<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-3008">
  <Title>Human-Computer Interaction for Semantic Disambiguation</Title>
  <Section position="3" start_page="112" end_page="112" type="metho">
    <SectionTitle>
2. Automatic Augmentation
</SectionTitle>
    <Paragraph position="0"> The KBMT-89 system consists of multiple components which run in separate Lisp processes (usually on separate workstations) in a distributed tashion. The distinct components (Figure 2) are a source-language analyzer, a source-language generator for paraphrasing (used for verification),* a target-language generator, the augmentor, and file ONTOS knowledge acquisition tool \[18\] (used for queries or updates of the ontological domain model).</Paragraph>
    <Paragraph position="1"> The analyzer provides the augmentor with a nested list representation (Figure 3) of file meaning and syntax for each of the possible interpretations of the input sentenee. The angmentor extracts the semantic information (itself in a nested list format within the (SEM ... ) sublists), removes any completely duplicated semantic parses, and converts the nested lists to an isomorphic set of trees of linked frames using the FRAMEKIT knowledge representation package \[5, 9\]. The hierarchies of fr,'unes produced by the format conversion form the candidate interlingua texts. At this stage, however, the ILTs are still &amp;quot;bare&amp;quot;, containing only that information which appears directly in the analyzer output (Figure 4) and a stub for the speech act information.</Paragraph>
    <Paragraph position="2"> The automatic augmentor and disambiguator in KBMT-89 consist of a pattern matchcr and a pronominal anaphora resolver described in \[4\]. The pattern mateher performs a number of structural rearrangements on the trees of linked FRAMEI, aT frames, as well as adding inlormafion which is readily derivable from other information already present in the parser output.</Paragraph>
    <Paragraph position="3"> After the pattern matcher completes its modificalions of the interlingua texts, the automatic disambigua-lion procedures me invoked. Currently, only the pronominal anaphora rcsolver MARS (Multiple  MARS attempts to find the referent for each pronoun and definite noun phrase in the interlingua texts, and adds a link to the referent if found. It is often possible to eliminate a candidate ILT during resolution, particularly if further processing of the parses is delayed until the next several sentences have been processed by the anaphora resolver.</Paragraph>
    <Paragraph position="4"> MARS employs a set of constraints and preference strategies 2 to determine the referent of a pronoun or definite noun phrase. The constraints are applied first to reduce the set of candidate referents, and then the preference strategies are applied using a voting scheme.</Paragraph>
    <Paragraph position="5"> The candidate with the largest total weight is considered the desired referent, unless there are other candidates within a predetermined threshold, in which case the anaphor is held to be ambiguous among those candidate referents. Possibly after an interactive disambiguation session (described below), the user is asked to confirm a paraphrase of the input.</Paragraph>
  </Section>
  <Section position="4" start_page="112" end_page="112" type="metho">
    <SectionTitle>
3. Interactive Disambiguation
</SectionTitle>
    <Paragraph position="0"> If multiple candidate ILTs remain after automatic disambiguation, a composite ILT (as described in detail in \[2, 10\]) is created by combining all candidate parses, and any parses which are proper subsets of other remaining candidates are removed. The composite ILT is then used to generate a set of multiple-selection menus which will be used in the interactive disambiguation. null A composite ILT retains the tree structure of each candidate interlingua text used to form it. Each frame in the composite contains all of the slots contained in each of the original ILT frames from which it was made. In turn, each slot of a composite frame contains all of the distinct fillers together with pointers to the original ILTs containing each distinct filler.</Paragraph>
    <Paragraph position="1"> To begin interactive disambiguation, the augmentor checks the slots of the composite ILT for multiple fillers. If there are multiple fillers, the augmentor builds a set of multiple-choice menus for the user to decide which of these fillers must remain in the final inter-lingua text. The user interface (Figures 2 and 5) displays as many as four menus at a time during disambiguation, and the user makes his selections on any of them. This puts the user in partial control of the order in which ambiguities are eliminated, allowing him to choose the menu which is simplest or most obvious.</Paragraph>
    <Paragraph position="2"> By allowing more than one choice from a menu to be selected, some disambiguation can occur even if the user is unsure which value is most appropriate.</Paragraph>
    <Paragraph position="3"> After a decision has been made on a menu by clicking the mouse button over the desired choices and then DONE, the augmentor examines the composite, ILT and determines which of the candidate interlingua texts conlain any of the selected values. The ILTs which do not contain any selected values al~ discarded, and the composite 1LT is adjusted by removing the discarded  entries. Finally, the menu contents are adjusted to reflect any possible reduction in choices, and menus with only a single entry are deleted. Because the menus are not entirely independent, it is not unusual for a single selection to cause the removal of multiple menus, even if the menu on which the selection was made still contains more than one choice. Once the menus have been adjusted, another set of menus is displayed, and the cycle of menu display and user input repeats until a single, unambiguous interlingua text remains, which is passed on to the generator.</Paragraph>
  </Section>
  <Section position="5" start_page="112" end_page="112" type="metho">
    <SectionTitle>
4. Augmentor Interface Features
</SectionTitle>
    <Paragraph position="0"> The augmentor user interface (Figure 5) consists of an input/output panel in the bottom half of the screen, a main menu to its right, and the query area in the top half of the screen. The input/output panel is further divided (from top to bottom) into the input window, the status line, the paraphrase window, and the translation window. The input window accepts all typed input; the status line informs the user of the progress of a translation or indicates what input the augmentor is expecting; the paraphrase window displays a paraphrase of the input text after all analysis and disambiguation is complete, ,and the translation window displays the final translation after the paraphrase has been accepted by the user.</Paragraph>
