<?xml version="1.0" standalone="yes"?>
<Paper uid="H86-1022">
  <Title>A LOGICAL-FORM AND KNOWLEDGE-BASE DESIGN FOR NATURAL LANGUAGE GENERATION 12 /</Title>
  <Section position="4" start_page="231" end_page="232" type="metho">
    <SectionTitle>
2. BASIC COMPONENTS
</SectionTitle>
    <Paragraph position="0"> The processes and representations we have employed include a unique linguistic component (Nigel), a frame-based network knowledge representation (NIKL), a propositional reasoner that can take advantage of the network knowledge representation (KL-TWO), and our own first order logic meaning representation.</Paragraph>
    <Section position="1" start_page="231" end_page="232" type="sub_section">
      <SectionTitle>
2.1. Nigel
</SectionTitle>
      <Paragraph position="0"> The Nigel grammar and generator realizes the functional systemic framework \[Halliday 76\] at the level of sentence generation. Within this framework, language is viewed as offering a set of grammatical choices to its speakers. Speakers make their choices based on the information they wish to convey and the discourse context they find themselves in. Nigel captures the first of these notions by organizing minimal sets of choices into systems. The grammar is actually just a collection of these systems. The factors the speaker considers in evaluating his communicative goal are shown by questions called inquiries \[Mann 83a\]. A choice alternative in a system is chosen according to the responses to one or more of these inquiries.</Paragraph>
      <Paragraph position="1"> For example, because processes with addressees are grammatically different from other processes, the grammar has an inquiry, VerbalProcessQ, to test whether the process is one of communication. Elsewhere, as part of deciding on number, Nigel has an inquiry MultiplicityQ that determines whether an object being described is unary or multiple. These are examples of information characterization inquiries.</Paragraph>
      <Paragraph position="2"> Another type of inquiry, called information decomposition, picks out of the environment the conceptual entities to be described. For example, at appropriate times, Nigel asks for descriptions of the causers of events, CauserlD, or the objects affected in them, AffectedlD.</Paragraph>
      <Paragraph position="3"> One very special inquiry, TermSpecificationlD, establishes the words that will be used. Nigel asks the environment for a set of lexical entries that can be used to describe an entity. So Nigel might find itself being told to describe some event as a &amp;quot;Send&amp;quot; or some object as a &amp;quot;Message&amp;quot;.</Paragraph>
      <Paragraph position="4"> Nigel currently has over 230 systems and 420 inquiries and covers a large subset of English.</Paragraph>
      <Paragraph position="5"> Up until the effort described here, the developers of Nigel had only identified the inquiries of the grammar, but not implemented them.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="232" end_page="233" type="metho">
    <SectionTitle>
LF and KB Design for Generation
2.2. NIKL
</SectionTitle>
    <Paragraph position="0"> NIKL is a network knowledge-base system descended from KL-ONE \[Brachman and Sohmolze 85\]. This type of reasoner supports description of the categories of objects, actions, and states of affairs that make up a domain. The central components of the notation are sets of concepts and roles, organized in IS-A hierarchies. The concepts are used to identify the categories of entities. The roles are associated with concepts (as &amp;quot;role restrictions&amp;quot;), and identify the relationships that can hold between actual individuals that belong to the categories. The IS.A hierarchies identify when membership in one category (or the holding of one relationship) entails membership in (or the holding of) another.</Paragraph>
    <Paragraph position="1"> We have experimented with a mail and calendar NIKL domain model developed for the Consul project \[Kaczmarek, Mark, and Sondheimer 83\]. It has a concept Send that is meant to identify the activity of sending messages. Send IS-A type of Transmit (intended to identify the general activity of transmission of information). Send is distinguished from Transmit by having s role restriction actee that relates Sends to Messages. The concept of Message is defined as being a kind of a communication object, through the IS-A relation to a concept Communication. In addition, role restrictions connect Message to properties of messages which serve to distinguish it from other communication objects. The overall model has over 850 concepts with over 150 roles.</Paragraph>
    <Paragraph position="2"> In flavor, NIKL is a frame system, with the concepts equivalent to frames and the role restrictions to slots. However, the NIKL representation can be given a formal semantics.</Paragraph>
