Predicting Oral Reading Miscues Export

@inproceedings{citeulike:309424, author = {Mostow, Jack and Beck, Joseph and Winter, Vanessa S. and Wang, Shaojun and Tobin, Brian}, booktitle = {ICSLP}, citeulike-article-id = {309424}, comment = {System to generate nbest list for likely child reading mispronunciations. Compare two different methods ("rote" and "extrapolative"). Training data : Colorado DB (Olson et al) of 112k transcribed child miscues (vocabulary 881 distinct words) Rote method selects miscues that have actually occured in the past (avg. 34.2 per word, pared down to 7.4 by looking at miscues that more than one student made). Extrapolative uses machine learning to extract other miscue words that are similar in pronunciation to the target words. Used a number of features (primarily a modified edit distance on pronunciation (more on that in another note). Results: as would be expected, rote works well in very common words (lots of data) and extrapolative works well for uncommon words. One big shortcoming was that miscues in extrapolative method seemed to necessarily be actual words in the dictionary. And the miscues were limitted to be words that started with the same phone as target. That last one seems like a mistake, because as they compare utility of different features in training the extrapolative model, the "same-first-phoneme" feature, while useful, scored rather low. ---=note-separator=--- They use a neat modified levenshtein (edit) distance: - 0-2 point penalty for substituting similar phones - 5 point penalty for substituting non-similar phonemes - unspecified penalty for insertion/deletion. And all normalized by phones in target word.}, keywords = {asr, call, child, cmu, good, language, levenshtein\_distance, listen, pedagogy, sla}, posted-at = {2005-08-31 21:47:46}, priority = {0}, title = {Predicting Oral Reading Miscues}, year = {2002} }

TY - CONF ID - citeulike:309424 L3 - citeulike-article-id:309424 TI - Predicting Oral Reading Miscues T2 - ICSLP KW - asr KW - call KW - child KW - cmu KW - good KW - language KW - levenshtein_distance KW - listen KW - pedagogy KW - sla N1 - System to generate nbest list for likely child reading mispronunciations. Compare two different methods ("rote" and "extrapolative"). Training data : Colorado DB (Olson et al) of 112k transcribed child miscues (vocabulary 881 distinct words) Rote method selects miscues that have actually occured in the past (avg. 34.2 per word, pared down to 7.4 by looking at miscues that more than one student made). Extrapolative uses machine learning to extract other miscue words that are similar in pronunciation to the target words. Used a number of features (primarily a modified edit distance on pronunciation (more on that in another note). Results: as would be expected, rote works well in very common words (lots of data) and extrapolative works well for uncommon words. One big shortcoming was that miscues in extrapolative method seemed to necessarily be actual words in the dictionary. And the miscues were limitted to be words that started with the same phone as target. That last one seems like a mistake, because as they compare utility of different features in training the extrapolative model, the "same-first-phoneme" feature, while useful, scored rather low. N1 - They use a neat modified levenshtein (edit) distance: - 0-2 point penalty for substituting similar phones - 5 point penalty for substituting non-similar phonemes - unspecified penalty for insertion/deletion. And all normalized by phones in target word. AU - Mostow, Jack AU - Beck, Joseph AU - Winter, Vanessa AU - Wang, Shaojun AU - Tobin, Brian PY - 2002/// ER -