<?xml version="1.0" standalone="yes"?>
<Paper uid="H92-1060">
  <Title>A Relaxation Method for Understanding Speech Utterances 1</Title>
  <Section position="3" start_page="299" end_page="300" type="metho">
    <SectionTitle>
ROBUST PARSING
MECHANISM
</SectionTitle>
    <Paragraph position="0"> The natural language component of the MIT ATIS system makes use of a semantic frame representation of the meaning which serves as the input for database access, spoken response generation, and history management. The frame design is flexible enough to be readily Jextended to other domains. Domain-dependent aspects of the system are entered mainly through table-driven mechanisms that seek certain patterns in the frame, with very little explicit programming required.</Paragraph>
    <Paragraph position="1"> Because the semantic frame is so central to our system, we felt it was appropriate to integrate the fragments provided by partial parse analysis at the frame level.</Paragraph>
    <Paragraph position="2"> Whenever a full linguistic analysis fails, a set of parse trees accounting for key phrases and clauses is recovered.</Paragraph>
    <Paragraph position="3"> Each parse tree is individually converted to a semantic frame, and the set of frames are combined to form a single semantic frame encoding the meaning of the entire sentence. This frame is then ready for integration into the existing mechanisms of the back-end component.</Paragraph>
    <Paragraph position="4"> The ability to provide partial parses was achieved by modifying the parser and the grammar in minor ways.</Paragraph>
    <Paragraph position="5"> The grammar is written as a set of context free rewrite rules with constraints, and is converted automatically to a network form, where each node in the network represents a particular category (which might be a semantic name such as a-place or a syntactic one such as predicate). In full-sentence analysis mode, only the sentence category is allowed to terminate, and only at the end of the sentence. In the relaxed mode, on the other hand, a set of categories representing important clauses and phrases are allowed to terminate, and such termination can occur anywhere in the sentence.</Paragraph>
    <Paragraph position="6"> W:hen operating in robust mode, the parser proceeds left-to-right, initially producing an exhaustive set of possible parses beginning at the first word of the sentence. The parse that consumes the most words is then selected 2.</Paragraph>
    <Paragraph position="7"> The parser begins again at the first subsequent word, repeating the procedure. Whenever no parses are returned, the parser advances by one word and tries again. Eventually a set of parsed phrases are returned.</Paragraph>
    <Paragraph position="8"> In order to combine parsed fragments, we need an inheritance mechanism that is similar in many respects to our discourse model. Since we already have the capability of responding appropriately to sentence fragments such as 'aircraft&amp;quot; or &amp;quot;first class,&amp;quot; we surmised that the same mechanism could be utilized effectively to fuse together parsed fragments within a single sentence. The only important distinction between such a sentence-internal history mechanism and the existing sentence-ezternal history mechanism is that nothing from the internal history can be overwritten, since answers have not yet been provided to the previous parsed fragments.</Paragraph>
    <Paragraph position="9"> In the standard history mechanism, the presence of certain attributes in the new frame masks inheritance of certain other attributes from the history. Furthermore, whenever a value for a given attribute occurs in the current frame and also in the history frame, the value of that attribute from the history is overwritten. The sentence-internal history mechanism remembers everything, however, since none of the pieces have as yet been answered.</Paragraph>
    <Paragraph position="10"> Whenever the frames are judged to be too disjoint, the system spawns additional top-level clauses, essentially producing a compound sentence. This would be the case, for example, for the input: &amp;quot;I'll take flight twelve oh nine. What ground transportation is available in Denver?&amp;quot; An example, shown in Figure 1, will help to explain the difference between the two history mechanisms. The sentence, &amp;quot;What are the meals and aircraft for flight two eighty one and also for flight two oh one,&amp;quot; is treated by the parser as three sequential entries: &amp;quot;What are the meals,&amp;quot; &amp;quot;aircraft for flight 281,&amp;quot; and &amp;quot;flight 201.&amp;quot; If this sequence were delivered to the sentence-external history mechanism, the last phrase would be interpreted as &amp;quot;aircraft for flight 201.&amp;quot; Sentence internally, however, the result would become &amp;quot;meals and aircraft for flights 281 and 201.&amp;quot; Once the sentence is fully fused, the external history is brought in, and the sentence may inherit further constraints from the dialogue context, as shown in the figure, where it picks up a source and destination.</Paragraph>
    <Paragraph position="11"> Further examples of robust parsing on sentences spoken by actual users are shown in Figure 2. In all three cases, we believe the system produced reasonable answers to the questions. The tables are omitted due to space lim- null itations, but the verbal response gives a clear indication of the system's interpretation.</Paragraph>
    <Section position="1" start_page="300" end_page="300" type="sub_section">
      <SectionTitle>
Rejection Criterion
</SectionTitle>
      <Paragraph position="0"> Because the DARPA evaluation mechanism currently penalizes systems for incorrect answers, we augmented the robust parser with a capability for detecting certain key words, such as &amp;quot;between,&amp;quot; which, if not properly understood, would most likely lead to an incorrect answer.</Paragraph>
      <Paragraph position="1"> Another heuristic, most relevant when a speech recognizer is included, was to refuse to answer if an unknown flight number was detected in the sentence. We used these sentences to update the discourse context, but gave a NO ANSWEa response for evaluation. In addition, when the input was judged overall to be sufficiently unreliable due to recognition errors, we used a more conservative rejection criterion that excluded answers for sentences that did not receive a full parse and were suspected to require context. We used a simple algorithm (flights with no source and destination) to distinguish this set.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>