最近在读一本神书，牛津大学出版的《Scientific Writing and Communication》，边阅读边检查书里面说的那些英文科学论文写作中经常出现的很俗的桥段，很多余的短语有没有出现在我正在写的科学论文中。检查的过程中发现书本第一章的内容非常有指导意义，只需要逐条搜索规则中出现的短语，删除或者替换即可。这样搜索替换了一阵后，感觉比较繁琐，寻思可以将规则收集起来，放在字典里，然后使用程序来全面检查论文内容。这样也可以帮到更多科研人员。花了一个晚上，将书中第一二章关于短语的规则搜集起来，制作了一个检查器。后面的章节更关注句子，段落和文章的结构，规则不容易建立，需要自然语言处理和深度学习训练。不过万事开头难，先迈出第一步，以后有时间慢慢填坑。

更新：书本前两章涵盖了第一到第六条原则, 建议删除冗余词，替换复杂和不准确的词。这些规则的实现收集在rule1to6_words.py中。最新版额外实现了位于第三章的第十条规则。这条规则判断一个语句是否头重脚轻，主语过长。为了判断主语的长度，程序使用了自然语言处理中的语法/依存分析。

更新：最新版本中加入了语法，单词拼写检查。

更新：加入对Latex文件的预处理。跳过方程，图片和表格，仅分析摘要，正文，图表描述等文字部分。LongGang Pang / proofread​gitlab.com

第一部分：简介

使用方法：

git clone https://gitlab.com/snowhitiger/proofread.git

cd proofread

python proofread.py path/*.tex

结果示例：

在默认情况下，proofread.py 仅对规则字典中要求删除或替换的词进行检查。

如果想要开启其他功能，可以在命令行添加选项。使用如下命令查看可用的选项，

python proofread.py --help

输出结果如下usage: proofread.py [-h] [--report_long_subjects] [--report_interruptions]

[--report_noun_clusters] [--check_grammars] [--brief]

f [f ...]

Proofread scientific papers to make it brief and simple

positional arguments:

f one file or a list of files for analysis

optional arguments:

-h, --help 显示帮助信息并退出

--report_long_subjects Principle 10: 避免过长主语，默认未开启

--report_interruptions Principle 11: 避免主语谓语距离过长，默认未开启

--report_noun_clusters Principle 18: 避免长串名词，默认未开启

--check_grammars 语法，拼写检查，默认未开启

示例，如果想对文件 example.tex 开启语法检查, 并报告过长的主语，使用

python proofread.py example.tex --report_long_subjects --check_grammars

请注意，语法检查速度比较慢，请谨慎开启。最新版本会对Latex文件自动分析，跳过方程，图，表格等与文字无关的内容。

替换字典：

这本书的本意是让我们在大脑中记住这些规则，潜移默化的改变我们的写作过程，而不是让我们写完之后修改。这里把搜集的字典文件放出来，大家可以先记在脑中，写作的过程中尽量避免。如果不想费力，或者不想按字典一条条检查自己学生写的初稿，就直接使用上面的 proofread.py 脚本来自动化搜索/建议修改过程。字典是一个 json 文件，第一个版本只有一些手动放进去的规则，

