Brill, CL 1995

Citation

Brill, E. 1995. Transformation-based error-driven learning and natural language processing: a case study in part-of-speech tagging. Computational Linguistics. 21. 4. p543-565

Online version

Transformation-based error-driven learning and natural language processing: a case study in part-of-speech tagging

Summary

This paper introduces a learning technique called "Transformation-based error-driven learning", a.k.a. Transformation Based Learning (TBL).

The key points from the paper are:

• Transformation Based Learning is an algorithm that learns a sequence of transformations to improve tagging on some baseline tagger
• Transformations are broken down into two components: a triggering event (such as if the previous word is a determiner) and a re-write rule (such as change tag from modal to noun)
• Authors described the use of Transformation Based Learning on POS Tagging.
  • Bottom-line of results: When compared to a markov-model based POS tagger, the TBL Tagger was able to achieve comparable tagging accuracy with a number of rules which is much smaller than the number of context probabilities calculated for stocastic tagger, and can do so with a much smaller sized training corpus. Also initial rules contribute most to tagging accuracy (say first 100 or 200), and rest improve performance marginally.
• Some advantages of Transformation based learning:
  • Simple conceptually
  • Can be adapted to different learning problems
  • Rich triggers/rules that can make use of specific information and context
  • Seemingly resistant to over-fitting
    • Empirical result, not entirely understood
    • Always learn on whole data set

• Some considerations:
  • Constructing all possible transformations: manually create rules or make templates?, potentially huge search space can be problematic, may need to use linguistic intuition to limit space
  • No probabilities/confidence associated with results
  • Transformations in one environment could affect application in another: should transformations be applied immediately or only after entire corpus is examined for triggering conditions? what order do we process corpus? (left-to-right or right-to-left)

Transformation-Based Learning

The learning algorithm is summarized as follows (see Figure 1 from paper as well):

• Pass un-annotated corpus (training data) through initial-state annotator
• Compare against truth to get current score (based on number of classification errors)
• Loop until no transformation can be found to improve score (Greedy search)
  • Consider all transformations rules applied to training data, select best
  • Apply to transformation to data & get current score
  • Add transformation to ordered transformation list

Transformations are applied as follows:

• Run initial-state annotator on unseen data
• Loop through ordered list of transformations, and apply each transformation.

Example usage: POS Tagging

• Initial-state:
  • Each word assigned to most likely POS tag based on training corpus
• Rules:
  • Non-lexical: Based mainly on tags of words located near the target position
  • Lexical: Based mainly on words in surrounding context

Related papers

• SEARN in Practice: This unpublished manuscript showcases three example problems where SEARN can be used - Daume_et_al,_2006.