[Feature]: 一种更好的LaTeX翻译方式，有望实现支持ChatGLM。已经完成部分代码，需要帮助

[azwphy]

Class | 类型

函数插件

Feature Request | 功能请求

基本思路是：用pylatexenc包的parser工具将LaTeX代码转为抽象语法树（AST），通过遍历整个抽象语法树找到需要翻译的纯文本，并记录其在AST中的位置（pos）。把文本丢给ChatGLM翻译，再将得到的代码重新构建为latex代码，导言区加入ctex包，直接编译即可。

尝试开发中的插件代码如下，其中pylatexenc==3.0a19:

'crazy_functions/latex_fns/latex_parser_ver.py'

from toolbox import update_ui, update_ui_lastest_msg    # 刷新Gradio前端界面
from toolbox import zip_folder, objdump, objload, promote_file_to_downloadzone
from .latex_toolbox import PRESERVE, TRANSFORM
from .latex_toolbox import set_forbidden_text, set_forbidden_text_begin_end, set_forbidden_text_careful_brace
from .latex_toolbox import reverse_forbidden_text_careful_brace, reverse_forbidden_text, convert_to_linklist, post_process
from .latex_toolbox import fix_content, find_main_tex_file, merge_tex_files, compile_latex_with_timeout
from pylatexenc.latexnodes import *
from pylatexenc.latexwalker import *
import os, shutil
import re, itertools
import numpy as np
from ..crazy_utils import request_gpt_model_in_new_thread_with_ui_alive

pj = os.path.join
white_macroname_list = [
    "emph",
    "textbf"
]
def parse_latex(latex_string):
   # 创建一个LatexWalker对象，用于解析LaTeX代码
   latex_walker = LatexWalker(latex_string)
   # 获取解析后的AST
   nodelist, parsing_state_delta = latex_walker.parse_content(parsers.LatexGeneralNodesParser())
   npos = nodelist.pos
   nlen = nodelist.len
   return nodelist, npos, nlen

   # 解析LaTeX代码

def begining_of_document(nodes):
    for node in nodes:
        if isinstance(node, LatexEnvironmentNode):
            if node.environmentname == "document":
                return node

def get_note_from_nodes(nodes, res_list):
    # 遍历AST
    if isinstance(nodes, list):
        for node in nodes:
            get_note_from_nodes(node, res_list)
    elif isinstance(nodes, LatexEnvironmentNode) or isinstance(nodes, LatexGroupNode):
        for child in nodes.nodelist:
            get_note_from_nodes(child, res_list)
    elif isinstance(nodes, LatexMacroNode):
        if nodes.macroname in white_macroname_list:
            get_note_from_nodes(nodes.nodeargd, res_list)
    elif isinstance(nodes, ParsedArguments):
        get_note_from_nodes(nodes.argnlist, res_list)
    elif isinstance(nodes, LatexCharsNode):
        if not "_" in nodes.chars:
            res_list.append([(nodes.chars).replace("\n", ""), nodes.pos, 0, 0])

def split_text_to_words(res_list, words):
    for line in res_list:
       currect_line = line[0].split()
       currect_line = [[x, line[1], line[2], line[3]] for x in currect_line]
       for i in currect_line:
            words.append(i)
    for index, element in enumerate(words):
        element[2] = index
    return words

def split_counted_res(split_res_list, max_limited_word=200):
    last_period_index = 0
    index = 0
    token_counter = 0
    line_counter = 0
    while True:
        if index == len(split_res_list):
            break
        split_res_list[index][3] = line_counter
        if "." in split_res_list[index][0]:
            last_period_index = index
        if token_counter > max_limited_word:
            line_counter += 1
            token_counter = 0
            index = last_period_index
        index += 1
        token_counter += 1
    return split_res_list