    <Paragraph position="1"> The user interface allows the user to consult the ontological domain model or the relevant dictionaries through the knowledge acquisition system ONTOS.</Paragraph>
    <Paragraph position="2"> The user may query the knowledge base, displaying either a glaphical representation of the heirarchy or the actual contents of the frame for a concept. A simpler query is possible even if ONTOS is not loaded; each menu which asks for a selection among ambiguous concepts for a word allows the user to display the synonymous terms which map into each concept rather than the definition of the concept. The augmentor performs the necessary extraction from the ontology itself. All of the windows in the KBMT-89 augmentor were implemented using the programmable editor HEMLOCK integlated into Carnegie Mellon University's Common Lisp system. As a result, the input, paraphrase, and translation windows are actually editor buffers and each retains the previous output even after it has scrolled out of the window. This nmkes reviewing earlier work simply a matter of placing the mouse cursor in the appropriate window and issuing editordeg movement commands (either from the keyboard or by pressing the mouse buttons). The entire transcript from a given window can also be saved to a file, if desired.</Paragraph>
    <Paragraph position="3"> Since the KBMT-89 system is modular, changing the direction of translation only requires reconnecting the various modules in different ways. This may be accomplished by executing the setup procedure (which occurs automatically when the augmentor is initially loaded, and may be selected from the main menu) and specifying the source and target languages. A shortcut has been placed on the main menu to switch between English-to-Japanese and Japanese-to-English translation, as those were the languages available to KBMT-89. Regardless of the source and target languages, the angmentor invokes the proper analyzer and generators to accomplish the desired u'anslation.</Paragraph>
    <Paragraph position="4"> One of the more interesting tbatures of the KMBT-89 augmentor is that the user interface language has been made completely independent of the source and target languages by passing all messages through a lookup function before displaying them. The language may be selected, during setup, from among those installed, and may either remain fixed or change to the new source language whenever the source language is changed. If the proper set of messages has been installed in the lookup table, it is possible for the user interface to be, for instance, in German while translating from Japanese to English. The main use of this feature, however, is to allow a user to translate from all unfamiliar language into his native language, though not as well as translating from his native language into an unfamiliar one.</Paragraph>
    <Paragraph position="5"> The definitions displayed in word-sense disambiguation menus are similarly translatable by placing definitions for the desired languages into the ontology along with the English definition. For both user interdeg face messages and definitions, the augrnentor automatically falls back to English if the message or definition is not available in the appropriate language.</Paragraph>
  </Section>
  <Section position="6" start_page="112" end_page="112" type="metho">
    <SectionTitle>
5. An Example
</SectionTitle>
    <Paragraph position="0"> We now describe an actual example of the use of the augmentor in the translation of a sentence from English into Japanese. This example begins when the user enters the sentence to be translated (number 19 in the test corpus: 7. Set the power switch on the system unit to On.) 3. The augmentor invokes the English analyzer with this sentence as input. Once the candidate parses are obtained, the augmentor converts each of them into a set of FRAMEKIT flames, which it then augments by making a variety of implicit inlbrmation explicit and performing structural rearrangements. The MARS anaphora resolver does not apply to this sentence, since the latter does not contain pronominal anaphora, and there is no prior context for attempting to determine coreferentiality of definite noun phrases. Therefore, all of the candidate parses remain after the automatic processing.</Paragraph>
    <Paragraph position="1"> After augmentation and disambiguation, any remaining anabiguity in the candidate parses invokes an interactive disambiguation session. In this case, four menus appear, indicating that there are at least four points on which the 14 candidate parses differ (Figure 5). We will work with the lower-left menu first, as it has the largest number Of entries, which, we hope, will reduce the ambiguity most quickly. After deciding on DISCRETE-ELECTRONIC-MOVE-LEVER as the meaning of SET and clicking on it and then on DONE, the augmentor di~ards those candidate parses which do not contain the selected value in the appropriate position (we could have selected multiple items if we had been unsure of the correct one). In this case, the number of candidate parses is reduced from fourteen to six, and another menu replaces the one just completed (unfortunately, space constraints prohibit inclusion of further screen images; a complete version of this example will appear in a forthcoming paper \[3\]).</Paragraph>
    <Paragraph position="2"> 3&amp;quot;lhe domain of KBMT.-89 is personal computer installation and maintellance guides.</Paragraph>
    <Paragraph position="3"> 4 45 We now select ON-POSITION in the upper left-hand menu as the meaning of DISCRETE-POSITION (rather than using the more general POSITION) 4, which reduces the number of candidates to two and removes three of the menus, as two of the other menus were not independent of the upper left-hand menu. A new menu appears, and we are left with just two menus.</Paragraph>
    <Paragraph position="4"> After making a total of three selections, only one candidate parse remains. This is passed on to the English generator for paraphrasing, and the paraphrase is displayed in the center window. The augmenter asks whether the paraphrase properly captures the meaning of the input, and an affirmative response triggers generation in the target language. The translation appears in the bottom-most window. A negative response would have restored all of the candidate parses (including any eliminated automatically) and started another disambiguation session.</Paragraph>
  </Section>
class="xml-element"></Paper>