    <Paragraph position="3"> 2.3. KL-TWO KL-TWO is a hybrid knowledge representation system that uses NIKL's formal semantics. KL-TWO links another reasoner, PENNI, to NIKL. For our purposes, PENNI can be viewed as restricted to reasoning using propositional logic 3. As such, PENNI is more restricted than those systems that use first order logic and a general purpose theorem prover. But it is also more efficient.</Paragraph>
    <Paragraph position="4"> PENNI can be viewed as managing a data base of propositions of the form (P a) and (Q a b) where the forms are variable free 4. The first item in each ordered pair is the name of a concept in an associated NIKL network and the first item in each ordered triple is the name of a role in that network. So the assertion of any form (P a) is a statement that the individual a is a kind of thing described by the concept P. Furthermore, the assertion (Q a b) states that the individuals a and b are related by the abstract relation described by Q.</Paragraph>
    <Paragraph position="5"> NIKL adds to PENNI the ability to do taxonomic reasoning. Assume the NIKL database contained the concepts just described in discussing NIKL. Assume that we assert just the following three facts: (Transmit x), (actee x y) and (Message y). Using the knowledge base, PENNI is able to recognize that any Transmit, all of whose actees are Messages, is a Send. So if we ask if (Send x) is true, KL-TWO will reply positively.</Paragraph>
    <Paragraph position="6"> KL-TWO can also retrieve information from its database. For example, if asked what individuals were the actees of x, it could respond with y.</Paragraph>
    <Paragraph position="7"> 2.4. THE LOGICAL LANGUAGE Our logical language is based on first order logic. To it, we have added restricted quantification, i.e., the ability to restrict the set quantified over. In addition, we allow for equality and some related quantifiers and operators, such as the quantifier for &amp;quot;there exists exactly one ...&amp;quot; (3!) and the operator for &amp;quot;the one thing that ...&amp;quot; (C/). We permit the formation and manipulation of sets, including a predicate for set membership (ELEMENT-OF). And we have some quantifiers and operators based on Habel's 'r/ operator \[Habel 82\].</Paragraph>
    <Paragraph position="8"> Figure 2-1 gives three examples of the forms accepted. Included are a few individuals: two people (RICHARD and SMITH), the computer (COMPUTER), a set of messages (MM33), and the current time (NOW). Later on, we will show how these are turned into English by our system.</Paragraph>
    <Paragraph position="9"> We include in our language a theory of the categories of conceptual entities and their relationships. We have taken what is often referred to as a Davidson approach \[Davidso n 67\]. This is marked by quantification over events and state of affairs. We</Paragraph>
    <Paragraph position="11"> Figu re 2-1 : Example Logical Expressions entities. We differ from Davidson by identifying a class of abstract Actions and Relations that are recorded by ActionOccurrences and RelationOccurrences 5. With Actions and Relations, we associate the participants and circumstances of the actions and states-of-affairs, e.g., the actor and actee.</Paragraph>
    <Paragraph position="12"> In addition to using the logical language for the demands for expression, we use it to maintain a database of factual information. Besides the &amp;quot;facts of the world&amp;quot;, we assume the availability of such knowledge as: &amp;quot;Hearer, speaker, time and place.</Paragraph>
    <Paragraph position="13"> *The theme of the ongoing discussion.</Paragraph>
    <Paragraph position="14"> &amp;quot;Objects assumed to be identifiable to the hearer.</Paragraph>
    <Paragraph position="15"> Work on maintaining this database is proceeding in parallel with the work reported here. Finally, we have allowed for a speech act operator to be supplied to the generation system along with the logical form. This can be ASSERT, COMMAND, QUERY, ANSWER or REFER. ANSWER is used for Yes/No answers. The others are given the usual interpretation.</Paragraph>
  </Section>
  <Section position="6" start_page="233" end_page="236" type="metho">
    <SectionTitle>
3. CONNECTING LANGUAGE AND LOGIC
</SectionTitle>
    <Paragraph position="0"> Restating the general problem in terms of our basic components, a logical form submitted as a demand for expression must be interpreted by the Nigel inquiries. Nigel must be able to decompose the expressions and characterize their parts.</Paragraph>
    <Paragraph position="1"> To achieve this, we-have-used N!KL to categorize the concepts (or terms) of the domain in terms of Nigel's implicit categorizations. We have written Nigel inquiries which use the structure of the logical language and the NIKL model to analyze the logical forms. To do this efficiently, we have developed a way to translate the logical form into a KL-TWO database and use its reasoning mechanisms.</Paragraph>
