Machine Learning and Sentiment Classification in Ruby

Transcript

1. .BDIJOF -FBSOJOH BOE 4FOUJNFOU $MBTTJGJDBUJPO JO 3VCZ by Matt Drozdzynski @matid
3. .BDIJOF -FBSOJOH … or how to teach your computer to do back flips for you.
4. 4FOUJNFOU $MBTTJGJDBUJPO … or how to quantify people’s opinions.
5. #euruko is definitely the most amazing Ruby conference ever!
6. I’ve been to many dreadful conferences, but #euruko is certainly not one of them.
7. Ruby is a true delight compared to how horrendous Java can be.
8. d JO 3VCZ
11. %BUB (BUIFSJOH
14. -BOHVBHF "DDVSBDZ 0% 25% 50% 75% 100% 2007 English Spanish German Italian Polish
15. "OOPUBUJPOT … or I have the tweets—now what?
16. %BUB $MFBOJOH … or how to separate wheat from the chaff.
17. 'FBUVSF 3FEVDUJPO … or Matt’s crash course in selective ignorance.
18. $MBTTJGJDBUJPO … and the ‘not so rocket’ science behind it all.
19. /BJWF #BZFT Simple and robust Assumes independence of features Scalable!
20. require "ankusa" require "ankusa/memory_storage" storage = Ankusa::MemoryStorage.new classifier = Ankusa::NaiveBayesClassifier.new(storage) training.each do |tweet| classifier.train tweet.sentiment, tweet.to_s end sentiment = classifier.classify tweet.to_s
21. .BYJNVN &OUSPQZ No independence assumptions Suffers from overfitting Substantially slower than Naive Bayes
22. require "maxent_string_classifier" classifier = MaxentStringClassifier::Loader.train(Classifier.root + "max_ent" + "data") classification = classifier.classify tweet.to_s
23. 4VQQPSU 7FDUPS .BDIJOFT Non-probabilistic binary linear classifier Only directly applicable to two-class problems “Works by constructing a set of hyperplanes in a high or infinite dimensional space”—what?
25. require "eluka" classifier = Eluka::Model.new training.each do |tweet| classifier.add(tweet.features, tweet.sentiment) end classifier.build sentiment = classifier.classify tweet.features
26. $PODMVTJPOT … or is the whole thing worth the hassle?
27. 2VFTUJPOT
28. @matid spkr8.com/t/7678 bit.ly/matid-dissertation bit.ly/matid-dissertation-pdf 5IBOLT