Toolkit for dictionary development

We all do similar things when making dictionaries. Make a load of scripts that we hack to do a specific job, then throw them away a the end. It would be nice to have some maintained scripts that could lighten the load.

When we're making a new dictionary, what resources do we have and use ?

• Descriptive grammar of some variety (best case, reference grammar, worst case collection of 'teach yourself stuff from the web')
• Glosses (either from the web, or from the grammar)
• Monolingual corpus
  • Frequency list
• Partially made analyser
• Spellchecker (list of validated surface forms)
• Bilingual wordlists
• "full form" lists
• partial-full form lists (e.g. category but not gender)
• Wikipedia (e.g. to get lists of categorised proper names, and to categorise proper names)
• Wiktionary
• Existing MT system(s)

Monolingual dictionaries

What kind of things might we want to do ?

• Assign possible categories to surface forms from the corpus
  • e.g. DET *UNK* ADJ → DET N ADJ
  • these classes/categories could be extracted somehow from a well-trained tagger.
• Assign possible features to surface forms in the corpus
  • e.g. DET *UNK* ADJ.M.SG → DET N.M.SG ADJ.M.SG
  • DET *UNK* ADJ.MF.PL → DET N.GD.PL ADJ.MF.PL
• Relate forms in the corpus between each other by means of paradigms
  • Sometimes you might get conflicts, in this case the paradigm<->stem combinations could get a score which is related to the predictive power according to the corpus.
    • features that are known to be ambiguous could get lower scores (this can be calculated from the partial dictionary)
• How to find good rules for determining POS ? -- Run the patterns through a corpus and count how frequency the POS you want comes in the slot you think is good.
• Sometimes you can just get good candidates from an ending + paradigm, e.g. all words that end in -ió, -joni, etc.
• Blacklist words from the corpus / hitparade, e.g. barbarismes.
• You might want to verify your analyser output on a corpus. Let's say you add a load of nouns and some of them have multiple gender -- it could be that some of them are erroneous, so you might want to run a corpus and pull out the instances of gender mismatch.

Making a dictionary can be an iterative process, generate some candidates, add them to the dictionary, run the scripts again because you have more context.

Payoffs

• You might want to add _all_ the possible analyses, but also be able to produce a "trimmed" dictionary.
  • For example, adding all the adjectives in -oso, even if they are low frequency -- e.g. annoying or hard to find translations for, because they might give you context for adding some higher frequency noun.
• In the end though, you want to be able to produce a dictionary trimmed to a bilingual dictionary. Also, if it needs to be manually revised, it should be trimmed.

Bilingual dictionaries

• Some kind of web/command line interface which gives you one word at a time from a missing-bidix frequency list, along with:
  • a source sentence with that word, -- it should be possible to toggle this with a keypress (e.g. cycle through the available sentences with the word.)
  • the output of the existing MT system,
  • possibilities from a probabilistic dictionary,
  • translation(s) of the sentence with other MT systems.