    <Paragraph position="2"> 3.1. Functional Systemic Categorizations in a NIKL Knowledge Base Our NIKL knowledge base is structured in layers. At the top are concepts and roles that reflect the structure we impose on our logical forms. Here we find concepts like ActionOccurrence and Action, as well as roles like records. At the bottom are the entities of the domain. Here we find concepts like Transmit and Send, as well as roles like requestedobject. All of these concepts and roles must be shown as specializing concepts at a third, intermediate level, which we have introduced to support Nigel's generation.</Paragraph>
    <Paragraph position="3"> Functional systemic linguists take a Whorfian view: that there is a strong connection between the structures of thought and language. We follow them in categorizing domain concepts in a way that reflects the different linguistic structures that describe them. For example, we have distinguished three types of actions (verbal, mental and material) because the clauses that describe these actions differ in structure. We have at least three types of relations (ascription, circumstantial and  LF and KB Design for Generation generalized possession) for the same reason s .</Paragraph>
    <Paragraph position="4"> Some of these categories are shown graphically in Figure 3-1. The double arrows are the IS-A links. The single arrows are role restrictions.</Paragraph>
    <Paragraph position="5"> Figu re 3-1 : Example Upper and Intermediate Layer Categories Relating these distinctions to our earlier examples, the concepts Transmit and Send are modelled as subclasses of MaterialAction. Message is a kind of NonConsciousThing.</Paragraph>
    <Paragraph position="6"> This modelling extends to the role hierarchy, as well. For example, the role requestedobject is modelled as a kind of actee role.</Paragraph>
    <Paragraph position="7"> The insertion of systemic distinctions does not compromise other factors, since non-linguistic categorizations can co-exist in the knowledge base with the systemic categories.</Paragraph>
    <Paragraph position="8"> Once the domain model is built we expect the systems using the generator to never have to refer to our middle level concepts. Furthermore, we expect Nigel to never refer to any domain concepts.</Paragraph>
    <Paragraph position="9"> Since the domain concepts are organized under our middle level, we can note that all domain predicates in logical forms are categorized in systemic terms. To be complete, the domain model must identify each unary predicate with a concept and each binary predicate with a role. The concepts in a logical form must either reflect the highest, most general, concepts in the network or the lowest layer. The domain predicates must therefore relate through domain concepts to systemic Fategories.</Paragraph>
    <Section position="1" start_page="234" end_page="235" type="sub_section">
      <SectionTitle>
3.2. Logical Forms in KL-TWO
</SectionTitle>
      <Paragraph position="0"> Gary Hendrix \[Hendrix 75\] developed the notion of Partitioned Semantic Networks in order to add the representational power of quantifier scoping, belief spaces, etc., to the semantic network formalism. This does not pay off in terms of faster inferences, but it allows us to separate the two structures inherent in logical formulas, the quantification scopes and the connections of terms. In partitioned networks, these are represented by hierarchically ordered partitions and network arcs, respectively.</Paragraph>
      <Paragraph position="1"> This separation of the scope and connection structure is needed. The connection structure can be used to evaluate Nigel's inquiries against the model, and the scope structure can be used to infer additional information concerning quantification.</Paragraph>
      <Paragraph position="2"> We translate a logical form into an equivalent KL-TWO structure. All predications appearing in the logical form are put into the PENNI database as assertions. Figure 3-2 shows the set of assertions entered for the formula in Figure 2-1A. These are 6Rougrdy. ascription relates ,in object to in intrinsic properly, such ilS its color. Orcumstantials involve time. place, inslnJment, etc. In addition to owner$hip. ger~raliz~ po~,ess4on includes such relat~onshil~ as i:wl/who4e anti social Nsoc~ation.</Paragraph>
      <Paragraph position="3">  LF and KB Design for Generation shown graphically in Figure 3-3 which includes the partitions. KL-TWO does not support partitions. Instead of cresting scope partitions, a tree is crested which reflects the variable scoping 7.</Paragraph>
      <Paragraph position="5"> Figu re 3-3: Sample Partition Structure During the translation, the variables and constants are given unique names so that these assertions are not confused with true assertional knowledge (this is not shown in our examples.). These new entities may be viewed as a kind of hypothetical object that Nigel will describe, but the original logical meaning may still be derived by inspecting the assertions and the scope structure.</Paragraph>
    </Section>
    <Section position="2" start_page="235" end_page="236" type="sub_section">
      <SectionTitle>
3.3. Implementation of Nigel Inquiries
</SectionTitle>
