#!/usr/bin/env python3 # -*- coding: utf-8 -*- import os import re import docx import magic import numpy as np import requests from sklearn.feature_extraction.text import TfidfVectorizer from sklearn.metrics.pairwise import cosine_similarity from typing import List, Tuple, Dict, Optional from docx.shared import Pt from docx.enum.text import WD_PARAGRAPH_ALIGNMENT from docx.enum.table import WD_TABLE_ALIGNMENT import subprocess import tempfile import json from docx.table import Table, _Cell from docx.text.paragraph import Paragraph from copy import deepcopy from docx.oxml import parse_xml from docx.oxml.ns import nsdecls class DocCleaner: def __init__(self, ollama_host: str = "http://192.168.1.18:11434"): """ 初始化文档清理器 Args: ollama_host: Ollama服务器地址 """ # 页眉页脚模式 self.header_footer_patterns = [ r'页码\s*\d+-\d+', # 页码格式：页码1-1, 页码2-1等 r'第\s*\d+\s*页\s*共\s*\d+\s*页', # 中文页码（第X页共Y页） r'Page\s*\d+\s*of\s*\d+', # 英文页码 ] # 特殊符号模式 self.special_char_patterns = [ r'©\s*\d{4}.*?版权所有', # 版权信息 r'confidential', # 机密标记 r'draft|草稿', # 草稿标记 r'watermark', # 水印标记 ] # 附录和参考文献标题模式 self.appendix_patterns = [ r'^附录\s*[A-Za-z]?[\s:：]', r'^Appendix\s*[A-Za-z]?[\s:：]', r'^参考文献$', r'^References$', r'^Bibliography$' ] # 初始化TF-IDF向量化器 self.vectorizer = TfidfVectorizer( min_df=1, stop_words='english' ) self.ollama_host = ollama_host self.embedding_model = "bge-m3:latest" # 使用nomic-embed-text模型进行文本嵌入 def _convert_doc_to_docx(self, doc_path: str) -> str: """ 将doc格式转换为docx格式 Args: doc_path: doc文件路径 Returns: str: 转换后的docx文件路径 """ # 创建临时文件路径 temp_dir = tempfile.mkdtemp() temp_docx = os.path.join(temp_dir, 'temp.docx') try: # 使用soffice（LibreOffice）进行转换 cmd = ['soffice', '--headless', '--convert-to', 'docx', '--outdir', temp_dir, doc_path] subprocess.run(cmd, check=True, capture_output=True) # 返回转换后的文件路径 return temp_docx except subprocess.CalledProcessError as e: raise Exception(f"转换doc文件失败: {str(e)}") def clean_doc(self, file_path: str) -> Tuple[List[str], List[str], List[Table]]: """ 清理文档并返回处理后的正文、附录和表格 Args: file_path: 文档文件路径 Returns: Tuple[List[str], List[str], List[Table]]: (清理后的正文段落列表, 附录段落列表, 表格列表) """ print(f"\n开始处理文档: {file_path}") # 检测文件类型 file_type = magic.from_file(file_path, mime=True) # 如果是doc格式，先转换为docx if file_type == 'application/msword': temp_docx = self._convert_doc_to_docx(file_path) doc = docx.Document(temp_docx) # 清理临时文件 os.remove(temp_docx) os.rmdir(os.path.dirname(temp_docx)) else: doc = docx.Document(file_path) # 提取所有内容（段落和表格） content = [] tables = [] table_count = 0 try: print("\n开始解析文档结构...") # 遍历文档体中的所有元素 for element in doc._element.body: if element.tag.endswith('p'): try: paragraph = docx.text.paragraph.Paragraph(element, doc) text = paragraph.text.strip() # 只添加非空段落 if text: # 检查是否是附录标题 is_appendix = any(re.match(pattern, text, re.IGNORECASE) for pattern in self.appendix_patterns) content.append({ 'type': 'paragraph', 'content': text, 'is_appendix_start': is_appendix }) if is_appendix: print(f"发现附录标题: {text}") except Exception as e: print(f"警告：处理段落时出错: {str(e)}") continue elif element.tag.endswith('tbl'): try: table = docx.table.Table(element, doc) # 验证表格是否有效 if hasattr(table, 'rows') and hasattr(table, 'columns'): tables.append(table) content.append({ 'type': 'table', 'index': table_count }) print(f"发现表格 {table_count}: {len(table.rows)}行 x {len(table.columns)}列") table_count += 1 except Exception