Summary prepared by Mark Jolly (additions and comments by Frawley; for additional comments go to Short summaries of presentations by Landau and Intraub
Prof. Barbara Landau lectured on object recognition and, more specifically, how the linguistic representations of recognized objects comes about. To begin, Landau noted how closely related both the linguistic and the visual aspects of recognition are. In both areas, the act of recognition occurs quickly and unconsciously, even among the least-developed among us, babies, but attempting to model the process in a computer has emphasized its true complexity, just as the geon theory has run into a hitch with the same area.Landau began with a deceptively simple question: what is the meaning of the word triangle? Early in the consideration of such semantic inquiries, Landau said, it was proposed that the word is closely linked to the material world in that the word refers to an internal image of the object it names. For reasons mirrored in Dr. Hoffman's lecture the class before, this shallow concept of word representation fails to adequately explain the phenomenon. Images are very specific representations, which implies for the visual end a difficulty in accounting for the ability to recognize objects at different angles or in different lighting. For the linguistic side of the problem, the specificity of material things is a problem because words are much more general. Chair, for instance, refers to any number of objects and can therefore not be tied to one image closely related to the world in the mind/brain.
Closely related to this problem are the objections to a second theory -- that word meanings are equivalent to what is being referred to, the referent. The largest problem here is the fact that any single object can be referred to under several names, i.e. a robin is also a bird and an animal, but each of those names has its own slightly different meaning. In addition to that problem, many words exist that have no referent in the world. Words such as and, but, truth, beauty have no concrete referent but can be understood nonetheless.
After examining these early theories, Landau proposed to stop for a moment and ask what a good semantic theory would entail. According to Landau, a satisfactory theory must explain five major areas: identification of the referent in a wide variety of conditions, the ability to use words in combinations, semantic relations between words in a sentence, language acquisition, and minute differences in processing time for certain words or combinations of words.
The major current theories of meaning and category structure can be categorized as either holistic or compositional. Holistic theories are along the lines of those examined above; they posit that linguistic representations are closely tied to visual images in memory. Compositional theories seek to explain word comprehension by deconstructing the word into features. The classical, or definitional, compositional theory holds that any word has necessary and sufficient conditions for membership. For example, a bachelor can be defined as an unmarried adult male. Without any of those conditions, the person being referred to cannot be a bachelor.
The basis for these features, however, supplies a problem for the compositional theories. John Locke proposed the features were based in the sensory organs, an argument echoed by Hoffman's account of the discovery of a "monkey paw receptor" in certain experiments. Generally, however, this is an impractical way to search for features and relies almost wholly on serendipity to advance at all. Instead, most researchers agree that attribute listing by subjects is the best way to produce features. Progress made in this vein, however, is challenged by the success of certain "take-away" experiments in which subjects are shown a robin, for example, that doesn't have some feature listed as integral to its being, i.e. wings. The subjects will usually still classify the animal as a robin, albeit a deformed one.
As an interesting aside, this difficulty leads to a philosophical concept that objects all possess some essence that lends things their names and that features therefore don't apply.
[Or it leads to Kripke's theory, that the meaning of a word is its direct reference to the world under some "dubbing ceremony."]It is a fascinating possibility, but an unsatisfactory explanation in terms of cognitive science since it makes language magical. Remember, no ghosts. A more worthwhile consideration to stem from this idea, however, is that meaning is nothing more than a series of beliefs people hold that can be easily revised if there is a reason for it.
Returning to the examination of current semantic theories, it is necessary to examine the second main type of compositional theory, the prototype concept. This theory claims that membership under a specific word is not an all-or-nothing proposition. Instead, there is an exemplar to which objects are compared, and if the object comes close enough, it is determined to belong to that word.
Overall, featural concepts are considered fairly strong and are used as the underlying assumptions behind work in artificial intelligence and parsing. These theories are far from perfect, however. When one applies the classical compositional theory to language acquisition, the fourth major area a satisfactory theory must explain, it falls short. Under a featural theory, the fewer units inherent in a word's definition, the easier that word would be to learn. Therefore bachelor, for instance, would be harder to both learn and process, since it consists of three necessary conditions, than boy, which is made up of two features -- `child' and` male.' Empirically, the theory is not verified, that is, bachelor has not, in experiments, proven to take longer to learn or process than boy.
One final theory examined in the lecture was drawn from the Armstrong, Gleitman and Gleitman paper. That theory proposes a graded model drawing from the prototype theory, that is, there are certain "perfect" objects within a class and that should be provable experimentally, which it appears to be. But words that appear to be definitional without a doubt, such as prime numbers, also produce a "perfect" example in experiments. This would suggest that prototypes, therefore, come about not because of some mind/brain structure, but due to other reasons such as familiarity or usefulness.
In the end, no semantic theory has yet been developed that can satisfactorily explain all five necessary areas, but evidence exists to support each of the current theories mentioned above.