      <Paragraph position="0"> Our implementation of Nigel's inquiries using the connection and scope structures with the NIKL upper structure is fairly straightforward to describe. Since the logical forms reflecting the world view are in the highest level of the NIKL model, the information decomposition inquiries use these structures to do search and retrieval. With all of the predicates in the domain specializing concepts in the functional systemic level of the NIKL model, information characterization inquiries that consider aspects of the connection structure can test for the truth of appropriate PENNI propositions. The inquiries that relate to information presented in the quantification structure of the logical form will search the scope structure. Finally, to supply lexical entries, we associate lexical entries with NIKL concepts as attached data and use the retrieval methods of PENNI and NIKL to retrieve the appropriate terms.</Paragraph>
      <Paragraph position="1"> Let's consider someexamples. The generation activity begins with a pointer to the major ProcessOccu rrence. By the time CauserlD is asked, Nigel has a pointer to what it knows to be a caused Action. CauserlD is realized by a procedure that finds the thing or things that are in actor type relationships to the Action. AffectedlD works similarly through the actee predicate. When VerbalProcessQ is asked, Nigel simply asks PENNI if a proposition with VerbalAction and the Action is true.</Paragraph>
      <Paragraph position="2"> These examples emphasize the use of the connection structure to analyze what functional systemic grammarians call the ideational content of an utterance. In addition, utterances are characterized by interpersonal content, e.g., the relation between the hearer and the speaker, and textual content, e.g., relation to the last utterance. We have been developing methods for stOring this information in a PENNI database, so that interpersonal and textual inquiries can also be answered by asking questions of PENNI.</Paragraph>
      <Paragraph position="3"> MultiplicityQ is an example of a more involved process. When it is invoked, Nigel has a pointer to an individual to be described. The inquiry identifies all sets as multiple and any non-set individuals as unitary. For non-set variables, it explores their scoping environment. Its most interesting property involves an entity whose quantification suggests an answer of unitary. If the entity is shown in the logical form as a property of or a part of some entity and it is inside the scope of the quantifier that binds that entity and this second entity must be treated as multiple, then both entities are said to be multiple.</Paragraph>
      <Paragraph position="4">  LF and KB Design for Generation TermSpecificationlD is unique in that it explores the NIKL network directly. It is given a pointer to a PENNI individual. It accesses the most specific genedc concept PENNI has constructed to descdbe the individual. It looks at this concept and then up through more general categodes until it finds a lexical entry associated with a concept.</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="236" end_page="236" type="metho">
    <SectionTitle>
4. EXAMPLE SENTENCES
</SectionTitle>
    <Paragraph position="0"> Space constraints forbid presentation of a complete example. Let's look at a few points involved in transforming the three example logical forms in Figure 2-1 into English. Assume for Example 2-1A, that, at this moment, the COMPUTER wishes to communicate to RICHARD the information as an assertion, and that SMITH is known by name through the PENNI database.</Paragraph>
    <Paragraph position="1"> The flow starts with x identified as the central ProcessOccurrence. From there, y is identified as describing the main process.</Paragraph>
    <Paragraph position="2"> TermSpecificationlD is applied to y in one of the first inquiries processed. This is stated to be a Transmit. However, we are also told that its actee is a Message. Assuming the model described in Section 2.2, PENNi concludes that y is not just a Transmit, but a Send as well. This leads TermSpecificationlD to look first at Send for a lexical entry.</Paragraph>
    <Paragraph position="3"> Next, Nigel asks for a pointer to the time being referred to and receives back p. Later this is evaluated against the speaking time to establish the tense.</Paragraph>
    <Paragraph position="4"> Further on, Nigel characterizes the process. The inquiries attempt to prove, in turn, that y is a Relation, a MentalActive, and a VerbalAction. When none of these queries are answered positively, it concludes that y is a MaterialAction.</Paragraph>
    <Paragraph position="5"> After establishing that y is a kind of event that is caused, Nigel uses CauserlD and AffectedlD. It receives back SMITH and z, respectively.</Paragraph>
    <Paragraph position="6"> The actual decision on how to describe SMITH and z are arrived at during separate passes through the grammar. During the pass for SMITH, TermSpecificationlD returns his name, &amp;quot;Smith&amp;quot;. MultiplicityQ is invoked and returns unitary. During the pass for z, TermSpecificationlD returns &amp;quot;message&amp;quot;, while MultiplicityQ returns unitary.</Paragraph>
