Tham khảo tài liệu 'sound patterns of spoken english phần 8', ngoại ngữ, kỹ năng nghe tiếng anh phục vụ nhu cầu học tập, nghiên cứu và làm việc hiệu quả | 108 Experimental Studies in Casual Speech representation features phonemes words and programmed to interact with other levels. These researchers claim that the resulting complex system TRACE shares features with human speech perception in being able to map acoustic input into words and in the kinds of mistakes which it makes. Gaskell Hare and Marslen-Wilson trained a PDP to match a citation form on one hand with place-assimilated forms on the other so for example screen will be the output associated with screem screeng or screen . Feature bundles were used instead of phonetic symbols in order to reduce the arbitrariness of the mapping so that . m and n were identical save one feature and to look at the question of underspecification which will not be pursued here . Results indicated that the PDP could learn that the response screen could come from the mentioned set of similar-but-not-identical stimuli. In a second experiment phonological variants were input along with the environments which condition them so the training data included material like screen play screen test screem play screeng colour . Again the citation forms were given as the correct output. The PDP showed signs of learning that non-coronal segments could be mapped into coronal representations but also made mistakes as do humans. While one is ill-advised to accept that humans and PDPs work along the same lines we can tentatively conclude that if PDPs can make generalizations of this sort it is very likely that humans can too. Gaskell and Marslen-Wilson advocate an algorithmic approach which works on underlying phonological principles but these experiments suggest only that many-to-one mappings are possible. If a lexical item is represented in the mind by a variety of traces with different pronunciations one of them being the citation form the same end is achieved as when active phonological processes are in operation but perception of a reduced form is passive rather than algorithmic if a trace
