TRACE
المؤلف:
John Field
المصدر:
Psycholinguistics
الجزء والصفحة:
P308
2025-10-21
31
TRACE
A leading computer simulation of spoken word processing (McClel land and Elman, 1986). TRACE is a connectionist model which operates on three different levels, corresponding to phonetic features, phonemes and words. It first encodes an incoming speech signal, not in terms of discrete phonemes or syllables but by dividing it into extremely short time slices which record the values of seven different acoustic features at any given moment. A series of phoneme units then samples this data for evidence that a given phoneme is present. The evidence upon which a phoneme unit draws covers three time-slices at a time, but the unit is linked, in all, to eleven consecutive slices. In this way, TRACE’s architecture takes account of the way in which features specifying different phonemes overlap in time (the non-linearity problem).
There is also a lexical level, containing a bank of units for each word in the lexicon. Each bank samples the phonemic data for evidence of the presence of its particular word. Like the phoneme units, the word units receive information from three time-slices at a time; and, also like the phoneme units, they are connected to, and continue to monitor, preceding time-slices. Though a word boundary may have been crossed, the lexical banks still have access to information which the system extracted before the boundary occurred. This enables the model to deal with the ‘right-context’ problem: the fact that many words cannot be confidently identified until well after their offsets.
However, TRACE’s critics suggest that its time-slice solution is bought at the cost of an extremely complex structure. The entire set of lexical units has to be duplicated every third time-slice. In the original TRACE format, there were only 211 words in the program’s lexicon; to simulate a typical listener they would have to be expanded to at least 30,000, all of which would need to be matched repeatedly against the input.
TRACE is a highly interactive model, with a continuous flow of information between its levels in both bottom-up and top-down directions. Evidence for a unit at one level lends support (activation) to a unit at a higher level but reduces support for competing units at the same level. Thus, evidence of /w/ at phoneme level would boost the activation of the word WORK, but scale down the activation of the competing phoneme /v/. This interactive structure enables TRACE to deal with misheard or mispronounced input. If there is a less than perfect match (e.g. if the model encounters a word like SHIGARETTE), the incoming evidence continues to build up activation for the most likely fit.
TRACE has performed impressively in lexical recognition tasks. But some reservations have been raised in relation to its ability to simulate the process of lexical segmentation.
See also: Connectionism, Interactive activation, Speech perception: phoneme variation
Further reading: Ellis and Humphreys (1999: 343–7); Harley (2001: 233–7)
الاكثر قراءة في Linguistics fields
اخر الاخبار
اخبار العتبة العباسية المقدسة