def combine_by_pos(finally_counted_res):
    combine_counted_res = []
    mid_res = [list(v) for k, v in itertools.groupby(finally_counted_res, key=lambda x: x[3])]
    for element in mid_res:
        combine_counted_res.append([list(v) for k, v in itertools.groupby(element, key=lambda x: x[1])])
    # 最后再合并一次
    finally_combine_counted_res = []
    for element in combine_counted_res:
        midmid_res = []
        for element2 in element:
            midmid_res.append([' '.join([x[0] for x in element2]), element2[0][1]])
        finally_combine_counted_res.append(midmid_res)
    return finally_combine_counted_res

def translate_func(combine_counted_res, project_folder, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt, mode='proofread', switch_prompt=None, opts=[]):
    # 翻译函数
    #  <-------- 根据需要切换prompt ----------> 
    inputs_array = []
    for element in combine_counted_res:
        for element2 in element:
            prompt = ""
            # prompt += "Context:\n"
            # for element3 in element:
            #     prompt += element3[0]
            prompt += "Below is a section from an academic paper, translate this section to Chinese."
            prompt += "Answer me only with the revised text:"
            prompt += element2[0]
            inputs_array.append([prompt, element2[1]])
    length_of_inputs_array = len(inputs_array)
    sys_prompt = "You are a professional translator."
    #  <-------- gpt 单线程请求 ----------> 
    gpt_response_collection = []
    for element in inputs_array:
        gpt_say = yield from request_gpt_model_in_new_thread_with_ui_alive(
            inputs=element[0],             # 提问的内容，给chatgpt看的
            inputs_show_user=element[0],   # 提问的内容，给用户看的（可以隐藏啰嗦的细节）
            llm_kwargs=llm_kwargs,    # 无聊的chatgpt内部参数
            chatbot=chatbot,          # 聊天框句柄，原样传递
            history=[],               # 之前的聊天内容，只有之前的聊天内容中有值得抽取的信息时，才是必要的
            sys_prompt=sys_prompt
            )
        gpt_response_collection.append([gpt_say,element[1]])
    return gpt_response_collection
    

def return_note_from_nodes(nodes, trans_res_list):
    # 遍历AST
    if isinstance(nodes, list):
        for node in nodes:
            return_note_from_nodes(node, trans_res_list)
    elif isinstance(nodes, LatexEnvironmentNode) or isinstance(nodes, LatexGroupNode):
        for child in nodes.nodelist:
            return_note_from_nodes(child, trans_res_list)
    elif isinstance(nodes, LatexMacroNode):
        if nodes.macroname in white_macroname_list:
            return_note_from_nodes(nodes.nodeargd, trans_res_list)
    elif isinstance(nodes, ParsedArguments):
        return_note_from_nodes(nodes.argnlist, trans_res_list)
    elif isinstance(nodes, LatexCharsNode):
        for i in trans_res_list:
            if i[1] == nodes.pos:
                i[1] = nodes.chars

def ProcessLaTeXMain(file_manifest, project_folder, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt, mode='proofread', switch_prompt=None, opts=[]):
    import time, os, re
    from .latex_actions import LatexPaperFileGroup, LatexPaperSplit

    #  <-------- 寻找主tex文件 ----------> 
    maintex = find_main_tex_file(file_manifest, mode)
    chatbot.append((f"定位主Latex文件", f'[Local Message] 分析结果：该项目的Latex主文件是{maintex}, 如果分析错误, 请立即终止程序, 删除或修改歧义文件, 然后重试。主程序即将开始, 请稍候。'))
    yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
    time.sleep(3)

