Automatic text normalisation

General ideas

• Diacritic restoration
• Reduplicated character reduction
  • How to learn language specific settings? -- e.g. in English certain consonants can double, but others cannot, same goes for vowels. Can we learn these by looking at a corpus ?

Code switching

• For the language subpart... we can actually train and keep copies of most frequently corrected words across languages and then refer to that list...
  • Maybe this will be too heavy for the on the run application ( needs discussion )
• Is it possible to identify sub-spans of text ? e.g.
  • LOL rte showin dáil in irish 4 seachtan na gaeilge, an ceann comhairle hasnt a scooby wots bein sed! his face is classic ha!
  • [en LOL rte showin dáil in irish 4] [ga seachtan na gaeilge, an ceann comhairle] [en hasnt a scooby wots bein sed! his face is classic ha!]
• Ideas: You will rarely have single word spans of X-Y-X-Y-X-Y e.g. "la family está in la house." "la família está in the house." is probably a more frequent structure.
  • So we can probably do this to a certain extent LR in a single pass.
  • We probably shouldn't consider a single word code switching, but perhaps a span of 2-3+
  • It's like a state machine, you are in state "en", and you see something that makes you flip to state "ga", then you see another thing that makes you flip to state "en".
  • It could also be that at some point you are not sure, so what you should do is keep both options open, e.g. you would keep adding to en/ga.

To do list

• Feed charlifter with n-grams ( works best with a trigram model ). This would improve the diacritics at the moment
• Make list of most frequently occurring non dictionary words (ga), these might be abbreviations. Check for these words
• add most frequently occuring english abbreviations to the list
• From Comments = tu -> tú, not tuilleadh
• change some_known capitals for diff. languages
• suggestions for including spelling correction
  • Example, Taisbeánta should be Taispeána
  • repetitions haha hehe can be included for this as well
  • thought for such a single repitition
  • should all replacements go through n-gram verification?
  • Words like CAP REM mean should stay the same. I'm over-reaching because of the trie implementation.. need to weigh down
  • Scope for addition of rules... vowels are not repeated
  • mhoiiiiiilllllll -> mhoill is the correct form.. I got mhoil.. Will have to look in the ngram model for this
  • only characters which get repeated are ll nn rr