A theory of syntactic recognition for natural language By Mitchell P. Marcus (review)

REVIEWS447 heeded his own warnings a little more: the notation used there for representing speakers' intentions seems to meet none of the criteria listed by W. Another good thing is W's criticism of Chomsky's failure to realize the semantic basis for many of the claims about syntax that have been made in transformational grammar (see, e.g., p. 325). Finally, let us say a few words about formal matters. In spite of the many formalized or semiformalized passages, we have been able to detect only one error in them, viz. on p. 122, where A & B = df (A D -B) should be (as in the German original) A & B = df ~ (A D -B) As for the English text, the translator, Roger Lass, has by and large succeeded in the aim which he formulates on p. xv, to produce 'a book that does not "read like a translation"'. We have noted a few passages, however, where the English text deviates from the German original without any clear motivation. The most glaring deviation is in the third paragraph of p. 96, where the omission of a whole sentence concerning the semantic basis of syntactic categories makes the argument a non-sequitur. There are also some unhappy choices of terms, as when 'truth-value condition' is used rather than the standard 'truth-condition' for 'Wahrheitswertbedingungen' (215), or when Austin's term 'felicitous' is re-translated as 'successful' (268) via the German 'gelingen'. ('Success' has been taken as a much more narrow notion than 'felicity' in many works on pragmatics.) 'Eine Funktion aus der Menge A in die Menge B' is 'a function from A into B' rather than 'in B', as the text on p. 308 has it. This mistranslation is particularly confusing in view of the fact that it occurs in a non-standard definition of the concept of 'operation'. Our most serious formal objection concerns the absence of a subject index, something which should really be required by law in a book of this kind. It is astonishing that the Cambridge University Press neglected to include one, although one exists in the German original. [Received 6 July 1981.] A theory of syntactic recognition for natural language. By Mitchell P. Marcus. Cambridge, MA: MIT Press, 1980. Pp. 335. $25.00. Reviewed by Michael B. Kac, University ofMinnesota The central thesis of this book is that natural language can be realistically parsed by a deterministic device. To understand exactly what this entails, it is necessary to examine more closely the notions 'deterministic' and 'realistic'. A parser operating on a given sentence may reach a point where it does not 'know' what to do next. Any such point constitutes what may be called a 'local indeterminacy'; an example of such an indeterminacy is the word you in (1) a. I believe you. b. I believe you did it. This is because, when the parser encounters this word, it cannot determine— assuming that it has access only to the current word and preceding material— whether it is the object of believe or the subject of a new clause acting as the complement of believe. A non-deterministic way of proceeding would be to impose an arbitrary choice of alternatives—but to leave open the possibility 448LANGUAGE, VOLUME 58, NUMBER 2 (1982) that, if later processing shows the choice to have been wrong, the parser can return to the point where the error was made and try another possibility. One might, alternatively, require that, whenever an indeterminacy arises, a decision be deferred until some later point in the parse at which it can be resolved. These two strategies have come to be known as the guess-and-back-up and wait-and-see approaches, respectively. Deterministic parsing, in M's sense, is more strongly defined, however. In his view, the determinism hypothesis is reduced to vacuity if one can wait indefinitely before making a decision; hence limitations are imposed on just how much additional material can become available before the parser is forced to act. It is also assumed that decisions, once made, are irrevocable. A realistic parser is simply one whose behavior mimics that of a human being with the same input. There is, ofcourse...