NLP enhanced technology

I met an old brother yesterday , Ask me what I know NLP Enhancement technology . I was stunned , Enhanced technology originated from The image processing , Later on 《 Baimian machine learning 》 I saw its detailed interpretation in this book .NLP Enhanced technology for ？ Actually , I used it before .
In the voice assistant , As input, expect , First, I enhanced the data of the definition . As the main model Fasttext, In fact, we also have data enhancement technology , So , I have also written several exploration summaries .
After you come back , I saw something about NLP Data enhancement technology , So I prepared to do some research .
This article mainly refers to Zhihu famous post 1

1. Word substitution
   1.1 Dictionary based substitution
   （1） Verb , Made into a small set; interjection 、 People and words have dictionaries , It can be replaced at will
   （2） Resource replacement is a difficult problem . such as , Children education This skill . Some resources are exclusive to children's education , Other resources do not belong to children's education , Therefore, attention should be paid to , Also add Counter sentences . And in this case , We also have words embedded Way to grasp .
   1.2. Word replacement based on word vector
   Never used , The pre training model is too big , This should not be easy to operate
   1.3 Based on mask language model （MLM augmentation）
   Difficult to operate , It's not easy to control
   1.4 be based on TF-IDF Replace with the word
   tf It is a word of high frequency ,IDF It is a word that distinguishes great power among articles . The original idea of this technique is Get rid of those Unimportant words .
2. Back translation
   This method is very useful in enhancing text similar data sets
3. Text form transformation
   This is specifically for To expand or abbreviate words
4. Random noise injection
   Based on a hypothesis ： A small amount of interference with the sample , The results predicted by the model are consistent
   4.1 Introduce spelling errors
   This is . We save the error results after speech recognition , As an alias for a formal resource name .
   4.2 Unigram The noise 、Blank The noise
   One is to join Useless high-frequency words ; One is to join Fixed symbol
   4.3 Other simple methods
   （1） Sentence order shuffle
   （2） Randomly exchange the order of two words
   （3） Insert words or sentences randomly
   （4） Random delete
5. Instance cross enhancement