as e: print(f"警告：处理表格时出错: {str(e)}") continue except Exception as e: print(f"警告：遍历文档内容时出错: {str(e)}") print(f"\n文档结构解析完成:") print(f"- 总元素数: {len(content)}") print(f"- 表格数量: {len(tables)}") # 分离正文和附录 main_content = [] appendix = [] is_appendix = False print("\n开始分离正文和附录...") for item in content: if item['type'] == 'paragraph': if item['is_appendix_start']: is_appendix = True print("进入附录部分") if is_appendix: appendix.append(item['content']) else: main_content.append(item['content']) elif item['type'] == 'table': table_placeholder = f'TABLE_PLACEHOLDER_{item["index"]}' if is_appendix: appendix.append(table_placeholder) print(f"添加表格到附录: {table_placeholder}") else: main_content.append(table_placeholder) print(f"添加表格到正文: {table_placeholder}") print(f"\n分离完成:") print(f"- 正文元素数: {len(main_content)}") print(f"- 附录元素数: {len(appendix)}") # 清理正文（保留表格标记） cleaned_content = [] print("\n开始清理正文...") for item in main_content: if item.startswith('TABLE_PLACEHOLDER_'): cleaned_content.append(item) print(f"保留表格标记: {item}") else: cleaned_text = self._clean_text([item])[0] if cleaned_text: cleaned_content.append(cleaned_text) print(f"\n清理完成:") print(f"- 清理后元素数: {len(cleaned_content)}") print("- 表格标记位置:") for i, item in enumerate(cleaned_content): if item.startswith('TABLE_PLACEHOLDER_'): print(f" 位置 {i}: {item}") return cleaned_content, appendix, tables def _clean_text(self, text: List[str]) -> List[str]: """ 清理文本内容 Args: text: 待清理的文本段落列表 Returns: List[str]: 清理后的文本段落列表 """ cleaned = [] for paragraph in text: # 如果是表格标记，直接保留 if paragraph.startswith('TABLE_PLACEHOLDER_'): cleaned.append(paragraph) continue # 跳过空段落 if not paragraph.strip(): continue # 检查是否是目录项（包含数字序号的行） is_toc_item = bool(re.match(r'^\s*(?:\d+\.)*\d+\s+.*', paragraph)) if not is_toc_item: # 移除页眉页脚 for pattern in self.header_footer_patterns: paragraph = re.sub(pattern, '', paragraph, flags=re.IGNORECASE) # 移除特殊符号 for pattern in self.special_char_patterns: paragraph = re.sub(pattern, '', paragraph, flags=re.IGNORECASE) # 如果段落不为空，添加到结果中 if paragraph.strip(): cleaned.append(paragraph.strip()) return cleaned def _split_content(self, paragraphs: List[str]) -> Tuple[List[str], List[str]]: """ 分离正文与附录/参考文献 Args: paragraphs: 文档段落列表 Returns: Tuple[List[str], List[str]]: (正文段落列表, 附录段落列表) """ main_content = [] appendix = [] is_appendix = False for p in paragraphs: # 检查是否是附录开始 if any(re.match(pattern, p, re.IGNORECASE) for pattern in self.appendix_patterns): is_appendix = True if is_appendix: appendix.append(p) else: main_content.append(p) return main_content, appendix def _get_embeddings(self, texts: List[str]) -> np.ndarray: """ 使用Ollama获取文本嵌入向量 Args: texts: 文本列表 Returns: np.ndarray: 嵌入向量矩阵 """ embeddings = [] for text in texts: try: response = requests.post( f"{self.ollama_host}/api/embeddings", json={ "model": self.embedding_model, "prompt": text } ) response.raise_for_status() embedding = response.json()["embedding"] embeddings.append(embedding) except Exception as e: print(f"获取文本嵌入失败: {str(e)}") # 如果获取嵌入失败，使用零向量 embeddings.append([0.0] * 768) # nomic-embed-text 模型输出维度为768 return np.array(embeddings) def _remove_duplicates(self, paragraphs: List[str], similarity_threshold: float = 0.92) -> List[str]: """ 删除重复段落，保持表格占位符的位置不变 Args: paragraphs: 段落列表 similarity_threshold: 相似度阈值，使用嵌入模型后可以设置更高的阈值 Returns: List[str]: 去重后的段落列表 """ if not paragraphs: return [] # 