Skip to content

Latest commit

 

History

History
31 lines (16 loc) · 1.03 KB

ideas.md

File metadata and controls

31 lines (16 loc) · 1.03 KB
Raw

Naive Bayes Classifier

Directly applying naive bayes classifer got 0.837 accuracy. - 2014.01.28

Use lexcicon and grouping words by category

Like use _RARE_ to replace all words occure less than 5 times. Use _DIGITS_ to replace all words that are digits, like 2014.

It improved the accuracy to 0.848. - 2014.01.29

Use sqrt of each log probability

Such that then difference of log probabilities that are too small(large in abs value) get weakened. It improve the accuracy to 0.872. - 2014.01.30

Treat multiple occurance of words as only once in decoding and training

Improve accuracy to 0.879. - 2014.01.31

Use average backoff from all domains

Improve accuracy to 0.880. - 2014.01.31

ideas failed

  1. Adjusting the _RARE_ threshold to other than 5 doesn't help.
  2. Adding Chinese surname lexicon didn't improve the model.
  3. Trying to use clustering method. Currently use KMeans(k = v/10) got 0.84 of accuracy. v is size of the vocalular.