Part-of-speech Tagging

1
Part-of-speech Tagging

• cs224n Final project
• Spring, 2008
• Tim Lai

2
POS Tagging 3 general techniques

• 1. Rule based system
• Relies on a hand-picked set of rules
• Performance is not very good
• 2. Stochastic methods
• HMM with Viterbi algorithm to determine best
  tagging
• Uses emission probabilities, i.e. P(word tag)
• and transition probabilities, i.e. P(prevTag
  currentTag)
• Maximum Entropy models also useful
• 3. Hybrid of the two
• Rules-based system to do POS tagging
• Uses rule templates and learns useful rules
  during training

3
Simple HMM vs Max-Ent

• HMM using bigrams for transition probabilities
• Max-Ent using simple features such as previous
  tag and current word

4
Error Analysis

• HMM and Max-Ent both perform well when tested on
  data from same domain
• Only 6.6 of words were ambiguous, making known
  words easy to tag
• Accuracy drops when using test data from another
  domain
• Most errors are caused by unknown words, or the
  POS tagging of words near unknown words.
• In sentences without unknown words, accuracy
  99!
• Most common mistake is mis-tagging JJ as NN
• Need to enhance both taggers to deal with
  unknowns.

5
Enhancement ideas

• For HMM
• Transition probabilities can be modeled using
  trigrams, taking more context information into
  account when word is unknown
• For Max-Ent
• Word shapes, word features, and more context can
  help
• Results
• HMM Switching from Unigram to Bigram helps a
  lot, but using Trigram doesnt help much.
• Max-Ent Hand picked features did not help much,
  but adding prefixes and suffixes were most
  helpful.

6
Transformation-based tagging

• One more idea to try using rule-based templates
  to learn POS tagging rules
• Sample rule template
• Change tag A to tag B when the preceding
  following word is tagged Z.
• Change tag A to tag B when the the tag Z appears
  within N positions of the current word.
• Result
• Using a very restricted set of rule templates,
  accuracy went up 0.5

7
Final results

• HMM with bigram and rule-based adjustments
• Max-Ent with prefix/suffix, word shape features
  and rule-based adjustments
• Max-Ent performs better, with 97 accuracy
  achievable