<?xml version="1.0" standalone="yes"?>
<Paper uid="H89-2031">
  <Title>PORTING TO NEW DOMAINS USING THE LEARNER *m</Title>
  <Section position="4" start_page="0" end_page="241" type="metho">
    <SectionTitle>
THE LEARNER
</SectionTitle>
    <Paragraph position="0"> Some researchers have addressed this problem by developing acquisition tools targeted for end users, to allow them to provide the syntactic and semantic information necessary for the NL system, but without requiring them to become experts in these areas. Such tools usually take the burden of providing detailed syntactic and semantic analyses off the user through a guided acquisition proccedure (either with menus or questions) and through the use of queries couched in terms of actual examples of language usage, rather than in terms from syntactic or semantic analysis (e.g. by asking &amp;quot;Can you say 'Someone  deployed something?'&amp;quot; rather than &amp;quot;Is 'deploy' transitive?&amp;quot;) Such acquisition tools include the acquisition component of TEAM (Grosz, 1983), the LapltUp lexical acquisition package for the JANUS system (Cumming and Albano, 1986), and the TELI system's semantic acquisiton facility (Ballard and Stumberger, 1986).</Paragraph>
    <Paragraph position="1"> BBN has developed a software package, the Learner, as a porting tool for non-expert users (Bates, 1989; BBN Parlance Learner Manual, 1989). The Learner creates a number of files that are used to configure the Parlance natural language processing system for a new application domain in a short time. Like the tools already mentioned, it uses an interactive, guided procedure to acquire syntactic and semantic information from the user. It also acquires information about the structure and content of the database directly from the database itself. Previous work with the Learner (Bates, 1989) has demonstrated a speed-up of ten times or more, compared to manual acquisition of the same information.</Paragraph>
    <Paragraph position="2"> Recently, we have begun porting our ACFG natural language system (Boisen, et al, 1989b) to a personnel database domain, a domain for which Parlance had been configured using the Learner. This raised the possibility of using the output files created by the Learner as knowledge sources for components of the ACFG system. This is a particularly interesting test, since the Learner is not designed to be a general acquisition tool for arbitrary natural language processing systems, but is optimized to produce information necessary for the Parlance system. While Parlance is an ATN-based system, the BBN ACFG utilizes a unification grammar formalism, similar to a Definite-Clause Grammar; the two systems, then, are quite different. If, despite this difference, the same Learner output could be used to configure the ACFG system, this would suggest that the speed-up of using the Learner to port to a new domain already demonstrated for the Parlance system, could be extended to the ACFG system and, perhaps, to similar unification based grammars.</Paragraph>
  </Section>
  <Section position="5" start_page="241" end_page="241" type="metho">
    <SectionTitle>
USING LEARNER OUTPUT
</SectionTitle>
    <Paragraph position="0"> We have taken the output files produced by the Learner and developed software tools to convert them into forms usable by the ACFG system. We have experimented with acquiring three kinds of knowledge:  In the rest of this section, we discuss in somewhat more detail the information provided by the Learner in each of these areas. In the next section, we discuss the results of our efforts.</Paragraph>
  </Section>
  <Section position="6" start_page="241" end_page="242" type="metho">
    <SectionTitle>
SYNTACTIC AND SEMANTIC INFORMATION
</SectionTitle>
    <Paragraph position="0"> We have used Learner output to acquire lexical entries for the open class categories: nouns (both common and proper), adjectives, and verbs, that include both syntactic and semantic information. For each of these categories, the Learner provides any necessary molphological information, such as inflectional paradigm (e.g. that city forms its plural by affixing -es), associated irregular forms (e.g. that got is the past tense of get), etc. In addition, the Learner outputs information specific to each category: Nouns: For ordinary entity type nouns, such as programmer, the Learner provides information about the semantic type of the noun and of the underlying concept with which it is associated (e.g. that a programmer is one of type person whose area of expertise is programming). For relational nouns, such as salary, the Learner provides information about the underlying relation associated with the noun, the domain of the relation, and the range of the relation (e.g. that the underlying relation of salary is salary-of, which applies to a person and produces a monetary-amount).</Paragraph>
    <Paragraph position="1"> Adjectives: For adjectives, the Learner provides informarion about the type of noun to which the adjective can be applied and about the underlying function with which the adjective is associated (e.g. the adjective asian american is applied to nouns of type person and is true of those whose ethnic group is asian-american).</Paragraph>