    <Paragraph position="7"> In the end, the sentence &amp;quot;Smith sent a message.&amp;quot; is generated.</Paragraph>
    <Paragraph position="8"> Looking at Example 2-1B, one difference on the level of the outermost ActionOccurrence is the absence of an actee relation. However, requestedobject is shown in the model as a type of actee relation and AffectedlD returns q. In order to describe q the grammar forms a relative clause, &amp;quot;which was sent by Smith&amp;quot;. There is no overt indication of the type of entity q is. However, from the model of Send, PENNI infers that (Message z) is true. TermSpecificationlD for z returns &amp;quot;message&amp;quot;. Treating the sentence as a command and assuming &amp;quot;show&amp;quot; is associated with Display, Nigel will produce &amp;quot;Show me the message which was sent by Smith.&amp;quot;.</Paragraph>
    <Paragraph position="9"> Example 2-1C allows us to consider the use of the scope structure in deciding the multiplicity of r. We are required to describe the displaying of the single inspection status (or read status) that is found for each message in a set of messages. As noted, we have modelled InspectionStatus as an Ascription relation. The grammar uses &amp;quot;of&amp;quot; to describe this sort of relation in a noun phrase. MultiplicityQ evaluates m as multiple. Because r is in m's scope, it too is evaluated as multiple and the noun phrase is rendered as &amp;quot;the read statuses of the messages&amp;quot;. If the scopings were reversed, the logical form would indicate that there was only one read status for all the messages. MultiplicityQ would evaluate r as unitary and the noun phrase would be &amp;quot;the read status of the messages&amp;quot;. If both the quantifiers were existential, then each scoping would result in MultiplicityQ evaluating both as unitary. The noun phrase would be rendered as &amp;quot;the read status of the message&amp;quot;. If m were bound by an 3t, bound by an ~ or replaced by a unitary constant, and r bound by a universal quantifier, the rendering would be &amp;quot;the read statuses of the message&amp;quot;.</Paragraph>
    <Paragraph position="10"> In Figure 4-1, we display a set of sentences to give the reader some idea of the generator's range as of January 1986. Nigel played the part of both participants in the dialogue, which used hand constructed logical forms and dialogue contexts in the absence of an associated software system.</Paragraph>
  </Section>
  <Section position="8" start_page="236" end_page="238" type="metho">
    <SectionTitle>
5. CONCLUSION
</SectionTitle>
    <Paragraph position="0"> asked me to display it.</Paragraph>
    <Paragraph position="1"> Figu re 4.1 : A Sample Set of, Generated Sentences</Paragraph>
    <Section position="1" start_page="237" end_page="237" type="sub_section">
      <SectionTitle>
5.1. Summary
</SectionTitle>
      <Paragraph position="0"> To summarize, we have developed a first-order predicate-calculus language which can be used to make demands for expressions to the Nigel grammar. This works by translating the logical forms into two separate structures that are stored in a PENNI database. Nigel inquiries are evaluated against these structures through the aid of a NIKL knowledge base. Discourse context is also stored in the data base and lexical entries are obtained from the knowledge base.</Paragraph>
      <Paragraph position="1"> Adding this facility to Nigel seems to have added only 10 to 20 percent to Nigel's run time.</Paragraph>
    </Section>
    <Section position="2" start_page="237" end_page="238" type="sub_section">
      <SectionTitle>
5.2. Limitations and Future Plans
</SectionTitle>
      <Paragraph position="0"> For the sake of presentation, we have simplified our description of the working system. Other facilities include an extensive tense, aspect and temporal reference system. There is also a facility for dynamically constructing logical forms for referring expressions. This is used when constants are found in other logical forms that cannot be referred to by name or through pronoun.</Paragraph>
      <Paragraph position="1"> There are also certain limitations in our approach. One of which may have occurred to the reader is that the language our system produces is ambiguous in ways formal logic is not. For example, &amp;quot;the read statuses of the messages&amp;quot; has one reading which is different from the logical form we used in our example. While scope ambiguities are deeply ingrained in language, they are not a problem in most communication situations.</Paragraph>
      <Paragraph position="2"> Related to this problem is a potentially important mismatch between logic and functional systemic grammars. These grammars do not control directly for quantification scope. They treat it as only one aspect of the decision making process about determiner choice and constituent ordering. Certainly, there is a great deal of evidence that logical scoping is not often a factor in the interpretation of utterances 8.</Paragraph>
