Filter multigram features based on how commonly they appear together
Argument and Default Value¶
PMI threshold will be set to it's default which is 3.0 (overridden by --set_pmi_threshold)
This filters multigram features based on their PMI value/(number of words - 1) and creates a new feature table that contains only the multi grams that were above a given threshold. The PMI value for a bigram b composed of word1 followed by word2 is calculated as follows:
In this case is the number of times x shows up divided by the total number of words in a document.
Intuitively the PMI should be a measure of how much a word pair "goes together". The PMI of a bigram like "happy birthday" will have a larger PMI. The PMI of a bigram like "bird purple" will have a smaller PMI because it was probably just a random occurence.
We divide the PMI by the number of words minus 1 in order to normalize.. otherwise 3grams have much higher PMI values than 2grams.
For more information see the article: Normalized (Pointwise) Mutual Information in Collocation Extraction by Gerlof Bouma
dlatkInterface.py -d fb22 -t messagesEn -c user_id -f 'feat$2to3gram$messagesEn$user_id$16to16$0_02' --colloc_filter --set_pmi_threshold 6.0