<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-3008">
  <Title>Human-Computer Interaction for Semantic Disambiguation</Title>
  <Section position="2" start_page="0" end_page="112" type="intro">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> Extraction and representation of text meaning is a central concern of natural language application developers. This goal still largely eludes computational linguists. Many problems remain unresolved. They include referential ambiguity resolution \[4, 12\], determining the nature of semantic dependency relations (as, tbr instance, in compound nouns in English \[8\]), treatment of novel language and ill-formed input \[21\], metaphor and metonymy \[6, 7\], discourse and pragmatic meanings \[11, 14, 17\], etc.</Paragraph>
    <Paragraph position="1"> Another set of tasks includes work on representation languages both for text meaning proper and for ontological domain models that underlie semantic analysis of texts \[1, 7, 13, 15\], problems of acquiring and working with domains and sublanguages of realistic s~e \[15, 16\] and taking into account requirements of particular applications, such as machine translation, natural language interfaces to databases and expert systems, etc.</Paragraph>
    <Paragraph position="2"> In the partial case of a particular application area, the representation problems ate alleviated. However, the treatment of a large number of linguistic phenomena is still a major problem. At this point, the developers of natural language processing (NLP) applications have a  choice of 1. not relying on results of semantic and pragmatic analysis; 2. providing semantic analysis for selected phenomena and limited domains only; or 3. using human help in determining facets of text meaning.</Paragraph>
    <Paragraph position="3">  In this paper we describe an environment facilitating human involvement in semantic and pragmatic analysis (Figure 1). This envhonment is applicable to mos~ corn-. prehensive lkq~P applications and consisL~ of ~_~ automatic analyzer of input text, a generator of output guage) and an augmentor module that bridges the two and facilitates the involvement of a human in the processing loop. The background knowledge for such a system consists of an ontological domain model, a grammar and a machine-tractable dictionary (MTD) t for each natural language involved in either analysis or generation.</Paragraph>
    <Paragraph position="4"> We will concentrate on the augmentor module, which consists of a human-computer interface with a dialog manager and a set of automatic semantic analysis components. The composition of the automatic components depends on the capabilities of the particular analyzer with which the augmentor is coupled. We proceed from the assumption that the format and content of the input to generation is fixed. It is this set of knowledge structures that we call the text meaning representation. Therefore, if the automatic analyzer is relatively shallow, the augmentor will have to perform more operations to fill the gaps in this representation. The role of the augmentor will diminish as the sophistication of the automatic analyzers increases. The above means that the environment we suggest is flexible and durable as a software configuration, because new findings and methods of treatment of the various linguistic phenomena will be accommodated in the architecture as they appear.</Paragraph>
    <Paragraph position="5"> The concept of the augmentor is also useful from the standpoint of building large software systems. In such applications it is usually desirable to incorporate as many existing software modules as possible, to avoid developing software from scratch. However, many such components expect their inputs and produce their outputs in an idiosyncratic formalism. An augmentor module can include special facilities for reformatting the output of one software module in accordance with the requirements on the input to another module. In the framework of natural language processing, the augmentor will usually reformat the results of the analyzer into the format expected by the generator.</Paragraph>
    <Paragraph position="6"> We now describe the augmentor module of the</Paragraph>
    <Paragraph position="8"> The augmentor components are shaded.</Paragraph>
    <Paragraph position="9"> in the analyzer module and partly in the augmentor, a division dictated largely by the requirements of computational efficiency and the reuse of an existing module. The KBMT-89 analyzer was built around a parser developed by Tomita in 1986 \[20\]. It proved to be essential to apply semantic constraints early in the parsing process to reduce the number of ambiguities; however, the semantic processing integrated into the analyzer was insufficient in many cases. Since the output of the KBMT-89 analyzer had to be reformatted in accordance with the requirements of the generator, additional ,semantic analysis was to be performed at the augmentation stage. Parts of this analysis and disambiguation can be performed automatically; for the remainder, human interaction is used. The user is asked to supply missing information and to choose among mnbiguous alternatives until a single, unambiguous result (called an interlingua text, or ILT) is obtained. This is in contrast to other systems such as TEAM, which selects a &amp;quot;best&amp;quot; parse to process based on a priori syntactic considerations \[19\].</Paragraph>
    <Paragraph position="10"> The KBMT-89 augmentor was thus designed with three main components to meet the criteria mentioned above: a format converter, an automatic augmentor/disan~biguator and an interactive disambiguator (Figure 2, previously published in \[2\]).</Paragraph>
  </Section>
class="xml-element"></Paper>