    #  <-------- 读取Latex文件, 将多文件tex工程融合为一个巨型tex ----------> 
    main_tex_basename = os.path.basename(maintex)
    assert main_tex_basename.endswith('.tex')
    main_tex_basename_bare = main_tex_basename[:-4]
    may_exist_bbl = pj(project_folder, f'{main_tex_basename_bare}.bbl')
    if os.path.exists(may_exist_bbl):
        shutil.copyfile(may_exist_bbl, pj(project_folder, f'merge.bbl'))
        shutil.copyfile(may_exist_bbl, pj(project_folder, f'merge_{mode}.bbl'))
        shutil.copyfile(may_exist_bbl, pj(project_folder, f'merge_diff.bbl'))

    with open(maintex, 'r', encoding='utf-8', errors='replace') as f:
        content = f.read()
        merged_content = merge_tex_files(project_folder, content, mode)

    with open(project_folder + '/merge.tex', 'w', encoding='utf-8', errors='replace') as f:
        f.write(merged_content)
    # 读取文件
    with open(project_folder + '/merge.tex', 'r', encoding='utf-8') as f:
        latex_code = f.read()
    # 解析LaTeX代码
    nodes, _, _ = parse_latex(latex_code)
    trnode= list(nodes)
    res_list = []
    get_note_from_nodes(begining_of_document(trnode), res_list)
    res_list = sorted(res_list, key=lambda x: x[2])
    res_list = [x for x in res_list if x[0] != ""]
    words = []
    split_res_list = split_text_to_words(res_list, words)
    finally_counted_res = split_counted_res(split_res_list)
    finally_combine_counted_res = combine_by_pos(finally_counted_res)
    gpttranres = translate_func(finally_combine_counted_res, project_folder, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt, mode='proofread', switch_prompt=None, opts=[])
    chatbot.append(str(gpttranres))
    trans_res_list = list(itertools.chain.from_iterable(gpttranres))
    # 生成LaTeX代码
    return_note_from_nodes(trnode, trans_res_list)
    with open(project_folder + f'/merge_{mode}.tex', 'w', encoding='utf-8', errors='replace') as f:
        if mode != 'translate_zh' or "binary" in final_tex: f.write(trans_res_list)
    #  <-------- 整理结果, 退出 ----------> 
    chatbot.append((f"完成了吗？", 'GPT结果已输出, 即将编译PDF'))
    yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
    #  <-------- 返回 ----------> 
    return project_folder + f'/merge_{mode}.tex'

crazy_functions/TeX论文翻译

import os
from toolbox import update_ui, trimmed_format_exc, get_conf, objdump, objload, promote_file_to_downloadzone
from toolbox import CatchException, report_execption, update_ui_lastest_msg, zip_result, gen_time_str
from functools import partial
from pylatexenc.latexnodes import *
from pylatexenc.latexwalker import *
from .latex_fns.latex_parser_ver import *
import glob, os, requests, time
pj = os.path.join
def import_requirements():
    try:
        from pylatexenc.latexwalker import LatexWalker, LatexCharsNode # 尝试导入依赖
    except:
        # 如果缺少依赖fitz，则给出安装建议
        report_execption(chatbot, history, 
            a = f"解析项目: {txt}", 
            b = f"导入软件依赖失败。使用该模块需要额外依赖，安装方法```pip install --upgrade pylatexenc```。")
        yield chatbot, history, '正常'
        return
ARXIV_CACHE_DIR = os.path.expanduser(f"~/arxiv_cache/")

# =================================== 工具函数 ===============================================
专业词汇声明  = 'If the term "agent" is used in this section, it should be translated to "智能体". '
def switch_prompt(translate_list, mode, more_requirement):
    # 切换prompt
    """
    Generate prompts and system prompts based on the mode for proofreading or translating.
    Args:
    - translate_list: Proofreader or Translator instance.
    - mode: A string specifying the mode, either 'proofread' or 'translate_zh'.

