Data and Annotation for Sociolinguistics (DASL)
Annotation Specification: -t/d deletion

Note: This specification represents our current practices, and was adopted with the tagging of the Switchboard Corpus.  An earlier version (click here to view it) of the annotation spec was in place during coding of the TIMIT Corpus.  TIMIT will eventually be re-coded to bring it into line with the current spec.

• Status of the dependent variable

• Deleted

-t/d segment has been completely deleted
• Retained

-t/d segment is retained.  Although a binary division of deleted/retained is usually adequate, in some cases the final -t/d segment is phonetically altered.  The segment is often unreleased, glottalized or (less commonly) flapped.  All of these variant realizations of -t/d are coded as retained, and the annotator notes the variation in the comments field.
• N/A

Because the regular expression query used to generate the list of potential -t/d tokens operates on orthographic transcripts, even with these filters in place some words which are not tokens of -t/d appear in the final list of words to be reviewed.  (For instance, based on its orthography, the word 'would' appears to be a possible -t/d token, but is not.)  These tokens are labeled N/A and are excluded from the final analysis.
 
• Morphological Category (revised from TIMIT spec)

• Monomorphemic

-t/d appears in a word constituting a single morpheme, e.g. 'old'
• Irregular/Semi-weak Past Tense

-t/d appears in an irregular past tense verb of semi-weak verbs, which undergo both stem change and -t/d suffixation, e.g. 'told'.  Note: Annotation practices in previous studies have varied with respect to the treatment of irregular verbs.  Originally in the TIMIT corpus, LDC annotators included all irregular verbs in this category (modals like must, strong verbs like built; the token went).  Further annotation excluded must from this category because of its high rate of deletion and high number of tokens in the corpus (n=34); a final round of recoding excluded both went and strong verbs.
• Strong Past Tense

-t/d appears in the past tense form of strong verbs, which undergo stem change but contain a -t/d in both present and past tense forms, and thus cannot be said to undergo -t/d suffixation, e.g., 'build/built'.
• Regular Past Tense Preterite

-t/d appears in a regular past tense verb, e.g. 'walked'.
• Regular Past Tense "Helping"

-t/d appears in a regular past tense verb, as part of a compound verb, e.g. 'have walked'.
(Note: Discussion of functional load argument here)
• Went

-t/d appears in the token 'went'.  We chose to treat this verb separately because it appeared in TIMIT to have an unusually high deletion rate, much higher than that of monomorphemes.
• Must

-t/d appears in the modal verb 'must' which also appeared in TIMIT to have an unusually high deletion rate.


Past -t/d studies have shown that monomorphemes are the most likely category to undergo application of the -t/d deletion rule, whereas regular past-tense forms are most resistant to deletion.

• Preceding Segment

The characteristics of the consonant immediately preceding the -t/d token influences the probability of deletion as well.  In previous studies, manner of articulation of the preceding segment was shown to be the most relevant constraint, and our study adopts a seven-way distinction.  Although some previous studies have adopted a five-way distinction, we chose a finer-grained coding scheme because much of the published -t/d data indicates that preceding /s/ and other alveolar segments favor deletion more strongly than their non-alveolar counterparts, so we chose to code preceding alveolars as separate factors.  The seven categories of preceding segment are:
 
• Lateral (/l/)
• Rhotic (/r/)
• Alveolar Nasal (/n/)
• Non-Alveolar Nasal (/m/ or /ng/)
• Stop or Affricate (/p, b, k, g/). Cases of preceding affricates were noted in the comments field.
• Alveolar Fricative (/s, z/)
• Non-Alveolar Fricative (/f, v, th, sh/)


In coding this factor group, the annotator is sometimes faced with a situation in which the preceding segment has been reduced or vocalized.  This typically occurs with the liquids /r/ and /l/.  If the preceding /l/ segment has been phonetically reduced in this way, the annotator will note this in the comments field, but still indicates the preceding environment as /l/, because the segment is phonemically present although phonetically altered.  With vocalized preceding /r/, no evidence of the /r/ remains in the acoustic signal, so the token is coded as N/A, and the annotator notes the vocalized /r/ in the comments field.  Less commonly, the annotator might encounter tokens where non-liquid preceding segments have been deleted (e.g., government being rendered as "governme_t").  As with vocalized preceding /r/, such cases are coded as N/A, with appropriate commentary.

• Following environment

This refers to which segment, if any, follows the -t/d token.  Past studies of -t/d deletion have consistently demonstrated that following consonants favor deletion, while following vowels inhibit deletion.  Among the consonants, glides and liquids favor deletion less than obstruents.  The effect of following pause has been shown to be variable based on geographic region.  We have adopted a seven-way distinction for following segment:
 
• obstruents (stops, fricatives and nasals)
• lateral (/l/)
• rhotic (/r/)
• labio-velar glide (/w/)
• palatal glide (/y/).

(Note: In the TIMIT corpus, glides were annotated as 'clustering' or 'non-clustering'.  This distinction proved unimportant; the relevant distinction seems to be place of articulation of the glide rather than the glide's ability to cluster with the -t/d segment.)
• vowel
• pause (silence follows -t/d segment).  Following pause usually occurrs sentence-finally.  In other cases, the annotator would indicate following pause if the speaker had a significant, noticable break in the flow of speech, usually to take a breath, etc.


In coding following segments, annotators occasionally encounter cases where the segment has been entirely deleted.  This happens most frequently with /h/, particularly when the following word is an unstressed pronoun (her, his).  In such cases, the annotator codes the following segment as a vowel, since the segment has been completely deleted, and makes a note in the comments field.
 

• Same preceding and following environment (New factor group)

This indicates whether or not the -t/d token is surrounded by two identical segments, e.g., 'processed soybeans'; and if so, what these segments are.  In coding TIMIT, annotators observed that an indentical segment in the preceding and following environments significantly favored deletion of -t/d (p < 0.001).  This effect seems stronger with surrounding obstruents -- especially the alveolars /s/ and /z/ -- than with surrounding liquids.  The current annotation specification adopts an eight-way distinction, which may be condensed later.  (Note:  Because of the way preceding and following environments are coded, it is not possible to simply observe after the fact whether the surrounding environments are identical.  For instance, preceding /s/ would be coded as "alveolar fricative" while following /s/ would be coded as "obstruent".  This mismatch is in keeping with previous studies' coding schemes for -t/d deletion.)
 
• N/A
• surrounding alveolar fricatives (/s, z/)
• surrounding alveolar nasals (/n/)
• surrounding stops (/p, b, k. g/)  Also, surrounding affricates, which are noted in the comments field.
• surrounding other fricatives (/f, v, th, sh/)
• surrounding other nasals (/m, ng/)
• surrounding rhotics (/r/)
• surrounding laterals (/l/)

• Stress (New factor group)

This refers to whether or not the syllable containing the -t/d token is stressed ('passed') or unstressed, ('discouraged').  Past studies have shown that unstressed syllables are more likely to undergo -t/d deletion than stressed syllables.
 
• Stressed syllable
• Unstressed syllable
• Cluster complexity (New factor group)

This refers to how many elements are present in the cluster containing -t/d.  Past studies have shown that clusters containing three or more elements ('mixed') are more likely to undergo -t/d deletion than clusters with just two elements ('helped').
 
• Cluster with two elements
• Cluster with three elements
• Cluster with more than three elements


Back to main DASL page