<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-2132">
  <Title>Island Parsing and Bidirectional Charts</Title>
  <Section position="3" start_page="0" end_page="636" type="intro">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> The goat of using &amp;quot;high level&amp;quot; knowledge sources in recognizing continuous speech is to reduce the hypotheses space generated by acoustic-phonetic analysis (and possibly to implement an interpretation of the utterance) (see for instance Walker 1976, Stringa 1988). Deeodifying the vocal signal is a process that must take into account phenomena such as the eoarticulatory processes typical of continuous speech and the presence of many sources of variability of the signal (anatomic characteristics of the speaker, emission speed, prosody and so on). These phenomena have as a consequence the fact that, at the level of aco.ustic-phonetic analysis, it is extremely uncertain how to segment the signal and what labels to give to the segments. Therefore acoustic-phonetic analysis generates a space of possible interpretative hypotheses of the signal; in general, the likelihood of each lexical hypothesis is given a score. A matrix of lexical hypotheses is provided by the lower level processes. Each hypothesis is characterized by: a) the hypothesized string that was recognized; b) the score of this hypothesis; c) the time interval that this hypothesis spans. We consider two different thresholds for the likelihoods: the word hypotheses with score above the higher threshold are to he / considere~ &amp;quot;very reliable&amp;quot;, and their role will be to drive the process. The word hypotheses with score between the two thresholds will be included in the analysis without a  driving role, while the word hypotheses below the lower threshold are not to be considered, at least in the first pass. This work is about parsing with the above constraints. In this connection it seems advantageous to anchor the recognizing process to those hypotheses that were given a high score. As we shall see, there is also a predictive aspect in our approach: this means that the parser will tell the lower level component to &amp;quot;do its best&amp;quot; to find in the given place an instance of what was predicted. In the simplest case we can think of a direct recovering of a word hypothesis with score below the lower threshold. Our starting point will be a very well founded technique, that has also been experimentally proved as valid, namely chart oarsinm Chart parsing works very well with well formed input, but the technique was not conceived for working with an uncertain input, and even worse, with a fragmentary input. Chart parsing is directional in the sense that it works from the starting point (usually the beginning of the sentence) extending its activity usually in a rightward manner. We shall introduce a different concept, that nonetheless will work with the same linguistic data. The concept is that of a chart that works outward from islands and makes sense of as much of the sentence as it is actually possible. Furthermore, where the signal was just not detected, predictions can be made on the basis of the configuration and of a set of heuristics. After the application of these heuristics, and the introduction of new low level hypotheses, the algorithm works on in the same way and if the situation was not unrecoverable concludes with one (or more) complete analysis of the sentence.</Paragraph>
    <Paragraph position="1"> It is worth noting that the proposed solution helps also in dealing with ill-formed written input. There is something more to it: in the general treatment we are giving in this paper we shall refer only to a grammar coded in the traditional form of rewriting rules, but the mechanism can work with a large number of formalisms. Some contemporary linguistic theories emphasize the role of particular words that play the role of head of a constituent (e.g. a noun in a noun phrase). As a matter of general parsing strategy it seems very interesting to couple the localization of the pivot with an island mechanism that guarantees local control of the process in all directions.</Paragraph>
  </Section>
class="xml-element"></Paper>