分离表格占位符和普通段落 table_placeholders = {} text_paragraphs = [] for i, p in enumerate(paragraphs): if p.startswith('TABLE_PLACEHOLDER_'): table_placeholders[i] = p else: text_paragraphs.append((i, p)) try: # 只对非表格段落进行去重 if text_paragraphs: # 获取文本嵌入 text_only = [p[1] for p in text_paragraphs] embeddings = self._get_embeddings(text_only) # 计算余弦相似度矩阵 similarity_matrix = cosine_similarity(embeddings) # 标记要保留的段落 keep_indices = [] for i in range(len(text_paragraphs)): # 如果当前段落没有与之前的段落高度相似，则保留 if not any(similarity_matrix[i][j] > similarity_threshold for j in keep_indices): keep_indices.append(i) # 保留的非表格段落 kept_paragraphs = [(text_paragraphs[i][0], text_only[i]) for i in keep_indices] else: kept_paragraphs = [] # 合并表格占位符和保留的段落，按原始位置排序 all_kept = list(table_placeholders.items()) + kept_paragraphs all_kept.sort(key=lambda x: x[0]) return [p[1] for p in all_kept] except Exception as e: print(f"使用Ollama嵌入模型失败，回退到TF-IDF方法: {str(e)}") # 如果使用Ollama失败，回退到原来的TF-IDF方法 return self._remove_duplicates_tfidf(paragraphs) def _remove_duplicates_tfidf(self, paragraphs: List[str], similarity_threshold: float = 0.85) -> List[str]: """ 使用TF-IDF方法删除重复段落（作为备选方案） Args: paragraphs: 段落列表 similarity_threshold: 相似度阈值 Returns: List[str]: 去重后的段落列表 """ if not paragraphs: return [] # 分离表格占位符和普通段落 table_placeholders = {} text_paragraphs = [] for i, p in enumerate(paragraphs): if p.startswith('TABLE_PLACEHOLDER_'): table_placeholders[i] = p else: text_paragraphs.append((i, p)) if text_paragraphs: # 计算TF-IDF矩阵 text_only = [p[1] for p in text_paragraphs] tfidf_matrix = self.vectorizer.fit_transform(text_only) # 计算余弦相似度矩阵 similarity_matrix = cosine_similarity(tfidf_matrix) # 标记要保留的段落 keep_indices = [] for i in range(len(text_paragraphs)): # 如果当前段落没有与之前的段落高度相似，则保留 if not any(similarity_matrix[i][j] > similarity_threshold for j in keep_indices): keep_indices.append(i) # 保留的非表格段落 kept_paragraphs = [(text_paragraphs[i][0], text_only[i]) for i in keep_indices] else: kept_paragraphs = [] # 合并表格占位符和保留的段落，按原始位置排序 all_kept = list(table_placeholders.items()) + kept_paragraphs all_kept.sort(key=lambda x: x[0]) return [p[1] for p in all_kept] def save_as_docx(self, cleaned_content: List[str], appendix: List[str], tables: List[Table], output_path: str): """ 将清理后的内容保存为docx格式 Args: cleaned_content: 清理后的正文段落列表 appendix: 附录段落列表 tables: 表格列表 output_path: 输出文件路径 """ print(f"\n开始保存文档: {output_path}") print(f"- 正文元素数: {len(cleaned_content)}") print(f"- 附录元素数: {len(appendix)}") print(f"- 表格总数: {len(tables)}") # 创建新文档 doc = docx.Document() # 添加正文内容和表格，保持它们的相对位置 print("\n处理正文内容...") # 创建一个列表来存储所有要插入的元素 elements_to_insert = [] for i, content in enumerate(cleaned_content): try: # 检查是否是表格占位符 table_match = re.match(r'TABLE_PLACEHOLDER_(\d+)', content) if table_match: table_index = int(table_match.group(1)) print(f"正在处理表格占位符: {content} (索引: {table_index})") if table_index < len(tables): table = tables[table_index] try: # 直接在XML级别复制表格 new_tbl = deepcopy(table._element) # 确保新表格有正确的命名空间 new_tbl.tbl = parse_xml(new_tbl.xml) elements_to_insert.append(('table', new_tbl)) print(f"准备插入表格 {table_index} 在位置 {i}") # 添加表格后的空行 elements_to_insert.append(('paragraph', doc.add_paragraph()._element)) except Exception as e: print(f"警告：复制表格时出错: {str(e)}") try: print("尝试使用备用方法...") p = doc.add_paragraph() self._copy_table_fallback(p._parent, table) elements_to_insert.append(('paragraph', p._element)) elements_to_insert.append(('paragraph', doc.add_paragraph()._element)) print("备用方法成功") except Exception as e2: print(f"警告：备用方法也失败: {str(e2)}") elements_to_insert.append(('paragraph', doc.add_paragraph("【表格处理失败】")._element)) else: # 添加普通段落 p = doc.add_paragraph(content) p.alignment = WD_PARAGRAPH_ALIGNMENT.JUSTIFY elements_to_insert.append(('paragraph', p._element)) except Exception as e: print(f"警告：处理段落或表格时出错: {str(e)}") continue # 按顺序将所有元素插入文档 for element_type, element in elements_to_insert: doc._body._element.append(element) # 如果有附录，添加分隔符和附录内容 if appendix: print("\n处理附录内容...") try: # 添加分页符 doc.add_page_break() # 添加附录标题 title = doc.add_paragraph("附录") title.alignment = WD_PARAGRAPH_ALIGNMENT.CENTER # 添加附录内容 appendix_elements = [] for content in appendix: # 检查是否是表格占位符 table_match = re.match(r'TABLE_PLACEHOLDER_(\d+)', content) if table_match: table_index = int(table_match.group(1)) print(f"正在处理附录中的表格占位符: {content} (索引: {table_index})") if table_index < len(tables): table = tables[table_index] try: # 直接在XML级别复制表格 new_tbl = deepcopy(table._element) new_tbl.tbl = parse_xml(new_tbl.xml) appendix_elements.append(('table', new_tbl)) print(f"准备插入附录表格 {table_index}") appendix_elements.append(('paragraph', doc.add_paragraph()._element)) except Exception as e: print(f"警告：复制附录中的表格时出错: {str(e)}") try: p = doc.add_paragraph() self._copy_table_fallback(p._parent, table) appendix_elements.append(('paragraph', p._element)) appendix_elements.append(('paragraph', doc.add_paragraph()._element)) print("备用方法成功") except Exception as e2: print(f"警告：附录表格的备用方法也失败: {str(e2)}") appendix_elements.append(('paragraph', doc.add_paragraph("【表格处理失败】")._element)) else: p = doc.add_paragraph(content) p.alignment = WD_PARAGRAPH_ALIGNMENT.JUSTIFY appendix_elements.append(('paragraph', p._element)) # 按顺序将附录元素插入文档 for element_type, element in appendix_elements: doc._body._element.append(element) except Exception as e: print(f"警告：处理附录时出错: {str(e)}") # 保存文档 try: doc.save(output_path) print("\n文档保存成功!") except Exception as e: print(f"错误：保存文档时出错: {str(e)}") raise def _copy_table_fallback(self, doc: docx.Document, table: Table): """ 表格复制的备用方法 Args: doc: 目标文档 table: 源表格 """ # 获取表格的行数和列数 rows = len(table.rows) cols = len(table.columns) # 创建新表格 new_table = doc.add_table(rows=rows, cols=cols) # 复制表格样式 if table.style: new_table.style = table.style # 复制表格属性 new_table._element.tblPr = deepcopy(table._element.tblPr) # 复制网格信息 new_table._element.tblGrid = deepcopy(table._element.tblGrid) # 创建单元格映射以跟踪合并 cell_map = {} # 第一遍：标记合并的单元格 for i in range(rows): for j in range(cols): try: src_cell = table.cell(i, j) # 检查是否是合并单元格的一部分 if src_cell._element.tcPr is not None: # 检查垂直合并 vmerge = src_cell._element.tcPr.xpath('.//w:vMerge') if vmerge: val = vmerge[0].get(qn('w:val'), 'continue') if val == 'restart': # 这是合并的起始单元格 span = self._get_vertical_span(table, i, j) cell_map[(i, j)] = ('vmerge', span) # 检查水平合并 gridspan = src_cell._element.tcPr.xpath('.