当前位置：首页 > web >正文

如何让AI自己检查全文？使用OCR和LLM实现自动“全文校订”（可DIY校订规则）

web 2025/6/27 12:40:41

详细流程及描述参见仓库（如果有用的话，请给个收藏）：

GitHub - xurongtang/DocRevision_Proj: A simple project about how to revist docment (such as your academic paper) in a automatic way with the help of OCR and LLM.A simple project about how to revist docment (such as your academic paper) in a automatic way with the help of OCR and LLM. - xurongtang/DocRevision_Projhttps://github.com/xurongtang/DocRevision_Proj

当完成很长的文字工作后，很害怕其中有明显的错别字或表述错误，但是又不想自己重头再读一遍怎么办？

我使用OCR和LLM做了一个自动检校脚本。其逻辑如下：

工作流设计如下：

from openai import OpenAI
import time
import yaml, json
import io
from PIL import Image
import base64
import fitz         # pip install pymupdf
from tqdm import tqdmdef get_apikey(path="apikey.yaml"):with open(path, 'r') as f:config = yaml.safe_load(f)res = config["apikey"]return resdef qwen_ocr(base64_image_code, addr_type):client = OpenAI(api_key=get_apikey()['dashscope'],base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",)completion = client.chat.completions.create(model="qwen-vl-ocr-2025-04-13",# model="qwen2.5-vl-72b-instruct",messages=[{"role": "user","content": [{"type": "image_url",# "image_url": "https://help-static-aliyun-doc.aliyuncs.com/file-manage-files/zh-CN/20241108/ctdzex/biaozhun.jpg","image_url": f"data:image/{addr_type};base64,{base64_image_code}","min_pixels": 28 * 28 * 4,"max_pixels": 1280 * 784},# 目前为保证识别效果，模型内部会统一使用"Read all the text in the image."作为text的值，用户输入的文本不会生效。{"type": "text","text": "Read all the text in the image"},]}])res_dict = json.loads(completion.model_dump_json())res_text = res_dict['choices'][0]['message']['content']return {"ocr_res": res_text}def qwen_max_repeat(content):client = OpenAI(api_key=get_apikey()['dashscope'],base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",)completion = client.chat.completions.create(model="qwen-max-2025-01-25",# model = "qwen3-235b-a22b",messages=[{"role": "system","content": """# 角色 # 你是一个专业的文字编辑，你需要对用户输入论文文本进行文字检查，以确保其准确。# 任务 #请逐句检查以下文本中的错别字、语法错误、用词不当及表达不清晰之处。按照指定格式输出结果：原文为："XXX"，其中"XXX"可能存在问题，可考虑修改为"XXX"。# 要求 # 只检查以下两类问题：1、同音异形字错误2、形近字错误# 限制 #1、确保每处问题独立呈现，不合并说明。2、不检查标点符号。3、只检测错误，不进行润色或改进。# 输出 # 只记录有问题的句子，并严格按模板输出结果，不输入任何其他的无关内容。"""},{"role": "user","content": f"""{content}"""}],max_tokens=4096,temperature=0.01,stream=False)res = json.loads(completion.model_dump_json())return {"response": res["choices"][0]["message"]["content"]}def main_process(base64_image_code,addr_type):content = qwen_ocr(base64_image_code, addr_type)out = qwen_max_repeat(content)return out["response"]def read_pdf2ImgLs(pdf_path) -> list:pdf = fitz.open(pdf_path)images_ls = []zoom_x = 2.0zoom_y = 2.0for i,pg in enumerate(pdf):mat = fitz.Matrix(zoom_x, zoom_y)pix = pg.get_pixmap(matrix=mat)img = Image.frombytes("RGB", [pix.width, pix.height], pix.samples)images_ls.append(img)return images_lsdef PILimage2base64(image):buffered = io.BytesIO()image_type = 'PNG'image.save(buffered, format=image_type)return base64.b64encode(buffered.getvalue()).decode(),image_typedef paper_revision(pdf_path,OUTPUT_PATH):# 设置输出txt路径# 设置名字命名为年月日时分output_txt = OUTPUT_PATHimage_ls = read_pdf2ImgLs(pdf_path)for page,image in enumerate(tqdm(image_ls, desc='Processing pages')):base64code,addr_type = PILimage2base64(image)repeat_response = main_process(base64code,addr_type)result = repeat_responsecleaned_string = result.strip('"')decoded_string = cleaned_string.replace('\\n', '\n').replace('\\\\', '\\')with open(output_txt, 'a', encoding='utf-8') as f:f.write(decoded_string+'\n')f.write(f'(page:{page+1})\n')if __name__ == '__main__':today = time.strftime("%Y-%m-%d-%H-%M", time.localtime())today = today.replace('-', '_')OUTPUT_PATH = f'{today}_output.txt'paper_revision('paper.pdf', OUTPUT_PATH)

其识别结果：