ARTIFICIAL INTELLIGENCE (AI)
Psychologists and computer scientists have joined forces to create computer simulations of human cognitive processes. The processes studied in this way include the understanding of language and the nature of expertise and how it is acquired.
Researchers working within AI require a detailed information processing model before they can simulate an activity. Hence the custom among psycholinguists of presenting theories in the form of models which resemble the step-by-step operations of a computer. The argument is not that a computer would operate in the same way as the human mind but that, in designing a computer program, we can obtain insights into the real-life process.
Sometimes AI researchers and psycholinguists have different goals. A distinction can be made between programs whose aim is to make computers ‘intelligent’ without regard to whether the processes involved resemble those of the human mind, and programs which attempt to shed light on human cognitive processes. For example, computer programs designed to parse written text can achieve their goals on the basis of frequency (the statistical likelihood of a particular word occurring in a particular type of text) and transitional probability (the statistical likelihood that word A will be followed by Word B). This ignores factors in natural comprehension (e.g. world knowledge and the existence of a meaning representation of the whole text) in the interests of efficient machine processing.
Another difference between many AI programs and natural language processing lies in the fact that linguistic information may have to be coded for presentation to the computer. Thus, some AI models of spoken word recognition depend upon the researcher transcribing the utterance into phonemes.
AI research explores a number of specific areas of human cognition which are relevant to language:
Knowledge representation. Knowledge systems simulate the form in which knowledge (including linguistic knowledge) is stored in the mind; in particular, the relationship between declarative knowledge (knowledge that) and procedural knowledge (knowledge how).
Learning. Learning systems simulate the way in which features of a first or second language might be acquired from the data that is available.
Inference. Expert systems apply inference to a store of knowledge in an attempt to model the way in which the human mind analyses data and arrives at conclusions. This may in time assist our understanding of how listeners and readers impose inferences upon discourse.
Search. Problem-solving systems attempt to trace the way in which thinking moves from an initial state to a goal state, choosing one or more paths and selecting sub-goals along the way. Here, there are potential insights into, for example, the way in which speakers construct a syntactic pattern to express a proposition.
A more applied area of AI research aims to develop speech recognition programs. These projects face a major problem in the fact that human voices vary enormously in pitch, in articulatory settings, in the shape and size of the articulators involved and in paralinguistic features such as breathiness. Some programs (e.g. phone answering systems) are designed to discriminate between a limited number of words uttered by a wide range of voices. Others (e.g. transcription programs) are designed to discriminate between a large number of words uttered by one voice.
Currently influential in AI is a computational approach to lexical recognition known as connectionism or parallel distributed processing (PDP). It is based upon the transmission of activation between different levels of processing. Connectionist models often include a learning process, back propagation, which enables the computer to adjust its priorities in the light of successful or unsuccessful outcomes. Their proponents argue that this can provide insights into the process of language acquisition.
See also: Augmented Transition Network, Connectionism, Model, Syntactic parsing
Further reading: Garnham (1985: 11–15)
