User:Francis Tyers/Sandbox2

Constraint-based lexical selection for rule-based machine translation

Corpus: cawiki-20110616-pages-articles.xml.bz2
          cleaned with `aq-wikicrp'

1758582 lines
531983  unique analyses 
531436  lines with >1 translation (30%)
2740    analyses with >1 translation
287     words (lemma+pos) with >1 translation in corpus
712     words in dictionary with >1 translation

1.03    fertility of dictionary over corpus (e.g. total number of word:word translations / total number of words)

Test corpus:

• 150 test words
• 1,500 sentences
• 10 per test word
• Randomly selected from the subset of sentences which were found in the corpus.
• Only words with >100 example sentences included
• Rationale: Dictionary doesn't provide good enough coverage to produce statistically significant results over a whole corpus.

• Discarding because of bad tagging: 'to', 'sol' (less than 50% correct)

Training corpus:

Baselines:

• TL Frequency-best
• TLM-best
• Linguist set

• Full analysis:Full analysis dic from Giza++
• Rules from phrase table

Process for using GIZA++:

• Tag both sides of the corpus (europarl, en-ca, first 1,700,000 sentences) with the Apertium language pair.
• Extract the model/lex.f2e file.
• Take the top scoring analysis:analysis results where the POS matches
• Where the word is already ambiguous in the Apertium dictionaries, add the possibilities from GIZA to the dictionary so that they may be chosen -- only added with POS tag.