Reducing the Annotation Effort for Letter-to-Phoneme Conversion
Dwyer, Kenneth and Kondrak, Grzegorz

Article Structure

Abstract

Letter-to-phoneme (L2P) conversion is the process of producing a correct phoneme sequence for a word, given its letters.

Introduction

The task of letter-to-phoneme (L2P) conversion is to produce a correct sequence of phonemes, given the letters that comprise a word.

Decision tree learning of L2P classifiers

In this work, we employ a decision tree model to learn the mapping from words to phoneme sequences.

Context ordering

In the L2P task, context letters that are adjacent to the focus letter tend to be more important than context letters that are further away.

Clustering letters

A decision tree trained on L2P data bases its phonetic predictions on the surrounding letter context.

Active learning

Whereas a passive supervised learning algorithm is provided with a collection of training examples that are typically drawn at random, an active learner has control over the labelled data that it obtains (Cohn et al., 1992).

L2P alignment

Before supervised learning can take place, the letters in each word need to be aligned with

Experimental setup

We performed experiments on six datasets, which were obtained from the PRONALSYL letter-to-phoneme conversion challenge.3 They are: English CMUDict (Carnegie Mellon University, 1998); French BRULEX (Content et al., 1990), Dutch and German CELEX (Baayen et al., 1996), the Italian Festival dictionary (Cosi et al., 2000), and the Spanish lexicon.

Results

We first examine the contributions of the individual system components, and then compare our complete system to the baseline.

Conclusions

We have presented a system for learning a letter-to-phoneme classifier that combines four distinct enhancements in order to minimize the amount of data that must be annotated.

Topics

random sampling

Appears in 7 sentences as: random sample (1) Random Sampling (1) random sampling (5)
In Reducing the Annotation Effort for Letter-to-Phoneme Conversion
  1. In order to speed up the experiments, a random sample of 2000 words was drawn from the pool and presented to the active learner each time.
    Page 5, “Experimental setup”
  2. The dashed curves in Figure 1 represent the baseline performance with no clustering, no context ordering, random sampling , and ALINE, unless otherwise noted.
    Page 6, “Results”
  3. For instance, on the Spanish dataset, random sampling reached 97% word accuracy after 1420 words had been annotated, whereas QBB did so with only 510 words — a 64% reduction in labelling effort.
    Page 6, “Results”
  4. It is important to note that empirical comparisons of different active learning techniques have shown that random sampling establishes a very
    Page 6, “Results”
  5. Query-by-Bagging __ Random Sampling
    Page 7, “Results”
  6. It is rarely the case that a given active learning strategy is able to unanimously outperform random sampling across a range of datasets.
    Page 7, “Results”
  7. The complete system consists of context ordering, clustering, Query-by-Bagging, and ALINE; the baseline represents random sampling with EM alignment and no additional enhancements.
    Page 8, “Results”

See all papers in Proc. ACL 2009 that mention random sampling.

See all papers in Proc. ACL that mention random sampling.

Back to top.

statistically significant

Appears in 4 sentences as: Statistically significant (1) statistically significant (3)
In Reducing the Annotation Effort for Letter-to-Phoneme Conversion
  1. Statistically significant improvements were realized on Dutch, French, and German.
    Page 6, “Results”
  2. The only case where it had no statistically significant effect was on English.
    Page 6, “Results”
  3. From this perspective, to achieve statistically significant improvements on five of six L2P datasets (without ever being beaten by random) is an excellent result for QBB.
    Page 7, “Results”
  4. The learning curves (not shown) were virtually indistinguishable, and there were no statistically significant differences on any of the languages.
    Page 7, “Results”

See all papers in Proc. ACL 2009 that mention statistically significant.

See all papers in Proc. ACL that mention statistically significant.

Back to top.

human annotator

Appears in 3 sentences as: human annotation (1) human annotator (2)
In Reducing the Annotation Effort for Letter-to-Phoneme Conversion
  1. It is often desirable to reduce the quantity of training data — and hence human annotation — that is needed to train an L2P classifier for a new language.
    Page 1, “Abstract”
  2. Words for which the prediction confidence is above a certain threshold are immediately added to the lexicon, while the remaining words must be verified (and corrected, if necessary) by a human annotator .
    Page 4, “Active learning”
  3. In the L2P domain, we assume that a human annotator specifies the phonemes for an entire word, and that the active learner cannot query individual letters.
    Page 4, “Active learning”

See all papers in Proc. ACL 2009 that mention human annotator.

See all papers in Proc. ACL that mention human annotator.

Back to top.

overfitting

Appears in 3 sentences as: overfitting (3)
In Reducing the Annotation Effort for Letter-to-Phoneme Conversion
  1. By biasing the decision tree learner toward questions that are intuitively of greater utility, we make it less prone to overfitting on small data samples.
    Page 2, “Context ordering”
  2. 5 The idea of lowering the specificity of letter class questions as the context length increases is due to Kienappel and Kneser (2001), and is intended to avoid overfitting .
    Page 6, “Results”
  3. Our expectation was that context ordering would be particularly helpful during the early rounds of active learning, when there is a greater risk of overfitting on the small training sets.
    Page 6, “Results”

See all papers in Proc. ACL 2009 that mention overfitting.

See all papers in Proc. ACL that mention overfitting.

Back to top.