//w:gridSpan') if gridspan: span = int(gridspan[0].get(qn('w:val'))) if span > 1: cell_map[(i, j)] = ('hmerge', span) except Exception as e: print(f"警告：处理合并单元格时出错 [{i},{j}]: {str(e)}") # 第二遍：复制内容并执行合并 for i in range(rows): for j in range(cols): try: src_cell = table.cell(i, j) dst_cell = new_table.cell(i, j) # 检查是否需要合并 if (i, j) in cell_map: merge_type, span = cell_map[(i, j)] if merge_type == 'vmerge': # 垂直合并 for k in range(1, span): if i + k < rows: dst_cell.merge(new_table.cell(i + k, j)) elif merge_type == 'hmerge': # 水平合并 for k in range(1, span): if j + k < cols: dst_cell.merge(new_table.cell(i, j + k)) # 复制单元格属性 if src_cell._element.tcPr is not None: dst_cell._element.tcPr = deepcopy(src_cell._element.tcPr) # 复制单元格内容 dst_cell.text = "" # 清除默认内容 for src_paragraph in src_cell.paragraphs: dst_paragraph = dst_cell.add_paragraph() # 复制段落属性 if src_paragraph._element.pPr is not None: dst_paragraph._element.pPr = deepcopy(src_paragraph._element.pPr) # 复制文本和格式 for src_run in src_paragraph.runs: dst_run = dst_paragraph.add_run(src_run.text) # 复制运行属性 if src_run._element.rPr is not None: dst_run._element.rPr = deepcopy(src_run._element.rPr) except Exception as e: print(f"警告：复制单元格时出错 [{i},{j}]: {str(e)}") continue def _get_vmerge_value(self, cell_element) -> str: """ 获取单元格的垂直合并属性 Args: cell_element: 单元格元素 Returns: str: 垂直合并属性值 """ vmerge = cell_element.xpath('.//w:vMerge') if vmerge: return vmerge[0].get(qn('w:val'), 'continue') return None def _get_gridspan_value(self, cell_element) -> int: """ 获取单元格的水平合并数量 Args: cell_element: 单元格元素 Returns: int: 水平合并的列数 """ gridspan = cell_element.xpath('.//w:gridSpan') if gridspan: return int(gridspan[0].get(qn('w:val'), '1')) return 1 def _get_vertical_span(self, table: Table, start_row: int, col: int) -> int: """ 计算垂直合并的行数 Args: table: 表格对象 start_row: 起始行 col: 列号 Returns: int: 垂直合并的行数 """ span = 1 for i in range(start_row + 1, len(table.rows)): cell = table.cell(i, col) if self._get_vmerge_value(cell._element) == 'continue': span += 1 else: break return span def _extract_table_text(self, table: Table) -> str: """ 提取表格中的文本内容 Args: table: docx表格对象 Returns: str: 表格内容的文本表示 """ table_text = [] for row in table.rows: for cell in row.cells: cell_text = cell.text.strip() if cell_text: table_text.append(cell_text) return ' '.join(table_text) def process_directory(input_dir: str, output_dir: str = None): """ 处理指定目录下的所有文档文件 Args: input_dir: 输入目录路径 output_dir: 输出目录路径，如果为None则使用输入目录 """ # 如果未指定输出目录，使用输入目录 if output_dir is None: output_dir = input_dir if not os.path.exists(output_dir): os.makedirs(output_dir) cleaner = DocCleaner() for root, _, files in os.walk(input_dir): for file in files: if file.endswith(('.doc', '.docx')): input_path = os.path.join(root, file) try: # 清理文档 main_content, appendix, tables = cleaner.clean_doc(input_path) # 创建输出文件名（统一使用docx扩展名） base_name = os.path.splitext(file)[0] output_path = os.path.join(output_dir, f"{base_name}_cleaned.docx") # 保存为docx格式 cleaner.save_as_docx(main_content, appendix, tables, output_path) except Exception as e: print(f"处理文件 {file} 时出错: {str(e)}") # 添加更详细的错误信息 if isinstance(e, subprocess.CalledProcessError): print(f"命令执行错误: {e.output}") elif isinstance(e, FileNotFoundError): print("请确保已安装LibreOffice并将其添加到系统PATH中") def qn(tag: str) -> str: """ 将标签转换为带命名空间的格式 Args: tag: 原始标签 Returns: str: 带命名空间的标签 """ prefix = "{http://schemas.openxmlformats.org/wordprocessingml/2006/main}" return prefix + tag if __name__ == '__main__': import argparse parser = argparse.ArgumentParser(description='文档清理工具') parser.add_argument('input_dir', help='输入目录路径') parser.add_argument('--output_dir', help='输出目录路径（可选，默认为输入目录）', default=None) args = parser.parse_args() process_directory(args.input_dir, args.output_dir)