strip())
-
-        # 第二步：根据字段类型决定是否需要API修正
-        if field_name == "groundtruth_content":
-            print(f"[DEBUG] 检测到groundtruth内容，仅使用正则提取公式")
-            formula_parts = regex_formulas
-        else:
-            print(f"[DEBUG] 检测到llm_webkit_md内容，使用正则+API修正模式")
-            # 对于llm_webkit_md，将正则结果传递给API进行修正
-            if regex_formulas:
-                # 将正则提取的公式作为输入传递给API
-                regex_formulas_text = '\n'.join(regex_formulas)
-                print(f"[DEBUG] 正则提取到 {len(regex_formulas)} 个公式，准备API修正")
-
-                cache_dir = os.path.join(os.path.dirname(os.path.abspath(__file__)), '.cache')
-                os.makedirs(cache_dir, exist_ok=True)
-
-                # 使用正则结果的哈希作为缓存文件名
-                text_hash = hashlib.md5(regex_formulas_text.encode('utf-8')).hexdigest()
-                cache_file = os.path.join(cache_dir, f'formula_correction_cache_{text_hash}.json')
-
-                try:
-                    from .formula_extractor import correct_formulas_with_llm
-                    corrected_formulas = correct_formulas_with_llm(regex_formulas, cache_file)
-                    formula_parts = corrected_formulas
-                    print(f"[DEBUG] API修正成功，最终得到 {len(formula_parts)} 个公式")
-                except Exception as e:
-                    print(f"[DEBUG] API修正失败: {type(e).__name__}: {e}，使用正则结果")
-                    formula_parts = regex_formulas
-            else:
-                print(f"[DEBUG] 正则未提取到公式，跳过API修正")
-                formula_parts = []
-
-        # 提取表格
-        table_parts = []
-
-        # ===== 1. 提取 HTML 表格 =====
-        # 用 BeautifulSoup 替代正则，防止嵌套或匹配不全
-        soup = BeautifulSoup(text, "html.parser")
-        for table in soup.find_all("table"):
-            # 判断当前表格的父级是否是表格内的标签（<td>、<tr>、<tbody>等）
-            parent_is_table_related = table.find_parent(["td", "tr", "tbody", "table"]) is not None
-            if not parent_is_table_related:  # 父级不是表格相关标签 → 是外层表格
-                html_table = str(table)
-                extracted_segments.append(html_table)
-                table_parts.append(html_table)
-
-        # ===== 2. 提取 Markdown 表格 =====
-        lines = text.split('\n')
-        table_lines = []
-        in_markdown_table = False
-        found_separator = False  # 是否已找到分隔行
-
-        def is_md_table_line(line):
-            """判断是否可能是 Markdown 表格行"""
-            if line.count("|") < 1:  # 至少三个竖线
-                return False
-            return True
-
-        def is_md_separator_line(line):
-            """判断是否为 Markdown 分隔行"""
-            parts = [p.strip() for p in line.split("|")]
-            # 检查是否所有部分都是分隔符格式
-            for p in parts:
-                if p and not re.match(r"^:?\-{3,}:?$", p):
-                    return False
-            return True
+        # 创建提取器配置
+        config = {
+            'llm_base_url': '',
+            'llm_api_key': '',
+            'llm_model': '',
+            'use_llm': False  # 使用时改为True
+        }
 
-        def save_table():
-            """保存当前表格并清空缓存"""
-            nonlocal table_lines
-            #     只有当表格行数大于等于2，且第二行是分隔行时才保存
-            if len(table_lines) >= 2 and is_md_separator_line(table_lines[1]):
-                md_table = '\n'.join(table_lines)
-                extracted_segments.append(md_table)
-                table_parts.append(md_table)
+        # 直接创建具体的提取器实例
+        from .code_extractor import CodeExtractor
+        from .formula_extractor import FormulaExtractor
+        from .table_extractor import TableExtractor
 
-        for line in lines:
-            if is_md_table_line(line):
-                table_lines.append(line)
-                in_markdown_table = True
-                if is_md_separator_line(line):
-                    found_separator = True
-            else:
-                if in_markdown_table:
-                    save_table()
-                    table_lines = []
-                    in_markdown_table = False
-                    found_separator = False
+        code_extractor = CodeExtractor(config)
+        formula_extractor = FormulaExtractor(config)
+        table_extractor = TableExtractor(config)
 
-        # 处理文档末尾的 Markdown 表格
-        if in_markdown_table:
-            save_table()
+        # 提取各类内容
+        code_content = code_extractor.extract(text, field_name)
+        formula_content = formula_extractor.extract(text, field_name)
+        table_content = table_extractor.extract(text, field_name)
 
-        # 提取剩余文本（移除所有已提取的内容片段）
-        clean_text = text
-        for segment in extracted_segments:
-            clean_text = clean_text.replace(segment, '', 1)
-
-        # 清理多余的空行
-        clean_text = re.sub(r'\n\s*\n', '\n\n', clean_text)
-        clean_text = clean_text.strip()
-
         return {
-            'code': '\n'.join(code_parts),
-            'formula': '\n'.join(formula_parts),
-            'table': '\n'.join(table_parts),
-            'text': text  # 原始全部文本
+            'code': code_content,
+            'formula': formula_content,
+            'table': table_content,
+            'text': text  # 保留原始全部文本
         }
-    
+
     def aggregate_results(self, results: List[MetricResult]) -> MetricResult:
         """
         Aggregate multiple metric results.