Mixed-Effects Regression Models in Linguistics by Dirk Speelman Kris Heylen & Dirk Geeraerts

Author:Dirk Speelman, Kris Heylen & Dirk Geeraerts
Language: eng
Format: epub
Publisher: Springer International Publishing, Cham

bejɣésa

subst. έ for é

1

bejɣέsa

delete a

1

bejɣέs

3

This sequence corresponds with the following alignment:

b

e

ɣ

é

s

a

b

e

j

ɣ

έ

s

1

1

1

The standard Levenshtein distance does not distinguish vowels from consonants and therefore could align these together. In order to prevent these (linguistically) undesirable alignments, a syllabicity constraint is normally added, allowing only alignments of vowels with vowels, consonants with consonants, and /j/ and /w/ with both consonants and vowels. It prevents alignments of other sounds, as these are assigned a very large (arbitrary) distance [22, 23].

It is clear that these Levenshtein pronunciation distances are very crude as the Levenshtein algorithm does not distinguish (e.g.,) substitutions involving similar sound segments, such as /e/ and /ε/, from more different sound segments, such as /e/ and /u/. Wieling et al. [24] proposed a method to automatically obtain more sensitive sound segment distances on the basis of how frequent they align according to the Levenshtein distance algorithm. Sound segments aligning relatively frequently obtain a low distance, while sound segments aligning relatively infrequently are assigned a high distance. The sound distances are based on calculating the Pointwise Mutual Information score (PMI; [25]) for every pair of sound segments. The automatically obtained sound segment distances were found to be phonetically sensible (based on six independent dialect data sets; [26]) and also improved pronunciation alignments when these sound segment distances were integrated in the Levenshtein distance algorithm [24]. A detailed description of the PMI-based approach can be found in Wieling et al. [26]. Similar to the study of Wieling et al. [27] on pronunciation differences between Dutch dialects and standard Dutch, our pronunciation distances are not based on the Levenshtein distance (with syllabicity constraint), but rather on the PMI-based Levenshtein distance. Using this phonetically more sensitive measure, the difference of the example alignment shown above is 0.107. The calculation is illustrated below:

b

e

ɣ

é

s

a

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.