1. POS Tagging Markov Models

2. POS Tagging Purpose: to give us explicit information about the structure of a text, and of the language itself, without necessarily having a complete understanding of the text To feed other NLP applications/processes: –Chunking (feeds IE tasks) –Speech Recognition –IR Stemming (to more accurately stem) QA –Adding more structure (Parsing – in all its flavors)

3. Tags Most common: PTB’s ~45 tags Another common one: CLAWS7 (BNC), ~140 tags (up from a historic 62 tags)

4. Approaches to Tagging Rule-based tagging –Hand constructed –ENGTWOL (Voutilainen 1995) Stochastic tagging –Tag probabilities learned from training corpus drive tagging Transformation-based tagging –Rule-based –Rules learned from training corpus –Brill’s tagger (Brill 1995)Brill 1995

5. A Really Stupid Tagger Read the words and tags from a POS tagged corpus Count the # of tags for any given word Calculate the frequency for each tag-word pair Ignore all but the most frequent (for each word) Use the frequencies thus learned to tag a text Sound familiar? –HW#3! (All but last 2 steps.)

6. A Really Stupid Tagger But Charniak 1993 showed: –Such a tagger has an accuracy of 90% An early rule-based tagger (Greene and Rubin 1971), using hand-coded rules and patterns got 77% right The best stochastic taggers around hit about 95% (controlled experiments approach 99%) Let’s just give up and go home!

7. A Smarter Tagger Assume that a word’s tag is dependent on what tags precede it. Therefore, we would assume that the “history” of a word affects how it will be tagged. So what is more likely: 1.a/DT truly/RB fabulous/JJ play/NN 2.a/DT truly/RB fabulous/JJ play/VB

8. A Smarter Tagger So what is more likely: 1.a/DT truly/RB fabulous/JJ play/NN 2.a/DT truly/RB fabulous/JJ play/VB C(JJ,NN) P(NN|JJ) = ------------- = 0.45 C(JJ) C(JJ,VB) P(VB|JJ) = ------------- = 0.0005 C(JJ)  1 is more likely than 2 (because P(NN|JJ) > P(VB|JJ) Nothing beyond the JJ,NN vs. JJ,VBD transitions matters (well, almost)

9. Stochastic Tagging Assume that a word’s tag is dependent only on the preceding tag(s) –Could be just one –Could me more than one Train on a tagged corpus to: –Learn probabilities for various tag-tag sequences –Learn the possible tags for each word (and the associated probabilities)

10. Markov Tagger What is the goal of a Markov Tagger? To maximize the following equation: P(w i |t j )  P(t j |t 1,j-1 )

11. Markov Tagger A sequence of tags in text can be thought of as a Markov chain Markov chains have the following property: Limited horizon P(X i+1 = t j |X 1 …,X i ) = P(X i+1 = t j |X i ) or, following Charniak’s notation: P(t i+1 |t 1,i ) = P(t i+1 |t i ) Thus a word’s tag depends only on the previous tag (limited memory).

12. Next Time For next time, bring M&S & Charniak 93 Read the appropriate sections in 9 and 10. Study 10 over 9 (for now).