Semantics and computation
One of the obvious differences between Jackendoff’s conceptual semantics and the theories examined in the previous chapter is the formal nature of Jackendoff’s theory. This formal character means that Jackendoff’s system is already in a form which would ideally allow it to be implemented on a computer. Computer technology has taken on great importance in linguistics generally since the 1970s, and semantics is no exception. Computers come closer than any other artificial system to matching the complexity and interconnectedness of the human brain, and it has often been assumed that we can learn important lessons about the way language is processed in real-life minds/brains by trying to simulate this ability computationally.
Computer simulations are not simply of theoretical interest, however. They also contribute to diverse practical applications, such as machine translation, the development of searchable corpora, speech recognition, spell-checking, and so on. Much of the research in computational linguistics is geared precisely towards these practical applications, and is con ducted just as much by computer scientists as by linguists. This means that computational linguistics is a highly cross-disciplinary field. Another result is that researchers are often more interested in satisfying practical needs rather than theoretical ones. A software engineer concerned to develop a functional natural-language processing system will have scant regard for the psychological plausibility of the result: what determines its success is simply whether it performs the task at hand, not whether it does this in a way that might mirror human abilities. It is therefore important not to expect to learn too much about the mind/brain from computer simulation. Computer architectures certainly are the closest simulations available for the complexity of the brain, but they still vastly underperform humans in linguistic ability, as anyone who has used such functions as grammar checkers, automatic translation programmes, voice recognition systems and the like will be able to agree. One reason for this may be that there are, in fact, many respects in which the analogy between mind and computer breaks down. (There have been many critiques of attempts to understand and model the mind on the analogy of computers. Searle 1980 and Dreyfus 1992 are two prominent examples.)
Another reason not to attach excessive significance to the mind–computer parallel comes from the history of technology. As has often been pointed out, it frequently proves to be the case that artificial simulations of natural abilities harness different underlying principles from those actually found in nature. For example, early attempts to build flying-machines tried to replicate bird flight, in the belief that the movable wing structures found in nature provided the best solution to the engineering problems involved. These attempts, however, never succeeded. The flying technology perfected by humans uses fixed-wing principles, which, of course, are unknown in birds. Fixed-wing flying is ultimately responsible to the same physical principles as bird flight, but makes use of them in a fundamentally different way. There is a cautionary lesson to be learnt here. An artificial system can accomplish similar goals to a natural one through quite different means. It may therefore be mistaken to look to artificial computer simulations of natural language for specific insights into the nature of the human language faculty: the neurological and psychological process of language may be entirely unlike the computational processes of a computer.
But the brain need not resemble a computer for the study of computer simulation of human linguistic ability to be fruitful. Whether or not humans and computers process information in fundamentally similar ways, we can look to computer simulation as a way of appreciating, often in fine detail, the tasks which any language-processor, people included, must perform. Computer programs are blind. They cannot rely on humans’ general intelligence in solving problems or applying general principles to particular cases. A computer will not fill in gaps using general common sense knowledge. Instead, every step of a programme must be explicitly spelled out in minute detail if the ‘right’ result is to be achieved. The process of automating natural linguistic abilities therefore demands a finegrained attention to the detail of linguistic processes. This requires computational linguistic programmes to be specified in extremely close detail. Only when language is simulated by a machine can we test the explicitness and completeness of a given linguistic theory. In this section, we will therefore concentrate on the aspects of computational linguistics which give insight into the nature of the task of language-processing as it concerns semantics.
