Automatic acquisition of pronunciation rules using error-driven learning.
Material type:
TextDescription: 95 pISBN: 0599927682Subject(s): Computer Science | Language, Linguistics | 0984 | 0290Dissertation note: Thesis (Ph.D.)--Stanford University, 2000. Summary: The introduction of error-driven learning, also known as transformation-based learning or Brill's algorithm, was an important development in part-of-speech (POS) tagging research, because it showed that very good tagging results could be attained with simple training and almost no human intervention. Other statistical approaches also showed this advantage, but error-driven learning displayed the added benefit that the output was a set of rewrite rules which were more understandable to human readers and which were interpretable into finite-state transducers. The results of error-driven learning taggers are equivalent to other systems, including much more labor-intensive systems which are based on human-encoded rules and grammars. In this work, I show how error-driven learning can be adapted and generalized to apply to grapheme-phoneme conversion (GPC), the task of determining the pronunciation of written words. After describing the fundamental differences between POS tagging and GPC which keep us from applying error-driven learning directly, I describe how to adapt the approach to handle these differences and I then give some experimental results. In addition, I discuss how a number of other tasks within Natural Language Processing can also be handled by this generalized form of error-driven learning.
Source: Dissertation Abstracts International, Volume: 61-09, Section: B, page: 4820.
Adviser: Martin Kay.
Thesis (Ph.D.)--Stanford University, 2000.
The introduction of error-driven learning, also known as transformation-based learning or Brill's algorithm, was an important development in part-of-speech (POS) tagging research, because it showed that very good tagging results could be attained with simple training and almost no human intervention. Other statistical approaches also showed this advantage, but error-driven learning displayed the added benefit that the output was a set of rewrite rules which were more understandable to human readers and which were interpretable into finite-state transducers. The results of error-driven learning taggers are equivalent to other systems, including much more labor-intensive systems which are based on human-encoded rules and grammars. In this work, I show how error-driven learning can be adapted and generalized to apply to grapheme-phoneme conversion (GPC), the task of determining the pronunciation of written words. After describing the fundamental differences between POS tagging and GPC which keep us from applying error-driven learning directly, I describe how to adapt the approach to handle these differences and I then give some experimental results. In addition, I discuss how a number of other tasks within Natural Language Processing can also be handled by this generalized form of error-driven learning.
School code: 0212.
There are no comments on this title.