<?xml version="1.0" standalone="yes"?>
<Paper uid="J80-2003">
  <Title>Responding Intelligently to Unparsable</Title>
  <Section position="11" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6. Conclusions
</SectionTitle>
    <Paragraph position="0"> We have drawn eight conclusions from our experience with the two systems on which our heuristics were tested. First, computing the presuppositions, or given information, of user input provides a means for detecting some of the user's assumptions inherent in the input. These may be checked against world knowledge in the system to recognize discrepancies between the user's model and the system's world model and to point out an incorrect assumption to the user.</Paragraph>
    <Paragraph position="1"> Second, an effective strategy for increasing the robustness of a parser is to allow relaxation of predicates (on ATN arcs) that the parser designer designates as relaxable, or &amp;quot;failable.&amp;quot; The system will prefer parses where no such predicates are false. If no parse can be found with all predicates true, the system will relax the predicates designated as failable, and will search for a parse with the fewest failable predicates false.</Paragraph>
    <Paragraph position="2"> The remaining conclusions regard our technique of assigning meanings to states as a means of generating responses when no parse can be found. The third conclusion is that the meanings of states, used with the longest path heuristic, can often pinpoint the cause of an input not parsing.</Paragraph>
    <Paragraph position="3"> Fourth, though the cause of the input not parsing can often be pinpointed with the technique, describing the cause to the user may be quite difficult because of the technical nature of the problem in the input.</Paragraph>
    <Paragraph position="4"> Fifth, the effectiveness of the longest path heuristic in correctly selecting the state corresponding to the actual problem in processing the input depends on the style of the grammar and the extent of the subset of language covered. The more constrained the language used in the application domain, the less possibility for the parser continuing beyond the point of the problem.</Paragraph>
    <Paragraph position="5"> Alternatively, the more syntactic and semantic constraints used as expectations by the parser, the greater the likelihood that the problem in the input will correctly correspond to a violated expectation, since violated expectations will help prevent the parser from going beyond the point of the problem. This does not conflict with the notion of relaxing predicates, since the longest path heuristic is used only after no parse can be found even after relaxing predicates. In our grammar, the longest path heuristic selected the correct state in over 90% of the test cases.</Paragraph>
    <Paragraph position="6"> Sixth, based on the two previous conclusions, the heuristic of responding using the meaning of states will be most effective in semantic grammars or in parsers that interact closely with semantic processes.</Paragraph>
    <Paragraph position="7"> Seventh, the longest path heuristic adds only a small fraction to the computing time and memory usage during parsing. Furthermore, adding the condition-action pairs to represent the meaning of states does not require a lot of programming, but does require a better understanding of the parser.</Paragraph>
    <Paragraph position="8"> Eighth and last, the technique of assigning meaning to states is applicable to explaining compile-time errors in programming languages as well.</Paragraph>
    <Paragraph position="9"> We also suggest four areas for further work. First, the heuristics should be tested in a parser that interacts closely with semantics while parsing. The purpose for that is twofold: (1) to more effectively respond to the user by paraphrasing the partial interpretation and semantic expectations when the input is unparsable and (2) to test further the effectiveness of the longest path heuristic. Second, the user's goals and intent are critical constraints which we have not incorporated in any of our heuristics. The aforementioned work on computational models of speech acts and dialogue games provide a starting point for this. A third area is to combine the ideas presented here with the heuristics in LIFER (Hendrix, et.al., 1978); the combination could provide a very user-oriented, flexible interface. Fourth, the effectiveness of our technique for responding to unparsable sentences should be examined in the domain of programming language compilers, because the user of a compiler knows many technical terms which the parser writer can employ in messages to convey effectively the cause of a blocked parse.</Paragraph>
  </Section>
class="xml-element"></Paper>