    Returns:
    - inputs_array: A list of strings containing prompts for users to respond to.
    - sys_prompt_array: A list of strings containing prompts for system prompts.
    """
    n_split = len(translate_list)
    if mode == 'proofread_en':
        inputs_array = [r"Below is a section from an academic paper, proofread this section." + 
                        r"Do not modify any latex command such as \section, \cite, \begin, \item and equations. " + more_requirement +
                        r"Answer me only with the revised text:" + 
                        f"\n\n{frag}" for frag in pfg.sp_file_contents]
        sys_prompt_array = ["You are a professional academic paper writer." for _ in range(n_split)]
    elif mode == 'translate_zh':
        inputs_array = [r"Below is a section from an English academic paper, translate it into Chinese. " + more_requirement + 
                        r"Do not modify any latex command such as \section, \cite, \begin, \item and equations. " + 
                        r"Answer me only with the translated text:" + 
                        f"\n\n{frag}" for frag in pfg.sp_file_contents]
        sys_prompt_array = ["You are a professional translator." for _ in range(n_split)]
    else:
        assert False, "未知指令"
    return inputs_array, sys_prompt_array

def desend_to_extracted_folder_if_exist(project_folder):
    """ 
    Descend into the extracted folder if it exists, otherwise return the original folder.

    Args:
    - project_folder: A string specifying the folder path.

    Returns:
    - A string specifying the path to the extracted folder, or the original folder if there is no extracted folder.
    """
    maybe_dir = [f for f in glob.glob(f'{project_folder}/*') if os.path.isdir(f)]
    if len(maybe_dir) == 0: return project_folder
    if maybe_dir[0].endswith('.extract'): return maybe_dir[0]
    return project_folder

def move_project(project_folder, arxiv_id=None):
    """ 
    Create a new work folder and copy the project folder to it.

    Args:
    - project_folder: A string specifying the folder path of the project.

    Returns:
    - A string specifying the path to the new work folder.
    """
    import shutil, time
    time.sleep(2)   # avoid time string conflict
    if arxiv_id is not None:
        new_workfolder = pj(ARXIV_CACHE_DIR, arxiv_id, 'workfolder')
    else:
        new_workfolder = f'gpt_log/{gen_time_str()}'
    try:
        shutil.rmtree(new_workfolder)
    except:
        pass

    # align subfolder if there is a folder wrapper
    items = glob.glob(pj(project_folder,'*'))
    if len(glob.glob(pj(project_folder,'*.tex'))) == 0 and len(items) == 1:
        if os.path.isdir(items[0]): project_folder = items[0]

    shutil.copytree(src=project_folder, dst=new_workfolder)
    return new_workfolder

def arxiv_download(chatbot, history, txt):
    def check_cached_translation_pdf(arxiv_id):
        translation_dir = pj(ARXIV_CACHE_DIR, arxiv_id, 'translation')
        if not os.path.exists(translation_dir):
            os.makedirs(translation_dir)
        target_file = pj(translation_dir, 'translate_zh.pdf')
        if os.path.exists(target_file):
            promote_file_to_downloadzone(target_file, rename_file=None, chatbot=chatbot)
            return target_file
        return False
    def is_float(s):
        try:
            float(s)
            return True
        except ValueError:
            return False
    if ('.' in txt) and ('/' not in txt) and is_float(txt): # is arxiv ID
        txt = 'https://arxiv.org/abs/' + txt.strip()
    if ('.' in txt) and ('/' not in txt) and is_float(txt[:10]): # is arxiv ID
        txt = 'https://arxiv.org/abs/' + txt[:10]
    if not txt.startswith('https://arxiv.org'): 
        return txt, None
    
    # <-------------- inspect format ------------->
    chatbot.append([f"检测到arxiv文档连接", '尝试下载 ...']) 
    yield from update_ui(chatbot=chatbot, history=history)
    time.sleep(1) # 刷新界面

    url_ = txt   # https://arxiv.org/abs/1707.06690
    if not txt.startswith('https://arxiv.org/abs/'): 
        msg = f"解析arxiv网址失败, 期望格式例如: https://arxiv.org/abs/1707.06690。实际得到格式: {url_}"
        yield from update_ui_lastest_msg(msg, chatbot=chatbot, history=history) # 刷新界面
        return msg, None
    # <-------------- set format ------------->
    arxiv_id = url_.split('/abs/')[-1]
    if 'v' in arxiv_id: arxiv_id = arxiv_id[:10]
    cached_translation_pdf = check_cached_translation_pdf(arxiv_id)
    if cached_translation_pdf: return cached_translation_pdf, arxiv_id

    url_tar = url_.replace('/abs/', '/e-print/')
    translation_dir = pj(ARXIV_CACHE_DIR, arxiv_id, 'e-print')
    extract_dst = pj(ARXIV_CACHE_DIR, arxiv_id, 'extract')
    os.makedirs(translation_dir, exist_ok=True)
    
