<?xml version="1.0" standalone="yes"?>
<Paper uid="C82-1030">
  <Title>TULIPS-2 - NATURAL LANGUAGE LEARNING SYSTEM</Title>
  <Section position="2" start_page="191" end_page="191" type="abstr">
    <SectionTitle>
192 M.G. MALKOVSKY
</SectionTitle>
    <Paragraph position="0"> in an unknown situation but to acquire mere knowledge, i.e. to learn the language and the user's speech peculiarities.</Paragraph>
    <Paragraph position="1"> The experimental system TULIPS (Malkovsky (1975)) and its new version TULIPS-2 (Malkovsky and Volkeva (1981)) both were designed in consideration of the above-mentioned demands.</Paragraph>
    <Paragraph position="2"> The AI system TULIPS-2 implemented in PLANNER for the BESM-6 computer is intended for further experiments in the field of the computer understanding of NL and for practical use. The system can help the user to form the conditions of a problem. In this case the user gives the system the unformalized description of the problem situation, whereas the system helps to specify this description and to find an adequate formal representation. Such a flexible dialogue using vague terms and loose concepts can be ccnviniently performed just in a NL (Russian - for TULIPS-2).</Paragraph>
    <Paragraph position="3"> ~oreover the TULIPS-2 system can work in problem-domains with various structures and degrees of formalization. That is another argument for the use of NL.</Paragraph>
    <Paragraph position="4"> A user's interaction with the system (via a terminal) is composed of several seances. At the begining of each seance the user have to identify himself and to indicate the problem-domain. This informstion guides the &amp;quot;tuning&amp;quot; of the system for the ~eance, i.e. fetching the relevant data from the external memory. This helps to reduce data used in conversation. On the other hand the tuning process introduces the user's speech peculiarities and specific NL items of the problem-domain. During the analysis of utterances these peculiarities and items are looked through before all the other data (lexical, syntactic, and semantic).</Paragraph>
    <Paragraph position="5"> Besides, there are the following methods of data representation and handling in the system: special tags define the measure of preferability of relevant data items and procedures and influence the order of their choise during analysis; the lexical items and the grammar rules contain the references to procedures that can be invoked when an item or rule is being handled; NL meta-level items describe the means and range of the Russian language rules alternation by the system; NL knowledge of the system includes both basic knowledge of the Russian language and &amp;quot;open&amp;quot; set of Russian grammar rules, Russian lexical items etc., that can be widened in a seance by the user or by the system itself (&amp;quot;selftaeching&amp;quot;)o null It should be noted that the basic knowledge is formed and input into the system by its authors or by its operators beforehand.</Paragraph>
    <Paragraph position="6"> Thus in a seance the system starts to learn NL, to acquire user's speech peculiarities, new terms and abbreviations having much knowledge of NL which make it possible for the system to act in unknown situations by itself. However, change of basic knowledge can be done only with user's permissiondeg The methods of representation and handling of NL knowledge are important to the system's analyzer which provides for the input message understanding from the context of the conversation. Syntactic, semantic, and pragmatic predictions are widely used on different levels of analysis. The predictions generated from context make it possible to attribute the expected (predicted) characteristics to unknown units, while the references to procedural elements provide for a flexible control, i.e. the pos-TULIPS-2 - NATURAL LANGUAGE LEARNING SYSTEM 193 sibility of passing on to a more informative (where predictions are more definite) level of analysis.</Paragraph>
    <Paragraph position="7"> If necessary the analyzer appeales to the meta-level knowledge invokes procedures which handle unknown units (words or phrases).</Paragraph>
    <Paragraph position="8"> These procedures classify such a unit (erroneous form of a known unit or an unknown correct unit) and prepare the information of a unit or an error for storing. The stored information is available both in this seance and in the subsiquent ones.</Paragraph>
    <Paragraph position="9"> Sometimes a deviant form can be passed on to further higher levels of analysis, as e.g. the module of spelling correction does. This module processes errors typical for the user working at the terminal (the missing, duplication, permutation of letters or an incorrect shift). However, usuall~ as the result of learning (self-teaching or teaching by user) new items are formed and the old items are changed. The following item types are formed and changed: NL words and phrases descriptions - lexical items and grammar rules, NL meta-level items, control structures - tags and procedures (e.g. special patterns for frequent and typical phrases).</Paragraph>
    <Paragraph position="10"> The methods of learning on morphological and lexical levels of Russian have been used in the TULIPS-2 system since 1980. The basic knowledge for these levels includes: a complete description of Russian inflexion, a description of some rules of Russian word-formation and of different typical mistakes made by users, a vocabulary of about 1000 stems, and vocabularies of affixes.</Paragraph>
  </Section>
class="xml-element"></Paper>