{"to_remove": ["actually",

"appear",

"absolutely",

"at all times",

"commonly",

"clearly",

"certainly",

"completely",

"much",

"usually",

"perhaps",

"possibly",

"definitely",

"basically",

"essentially",

"fairly",

"very",

"just",

"really",

"seem",

"practically",

"quite",

"rather",

"virtually",

"as it is well known",

"for all intents and purposes",

"it is well known",

"it is shown that",

"it is shown to",

"it is observed that",

"it is reasonable to assume that",

"evidence is presented that shows that",

"it is speculated that",

"it is found that",

"it is demonstrated that",

"it is reported that",

"it has long been known that",

"is observed to",

"is determined to",

"note that",

"needless to say",

"it is noted that",

"it is also worth noting",

"it is worth noting",

"respectively",

"so-called",

"putative",

"in shape",

"in a previous study, it is demonstrated that",

"it is important to realize",

"it is important to note",

"i believe",

"we believe",

"in my opinion",

"in our opinion",

"there are many papers stating"],

"to_replace": {

"a large majority of":"most",

"a very large number of":"many",

"a large number of":"many",

"a considerable number of":"many",

"an adequate amount of":"enough",

"are in agreement":"agree",

"an example of this is the fact that":"for example",

"as a consequence of":"because",

"at no time":"never",

"at the present time":"at present",

"already existing":"existing",

"all of the":"all the",

"as to whether":"whether",

"basic fundamentals":"fundamentals",

"blue in color":"blue",

"based on the fact that":"because",

"by means of":"by",

"considerable amount of":"much",

"cold temperature":"cold",

"currently underway":"underway",

"complete eliminate":"eliminate",

"conduct research":"research",

"make a decision":"decide",

"do not overlook":"note",

"despite the fact that":"although",

"due to the fact that":"due to",

"during the time that":"while/when",

"employ":"use",

"elucidate":"show",

"etiology":"cause",

"each individual":"each",

"end result":"result",

"evolve over time":"evolve",

"first of all":"first",

"for the purpose of":"to",

"has the capability of":"can/is able to",

"has the capacity to":"can",

"Has the ability to":"can",

"in a careful manner":"carefully",

"in light of the fact that":"because",

"in many cases":"often",

"in order to":"to",

"in some cases":"sometimes",

"in the absence of":"without",

"in the event that":"if",

"in terms of":"regarding",

"in view of the fact that":"because/since",

"it is interest to note that":"note that",

"it is often the case that":"often",

"in the event that":"if",

"in the process of":"while/when",

"it is possible that":"restructure using can/could/may/might",

"the fact that":"delete and restructure",

"may indicate":"suggest",

"majority of":"most",

"methodology":"method",

"no later than":"by",

"not different":"similar",

"not infrequently":"frequently",

"not many":"few",

"not the same":"different",

"not unimportant":"important",

"on the basis of":"by",

"referred to as":"called",

"regardless of the fact that":"even though",

"so as to":"to",

"small in size":"small",

"large in size":"large",

"take advantage of":"use",

"take into consideration":"consider",

"make a decision":"decide",

"the question as to whether":"whether",

"there is no doubt but that":"doubtless",

"this is a subject that":"this subject",

"here is":"Delete and restructure to create a stronger active subject/verb.",

"ultilize":"use",

"used for * purposes":"used for *",

"utilization":"use",

"with reference to":"about (or omit)",

"with regard to":"about (or omit)",

"with the exception of":"except",

"whether or not":"whether",

"prior to":"before",

"past history":"past",

"plays a key role in":"is essential to",

"prioritized":"given priority",

"subsequent to":"after",

"performed an analysis":"analyzed",

"period of time":"period",

"at this point in time":"now",

"due to the fact that":"because",

"in the event that":"if",

"a new initiative":"an initiative",

"nearly unique":"unique/rare",

"both * were equally affected":"the * were equally affected",

"in spite of the fact that":"though",

"owing to the fact that":"because",

"enhanced":"increased",

"man":"humans",

"quite sufficiently large":"large",

"reflect deviations":"deviate",

"estimated roughly at":"estimated at",

"first and for most":"first",

"final outcome":"outcome",

"frequently":"make it more precise, e.g., frequently ---> every 12 hours",

"future plans":"plans",

"main essentials":"essentials",

"still persists":"persists",

"never before":"never",

"reason is because":"reason is",

"situation":"can be omitted, perhaps",

"several":"make it more precise, e.g., several hours ---> 6 hours",

"true facts":"facts",

"the nature of":"can be omitted, perhaps"},

"other_rules":{"there is":"Delete and restructure to create a stronger active subject/verb. E.g., There have been many discussions among the scientific community about ethical boundaries in gene-splicing research. → The scientific community has frequently discussed the ethical boundaries in gene-splicing research.",

"even":"sometimes can omit",

"often":"sometimes can omit",

"with respect to":"about"}

}

第二部分：实现方法与思路

基于规则的替换建议

这部分使用规则匹配来建议删除多余词，替换不准确词。举例说明，

1. "As it is well known"：建议删除。

2. "It is shown that": 建议删除。

3. "is related to": 建议修改成 "is associated with"。

4. "provides a review": 建议修改成 "reviews"。

这些规则会有很多变体，基于规则的部分会尽量考虑所有的可能并提出相应的替换方案。比如，在字典中只会出现 “As it is well known“，程序中会通过数据增广，生成新的规则，如 "As it was well know" , "As it has been well know" , "As it have been well know" , "As it had been well know" 。

rule1to6_words.py 实现了所有词的替换和删除建议。直接运行 proofread.py 会默认使用定义在此脚本中的所有规则。

基于自然语言处理的语法判断

避免主语太长或者主语之前有过多的限定词

写作的过程应尽量使主语简短，接近谓语，否则会带来理解上的困难。这是《Scientific Writing And Communication》书中提到的第10条科学写作原则。

举例说明：

反例：Due to the nonlinear and hence complex nature of ocean currents, modeling these currents in the tropical pacific is difficult.

正例： Modeling ocean currents in the tropical Pacific is difficult due to their nonlinear and hence complex nature.

目前使用Spacy库对文章的每句话进行语法分析，找到主语谓语。spacy 库是与nltk类似的自然语言处理 python 库。如果谓语之前的部分过长，影响理解，则提出修改请求。具体实现在 rule10_long_subject.py 中。

proofread.py 默认不会对主语过长进行检查，需要在命令行中加入--report_long_subjects选项开启，

python proofread.py example.txt --report_long_subjects

如果 Spycy 库没有安装，第一次开启了 --report_long_subjects 之后，python会自动调用命令 pip install spycy --user 安装spycy 并下载其需要的英文语法库。

基于自然语言处理的语义分析

ToDo

对于单词和短语来说，有些需要语义判断的地方，可能会用到额外的自然语言处理技术。比如 since 一般被用作与时间相关的语句，但有时候它也会被用作 “because”，如果我们想将所有与时间无关的用例都替换成 "because" 或者 "as", 那么就需要使用一定的机器学习或人工智能，来判断语句是否表达了自某一时刻起。这个时候，如果使用上下文相关的词向量，当可判断“since”是否有时间的分量在里面。(注：此处BERT的词向量不可用，因为有位置嵌入)。

对于句法，段落以及文章结构，自然语言处理的方法必不可少。比如下面这些规则，将新的信息放在语句的最后

审慎使用消极或被动语态

确保段落的第一句与最后一句匹配

将名词换成动词，使语句看起来更加积极

这些规则的构建都需要对语义的理解和判断。比如如何判断一段话中哪句话引入了新的信息，哪里的消极或被动语态可以被同义词替换。可能用到的方法有简单的隐马尔可夫链/条件随机场，复杂点的RNN，或者是更复杂的BERT。虽然困难，但也不是做不了，随着自然语言理解的发展，这些有关句法段落的规则可能也会慢慢的被构建进代码。

对于句法和段落来说，最难的是训练语料的获取。如果预训练的语言模型，如BERT，能够将微调时需要的训练语料降低到10个或100个左右，那么就可以使用少量的样例，来训练一个神经网络，一条条履行这些事关语法和段落的规则。

规则资源