    # <-------------- download arxiv source file ------------->
    dst = pj(translation_dir, arxiv_id+'.tar')
    if os.path.exists(dst):
        yield from update_ui_lastest_msg("调用缓存", chatbot=chatbot, history=history)  # 刷新界面
    else:
        yield from update_ui_lastest_msg("开始下载", chatbot=chatbot, history=history)  # 刷新界面
        proxies, = get_conf('proxies')
        r = requests.get(url_tar, proxies=proxies)
        with open(dst, 'wb+') as f:
            f.write(r.content)
    # <-------------- extract file ------------->
    yield from update_ui_lastest_msg("下载完成", chatbot=chatbot, history=history)  # 刷新界面
    from toolbox import extract_archive
    extract_archive(file_path=dst, dest_dir=extract_dst)
    return extract_dst, arxiv_id
# ========================================= 插件主程序 =====================================================    
@CatchException
def 更好的Latex翻译中文并重新编译PDF(txt, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt, web_port):
    # <-------------- information about this plugin ------------->
    chatbot.append([
        "函数插件功能？",
        "对整个Latex项目进行翻译, 生成中文PDF。函数插件贡献者: Binary-Husky, azwphy。"])
    yield from update_ui(chatbot=chatbot, history=history) # 刷新界面

    # <-------------- more requirements ------------->
    if ("advanced_arg" in plugin_kwargs) and (plugin_kwargs["advanced_arg"] == ""): plugin_kwargs.pop("advanced_arg")
    more_req = plugin_kwargs.get("advanced_arg", "")
    _switch_prompt_ = partial(switch_prompt, more_requirement=more_req)

    # <-------------- check deps ------------->
    import_requirements()
    try:
        import glob, os, time, subprocess, pylatexenc, itertools
        subprocess.Popen(['pdflatex', '-version'])
    except Exception as e:
        chatbot.append([ f"解析项目: {txt}",
            f"尝试执行Latex指令失败。Latex没有安装, 或者不在环境变量PATH中。安装方法https://tug.org/texlive/。报错信息\n\n```\n\n{trimmed_format_exc()}\n\n```\n\n"])
        yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
        return
    # <-------------- clear history and read input ------------->
    txt, arxiv_id = yield from arxiv_download(chatbot, history, txt)
    if txt.endswith('.pdf'):
        report_execption(chatbot, history, a = f"解析项目: {txt}", b = f"发现已经存在翻译好的PDF文档")
        yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
        return
    

    # <-------------- clear history and read input ------------->
    history = []
    if os.path.exists(txt):
        project_folder = txt
    else:
        if txt == "": txt = '空空如也的输入栏'
        report_execption(chatbot, history, a = f"解析项目: {txt}", b = f"找不到本地项目或无权访问: {txt}")
        yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
        return
    file_manifest = [f for f in glob.glob(f'{project_folder}/**/*.tex', recursive=True)]
    if len(file_manifest) == 0:
        report_execption(chatbot, history, a = f"解析项目: {txt}", b = f"找不到任何.tex文件: {txt}")
        yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
        return

    # <-------------- if is a zip/tar file ------------->
    project_folder = desend_to_extracted_folder_if_exist(project_folder)


    # <-------------- move latex project away from temp folder ------------->
    project_folder = move_project(project_folder, arxiv_id=None)


