Citation Context

...ple, capture aspects of language and language processing. These models showed that connectionist networks could acquire parts of linguistic structure without extensive ‘innate’ knowledge1. Recent work has moved towards a ‘connectionist psycholinguistics’, which captures detailed psychological data2. Criteria for connectionist psycholinguistics We review progress in connectionist psycholinguistics in four key areas: speech processing, sentence processing, language production, and reading aloud. We suggest that computational models, whether connectionist or symbolic, should meet three criteria: (1) data contact, (2) task veridicality, and (3) input representativeness. Data contact refers to the degree to which a model captures psycholinguistic data. Of course, there is more to capturing the data than simply fitting existing empirical results; for example, a model should also make non-obvious predictions (see Ref. 3 for discussion). Task veridicality refers to the match between the task facing people and the task given to the model. Although a precise match is difficult to obtain, it is important to minimize the discrepancy. For example, many models of the English past tense4 have low ta...

Citation Context

...heard before. Generalization to new cases is thus a crucial testa for many connectionist models. (3) Representation Because most connectionist nets learn, their internal codes are devised by the network to be appropriate for the task. Developing methods for understanding these codes is an important research strand. Whereas internal codes may be learned, the inputs and outputs to a network generally use a code specified by the researcher. The choice of code can be crucial in determining network performance. How these codes relate to standard symbolic representations of language is contentious. (4) Rules versus exceptions Many aspects of language exhibit ‘quasi-regularities’: regularities which usually hold, but which admit exceptions. In a symbolic framework, quasi-regularities may be captured by symbolic rules, associated with explicit lists of exceptions. Symbolic processing models often incorporate this distinction by having separate mechanisms for regular and exceptional cases. In contrast, connectionist nets may provide a single mechanism that can learn general rule-like regularities, and their exceptions. The viability of such ‘single route’ models has been a major point of contr...

Citation Context

... processing. These models showed that connectionist networks could acquire parts of linguistic structure without extensive ‘innate’ knowledge1. Recent work has moved towards a ‘connectionist psycholinguistics’, which captures detailed psychological data2. Criteria for connectionist psycholinguistics We review progress in connectionist psycholinguistics in four key areas: speech processing, sentence processing, language production, and reading aloud. We suggest that computational models, whether connectionist or symbolic, should meet three criteria: (1) data contact, (2) task veridicality, and (3) input representativeness. Data contact refers to the degree to which a model captures psycholinguistic data. Of course, there is more to capturing the data than simply fitting existing empirical results; for example, a model should also make non-obvious predictions (see Ref. 3 for discussion). Task veridicality refers to the match between the task facing people and the task given to the model. Although a precise match is difficult to obtain, it is important to minimize the discrepancy. For example, many models of the English past tense4 have low task veridicality because they map verb stems t...

Citation Context

...ts of language and language processing. These models showed that connectionist networks could acquire parts of linguistic structure without extensive ‘innate’ knowledge1. Recent work has moved towards a ‘connectionist psycholinguistics’, which captures detailed psychological data2. Criteria for connectionist psycholinguistics We review progress in connectionist psycholinguistics in four key areas: speech processing, sentence processing, language production, and reading aloud. We suggest that computational models, whether connectionist or symbolic, should meet three criteria: (1) data contact, (2) task veridicality, and (3) input representativeness. Data contact refers to the degree to which a model captures psycholinguistic data. Of course, there is more to capturing the data than simply fitting existing empirical results; for example, a model should also make non-obvious predictions (see Ref. 3 for discussion). Task veridicality refers to the match between the task facing people and the task given to the model. Although a precise match is difficult to obtain, it is important to minimize the discrepancy. For example, many models of the English past tense4 have low task veridicality be...