Computation of probabilities for island-driven parsers

(A(x) j n s) is essentially a performance model obtainable by experiments of a system that segments a speech signal into syllables, n s being the number of actual syllables in the sentence corresponding to the signal. Pr(n s j m) is the probability that a string of m words is made up of n s syllables; it can be estimated from a written text. Once the above probabilities have been computed, it is possible to estimate Pr(S < uxvy >) as follows: (14) where Pr(A(x) j m) < " for m m min and m m max. It is also possible to introduce other heuristics, such as weight-ing the latter probability Pr(S < ux (m) vy () >) with some quantities obtained from Pr(A(x) j m). DISCUSSION We think that the presented theory will be eeective when the task at hand has certain characteristics. In particular , a good task should be sentence interpretation in restrict domains, because in this case island predictions can be guided by extrasyntactical sources of knowledge (for example semantic and pragmatic heuristics). In addition, a word spotter should be required to nd the words predicted by the system within the speech signal; the performance of the word spotter is improved by the fact that we only look at a limited number of \higly informative keywords". Our theory accounts for bidirectional expansion of partial analyses; this results in an improvement of the predictive capabilities of the system. In fact, bidirectional strategies can be used in restricting the (stochastic) syntactic search space for gaps surrounded by two partial analyses. This point has been discussed in 7] for the cases of one word length gaps. Notice that our theory generalizes such cases to m-length gaps as well as to case when partial analyses are not partially instantiated context-free productions, but partial derivation trees in all their generality. So far we have discussed how to score a given theory. It is also possible to use the presented framework in predicting words adjacent to an already recognized string, in order to extend the corresponding theory. This can be done by selecting the word(s) which maximize the probability of the theory augmented with it. Instead of computing these probabilities for all symbols in the dictionary, it is possible to restrict such an expensive process to the preterminal symbols (see 2]).