The structure of this framework is based on a survey [1] published by Markus Bayer et al. in which alot of different methods for data augmentation in the NLP subject were compared to eachother. A categorization of this methods was done too and this categorization builds the core structure.
There are four levels of augmentation methods in the data space. Following I will document them.
Character Level
Noise -> Introduces errors into data
Rule based -> Insertion of spelling mistakes, data alterations, entity names and abbreviations
Word Level
Noise
Unigram noising -> Replacing words by different random words
Blank noising -> Replacing words by "_"
Syntactic noise -> Shortening, alteration of adjectives
Semantic noise -> Lexical substitution of synonyms (See next point)
Random swap (EDA)
Random deletion (EDA)
Noise instead of zero-padding
TF-IDF -> Replace uninformative words by other uninformative
Synonyms
Page 9-11 of [1] delivers table with different replacement methods and synonym selections. Choose ones that are delivering positive results and implement 3-4
Embeddings
Page 12-14 similar table
Personally I still have problems totally grasping what this approach is exactly doing, so this will be interesting when I will go into details
Language Models
Generate similar words with embeddings, higlhy contextualized
Page 15 method table
Phrase Level
Structure
POS-Tagging
Cropping -> shorten sentences by putting focus und object and subject
Rotation -> Move flexible fragments
Semantic Text Exchange method
Interpolation
Substructure substitution -> Substitute substructures if same tagged label; 4 replacement rules that can be used in any combination
Document Level
Translation
Round-trip translation (RTT) -> Translate word, phrase, sentence or document into one language, then translate back -> Augmented data
Generative
Generate new data completely artificial
Pages 21-22 give some possibilities
This is the most complicated and new approach, so further information will be written down in an own issues
References
[1]
Markus Bayer, Marc-André Kaufhold, Christian Reuter (2021)
A Survey on Data Augmentation for Text Classification
djaszak
commented
2 years ago
The structure of this framework is based on a survey [1] published by Markus Bayer et al. in which alot of different methods for data augmentation in the NLP subject were compared to eachother. A categorization of this methods was done too and this categorization builds the core structure.
There are four levels of augmentation methods in the data space. Following I will document them.
References
[1] Markus Bayer, Marc-André Kaufhold, Christian Reuter (2021) A Survey on Data Augmentation for Text Classification