Volume 14, Number 3
PATTERN RECOGNITION LETTERS
statistical techniques. In a March 1992 paper (Bahl et al. (1992)) they have published results on continuous speech dictation. Their techniques produce an error rate of 4.8% in transcribing natural English sentences from a vocabulary of 5000 words, spoken as in a normal conversational mode, i.e., no forced pauses between words. If "perplexity" is viewed as the average branching factor at any point in a partially expanded sentence, this is a high perplexity task, i.e., it has a very large search space. This system is speaker-dependent requiring sufficient training data from each speaker. The Defense Advanced Research Projects Agency (DARPA) of the United States is sponsoring research on spoken language systems and the task chosen for this effort is an airline transport information system (ATIS). A February 1992 Proceedings of a DARPA workshop on speech and natural language processing contains a paper by F. Kubala et al. of BBN Corp. (Kubala et al. (1992)) on a task with a vocabulary of about 1500 words. The task has low perplexity. However the sentences are spoken spontaneously, not read from a prepared text; this leads to all types of hesitation and stammering. Also there is noise and the task is speaker-independent. The BBN system reported in the paper achieved a 6.2% error rate on a subset of the test data, with some noisy data removed. Certainly these achievements are significant when compared to past performance but the benchmark in such domains is human performance.
Suppose the above speech recognition systems scaled up and even learned, but we knew that they use hidden Markov modeling and certain statistical techniques and learning procedures. Would we call them intelligent? Perhaps some of us would and others would not. After all, we can search very large spaces using some search algorithms. Search is a very useful contribution of mathematics, OR and Al, and a key ingredient of Al systems. But even many within the field of Al are unwilling to view search as an important part of Al, and many in OR are unable to appreciate the value of heuristic search. It seems that the argument of what is or is not AI will continue even as the contributions from the field known as Al become an integral part of computer science, engineering and other disciplines.
Schank's article ends by suggesting case-based reasoning and case-based teaching as promising areas for Al. Currently case-based reasoning is a popular topic in Al. Relevant in this context is a theorem enunciated many years ago by Satosi Watanabe, a physicist-engineer and philosopher of pattern recognition.
Assessing similarity and using near-neighbor classification underlies case-based reasoning. Watanabe's "Theorem of the Ugly Duckling" proved many years ago that if the resemblance or similarity between two objects is measured by the maximum number of predicates shared by them, then the degree of similarity between any pair of arbitrary objects is the same. Thus a swan and a duck, and two swans are equally similar. This situation arises because all predicates of the same rank are treated equally. As Watanabe discussed in his papers and books (Watanabe (1969) and (1985)), performing logical manipulation on raw data resulting from observation does not provide a grouping among the observed objects because unless some predicates are considered more important than others, i.e., weighted more heavily, the above theorem holds. Logic may be "a ballet dance of bloodless categories" (Bradley (1922)), but the preferential weighing of predicates has its origins in human values and in the objective of performing a classification. As Watanabe puts it, "What makes human cognition possible is the evaluative weighing whose origin is aesthetic and emotional in the broadest sense of the terms." In comments on non-similarity grouping and object-predicate inversion, he also points out the weakness of simple-minded similarity theory as a foundation for pattern recognition and gives examples where the relationship among elements of a group is different from similarity. Interesting examples include trilateral circular relations and multilateral relations. In the first several chapters of his 1985 book titled, Pattern Recognition-Human and Mechanical, he summarizes earlier papers of his that cover a variety of philosophical views on categorization, from the Greeks and Western philosophers to Brahmanism and Buddhism.
Many of the points about categorization touched on in Watanabe's papers and books are addressed at length in an excellent book by George Lakoff, which brings together thinking and empirical evidence from several disciplines, languages and cultures, on how humans form categories (Lakoff (1987)). The provocative2 title Women, Fire and Dangerous Things, comes from the fact that these three items are placed in the same category in
2 Lakoff’s title invariably provokes a response. After my talk the following anonymous note showed up on bulletin boards in the convention center: “Women, Fire and Dangerous Things; Men Dust and Uninteresting Things.”