Briefly Noted

ation of many strings. Rule-based systems can do better , but the rules are based on years of s tudy by phoneticians, and are not only complex, but are supplemented by large lists of lexical exceptions. Presumably a connectionist system that could be taught rules as well as learn from example would outperform both versions; and that is what a human seems to b e a system that can be taught rules and that can learn by example how to apply and when to bend the rules. Levelt 's machine is not human-l ike in this respect. I t knows and uses rules. I t may even abs t rac t rules from mult iple examples. But the rules of the machine do not allow easy context-dependence, analogy, or the use of s imilar i ty. One of our annoyances with this fine book is that very often, a point is settled, without evidence, by a comment that so-and-so is very unlikely. In many of these cases we find so-and-so to be quite probable. A typical example is the comment (p. 18) that it is "an unlikely assumption that the speaker ar t iculates sentence i while formulat ing sentence i + 1." On the contrary, it seems to us unlikely that the speaker ever ar t iculates jus t one sentence without at the same t ime formulat ing future sentences that contr ibute to a developing argument . This is a difference of opinion that is unlikely to be resolved by experiment , but it i l lustrates the general point. In sum, our react ion to the book as an engineering design for a m a c h i n e t h a t m i g h t s p e a k l ike a h u m a n is "magnif icent ," but as a description of how people speak, "we remain unconvinced."