    # <-------------- if merge_translate_zh is already generated, skip gpt req ------------->
    if not os.path.exists(project_folder + '/merge_translate_zh.tex'):
        res_test_file = yield from ProcessLaTeXMain(file_manifest, project_folder, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt)
    
    promote_file_to_downloadzone(file=res_test_file, chatbot=chatbot)
    # # <-------------- compile PDF ------------->
    # success = yield from 编译Latex(chatbot, history, main_file_original='merge', main_file_modified='merge_proofread_en', 
    #                          work_folder_original=project_folder, work_folder_modified=project_folder, work_folder=project_folder)
    

    # <-------------- zip PDF ------------->
    # zip_res = zip_result(project_folder)
    # if success:
    #     chatbot.append((f"成功啦", '请查收结果（压缩包）...'))
    #     yield from update_ui(chatbot=chatbot, history=history); time.sleep(1) # 刷新界面
    #     promote_file_to_downloadzone(file=zip_res, chatbot=chatbot)
    # else:
    #     chatbot.append((f"失败了", '虽然PDF生成失败了, 但请查收结果（压缩包）, 内含已经翻译的Tex文档, 也是可读的, 您可以到Github Issue区, 用该压缩包+对话历史存档进行反馈 ...'))
    #     yield from update_ui(chatbot=chatbot, history=history); time.sleep(1) # 刷新界面
    #     promote_file_to_downloadzone(file=zip_res, chatbot=chatbot)

    # # <-------------- we are done ------------->
    # return success

请注意，部分代码未完成。遇到的主要问题为：

每个单词都有一个pos位置信息，所以必须在翻译后的片段也包含这种位置信息。例如，\emph{}用于加粗时会将一句话分割为三个部分。如果要保证翻译后的文本还能在适当的位置加粗，则必须将这句话的每一个片段都有对应的翻译（而不是有整段话的翻译）。为了保持上下文的连贯性，我试图使用这样一种方法：按照token数限制将文本分割为片段，每一段都以句点结束。把整段话提供给GLM,并让它只翻译其中的某一片段。把每一个片段都重复这个操作。但无论怎么写prompt,GLM似乎都无法理解我的意图。所以写issue向各位求助。

最后还有关于pylatexenc包的一个问题。该包不支持将针对AST的修改重构为LaTeX代码。我能想到的替代方案为，按照文本顺序对收集得到的英文文本逐一替换为相应的中文文本。这个思路应该是比较可行的。最主要的问题还是如何连贯的翻译。

[binary-husky]

AAAAA\emph{BBBBBBB}CCCCCCCCC这种情况确实不太容易处理

[binary-husky]

我之前的一个思路时借助pandoc实现ast，这样做出来的代码可以复用到word，ppt等各种文档上

不过需要解决的难点太多，需要花大量时间，只能暂缓

[azwphy]

我之前的一个思路时借助pandoc实现ast，这样做出来的代码可以复用到word，ppt等各种文档上

不过需要解决的难点太多，需要花大量时间，只能暂缓

pandoc感觉还是有些局限性，LaTeX转Markdown容易，但翻译后很难转回LaTeX
如果只需要翻译结果的话，或许pandoc也行？

[reonokiy]

我的想法是能否按AST分割到段（分割到小于GPT的长度限制）之后直接交给GPT处理，处理完成后验证LaTeX语法的正确性，错误就回退到纯文本模式。

这里是我直接引导GPT输出内容的一个例子，不知道符不符合要求：

这个例子只有第一段话是我直接输入的，后续它能够知道每段话中的LaTeX代码并且不翻译（尤其是对于emph进行了翻译而cite不翻译）

另一个例子是使用Few Shot，想法是可以直接引导GPT输出JSON

在这个例子里，第一个要求以及hello，meaning两个单词的回复都由我手动撰写。接下来在输入英文单词，它能够正确输出JSON格式的回复。

相关链接：

• Zero, One and Few Shot Prompting ChatGPTNextWeb/ChatGPT-Next-Web#138 (comment)