    <Paragraph position="2"> Verbs: For verbs, the Learner provides information about the underlying relation associated with the verb, the type restrictions associated with its noun phrase arguments, as well as any prepositions that the verb may select; for example, for the verb graduate, a noun of type person does the graduating,  the preposition from is used to mark the place from which the graduation took place, which is a noun of type school.</Paragraph>
    <Paragraph position="3"> In addition to information about specific lexical items, the Learner produces information about the underlying concepts of the domain; for example, that there is a relation gender-of that applies to persons and produces a result of type genders. We have also used the Learner output to acquire this information.</Paragraph>
  </Section>
  <Section position="7" start_page="242" end_page="242" type="metho">
    <SectionTitle>
DATABASE INFORMATION
</SectionTitle>
    <Paragraph position="0"> As part of its output, the Learner produces a file of pattern transformation rules, which map from concepts in the semantic domain model---and, so, ultimately, from words associated with those concepts--to fields in the data base. This file contains the information that allows the associated natural language system to actually obtain an answer from the database. Since the database system used in the ACFG system--described in (Boisen, 1989)---is essentially a modified version of that used by Parlance, these rules are straightforwardly usable.</Paragraph>
  </Section>
  <Section position="8" start_page="242" end_page="242" type="metho">
    <SectionTitle>
WORDS AND WORD CLASSES
</SectionTitle>
    <Paragraph position="0"> We have also used the Learner to acquire a set of vocabulary items and word classes for a class grammar (Derr and Schwartz, 1989) for use in HARC, the BBN Spoken Language System (Boisen, et al, 1989a), which incorporates the ACFG system as its natural language component. Though the Learner does not directly produce a class grammar, we used the syntactic categories, semantic classes, and inflectional paradigms and forms which it provides to produce a class grammar. To create a complete set of words for our speech recognition system, we did not use the Learner output directly, since it only contains the base and irregularly inflected forms of words, but a lexicon that was created on the basis of the Learner output and which included inflected forms, as well.</Paragraph>
  </Section>
  <Section position="9" start_page="242" end_page="242" type="metho">
    <SectionTitle>
RESULTS
</SectionTitle>
    <Paragraph position="0"> Since our experiment in obtaining information from existing Learner output was performed at the same time that we were writing the code to perform the necessary translations, we cannot measure the efficiency of using the Learner output in terms of elapsed time or person weeks. Therefore, as a rough measure of the benefits of using Learner output as a source, we propose to compare the number of items obtained from the translation program which were usable without further manual modification with the number that required some hand editing.</Paragraph>
  </Section>
  <Section position="10" start_page="242" end_page="242" type="metho">
    <SectionTitle>
SYNTACTIC AND SEMANTIC INFORMATION
</SectionTitle>
    <Paragraph position="0"> A total of 1499 lexical items were acquired from the Learner output; of these 1379 (91%) were directly usable, without any human intervention. Since there were many proper nouns in the lexicon obtained, and since proper nouns typically did not need to be edited, we also present the percentage of lexical items that were immediately usable excluding the proper nouns, so that their presence does not bias the result; in this case, 74% of the derived lexicai items were directly usable. These results are shown in the following table:  Some comments are in order about the manual editfing required. In the case of nouns and adjectives, some editting was required for syntactic and morphological features, owing to differences between the grammars of the Parlance and ACFG systems. All of the semantic information for nouns and adjectives, however, was left untouched. In the case of verbs, on the other hand, the difference between the Learner and ACFG semantic representations was too great to allow automatic acquisition of semantic information; for verb entries, the semantic portion was written by hand. However, even in the case of verbs, the semantic information was not written from scratch; rather, the semantic entry for each verb was a manual translation of the information in the Learner output.</Paragraph>
    <Paragraph position="1"> We also obtained a total of 109 semantic concepts from the Parlance output, which were used in the semantic entries of lexical items. These concepts required no manual editting at all.</Paragraph>
  </Section>
  <Section position="11" start_page="242" end_page="243" type="metho">
    <SectionTitle>
DATABASE INFORMATION
</SectionTitle>
    <Paragraph position="0"> For this domain, the Learner produced 65 translation rules, all of wMch were usable unedited.</Paragraph>
  </Section>
  <Section position="12" start_page="243" end_page="243" type="metho">
    <SectionTitle>
WORDS AND WORD CLASSES
</SectionTitle>
    <Paragraph position="0"> The lexicon derived from the Learner was used to create a speech vocabulary of 2170 items, with an associated class grammar of 637 classes with a perplexity of 89, with a small amount of manual editing.</Paragraph>
  </Section>
class="xml-element"></Paper>