Word Level Synonyms

[djaszak]

Synonym replacement should be the core of my framework, as IMO this is the most interesting as easiest approach. We want to approach 3 different layers:

• Synonym Database
• Replacement Method
• Synonym Selection

For these layers I am choosing 3-4 method proven good in previous work, stated by Table 2 in [1].

Synonym Database

• Wordnet and its python API
• Thesaurus

Replacement Method

• Substitutable words are nouns, verbs, adjectives, or adverbs that are not part of a named entity. Each word is replaced with a certain probability.
• Only adverbs/adjectives, sometimes nouns, more rarely verbs
• No time words, prepositions and mimetic words
• No stop words. n random words are replaced (SR) or synonyms are insterted at random position (RI) The last point is just referncing to 2 techniques defining EDA. All 4 defining EDA should be implemented and used in different behaviours but for this issue, the ones referencing on synonyms should be implemented.

  Synonym Selection

• Remaining probabilty shared among synonyms based on language model score
• Uniform random
• Chi-square statistics method (TBD)

[1] Markus Bayer, Marc-André Kaufhold, Christian Reuter (2021) A Survey on Data Augmentation for Text Classification

[djaszak]

FIrst implementation in be40fbd9c978f3357b5a4a826a11d938d2ece3be . Next steps:

• Get rid of whitespaces around special characters
• Implement different replacement methods as well as synonym selection