Classifying Parts of Speech Based on Sparse Data

1
Classifying Parts of Speech Based on Sparse Data

Katherine Brainard
2
The Problem

• Sparse data has little contextual information
• Many words fall into this category
• Automatic PoS taggers and finders are useful

3
Approach

• Relatively easy to learn categories from frequent
  words
• Infrequent words often more regular than their
  common counterparts
• Learn frequent words, then use these to classify
  infrequent
• Uses clustering for the frequent words

4
Evaluating the Model

• Somewhat tricky - want eval function that doesnt
  encourage degenerate behavior
• Evaluation separated from clustering
• Used both bigram probability model and comparison
  with already-tagged data

5
Results

• Improvement of 36 from delaying processing of
  data
• About 2.5 times better than classifying
  infrequent words into one lump
• Using just contextual data produced the best
  performance