      <Paragraph position="3"> Another set of problems concern the limits we place on logical connectives in logical forms. One limit is the the position of negations: we can only negate ProcessOccurrences, e.g., &amp;quot;John didn't send a message.&amp;quot;. Negation on other forms, e.g., &amp;quot;John sent no messages.&amp;quot;, affects the basic connection with the NIKL model. Furthermore, certain conjunctions have to be shown with a conjunctive Relation as opposed to logical conjunction. This includes conjunctions between ProcessOccu rrences that lead to compound sentences, as well as all disjunctions.</Paragraph>
      <Paragraph position="4"> Furthermore, we impose a condition that a demand for expression must concern a single connected set of structures. In operation the system actually ignores parts of the logical form that are independent of the main ProcessOccurrences.</Paragraph>
      <Paragraph position="5"> Because the underlying grammar can only express one event or state of affair(not counting dependent processes) and its associated circumstances at a time, in order to fit in one sentence all the entities to be mentioned must be somehow connected to one event or state of affair.</Paragraph>
      <Paragraph position="6"> We expect that the limitations in the last two paragraphs will be overcome as we develop our text planning system, 8For example, Keenan and Faltz state &amp;quot;We feel that the reason for the poor correspondence is that NP Scope differences in natural language are not in fact coded or in general reflected in the derivational history of an expression. If so, we have a situation where we need something in LF which really doesn't correponcl to anything in SF&amp;quot; \[Keenan 85, p. 21\].</Paragraph>
      <Paragraph position="7">  LF and KB Design for Generation Penman \[Mann 83b\]. A theory of text structure is being developed at USC/ISI that will take less restrained forms and map them into multi-sentence text structures \[Mann 84\]. The use of this intermediate facility will mediate for logical connectives and connectivity by presenting the sentence generator with normalized and connected structures.</Paragraph>
      <Paragraph position="8"> The word choice decisions the system makes also need to be enhanced. It currently takes as specific a term as possible. Unfortunately, this term could convey only part of the necessary information. Or it could convey more information than that conveyed by the process alone, e.g., in our transmit/send example, &amp;quot;send&amp;quot;, unlike &amp;quot;transmit&amp;quot;, conveys the existence of a message. We are currently developing a method of dealing with word choice through descriptions in terms of primitive concepts that will support better matching between demands and lexical resources.</Paragraph>
      <Paragraph position="9"> A related limit is the requirement in the current NIKL that all aspects of a concept be present in the logical form in order for the NIKL classifier to have effect. For example, the logical forms must show all aspects of a Send to identify a Transmit as one. A complete model of Send will certainly have more role restrictions than the actee. However, just having an actee which is a Message should be sufficient to indicate that a particular Transmit is a Send. We are working with the developers of NIKL to allow for this type of reasoning.</Paragraph>
      <Paragraph position="10"> Two other areas of concern relate directly to our most important current activity. First, it is not clear that first-order logic will be sufficiently expressive for all possible situations. Second, it is not clear the use of hand-built logical forms is sufficient to test our design to its fullest exteot.</Paragraph>
    </Section>
  </Section>
  <Section position="9" start_page="238" end_page="238" type="metho">
    <SectionTitle>
5.3. JANUS
</SectionTitle>
    <Paragraph position="0"> The success of our work to date has led to plans for the inclusion of this design in the Janus natural language interface.</Paragraph>
    <Paragraph position="1"> Janus is a joint effort between USC/ISI and BBN, Inc., to build the next generation natural language interface within the natural language technology component of the Strategic Computing Initiative \[Walker 85\]. One feature of the system will be the use of higher-order logics. Plans are underway to test the system in actual use. The future direction of the work presented here will be largely determined by the demands of the Janus effort.</Paragraph>
  </Section>
  <Section position="10" start_page="238" end_page="239" type="metho">
    <SectionTitle>
Acknowledgments
</SectionTitle>
    <Paragraph position="0"> We gratefully acknowledge the assistance of our colleagues Bill Mann, Richard Whitney, Tom Galloway, Robert AIbano, Susanna Cumming, Lynn Poulton, Christian Matthiessen and Marc Vilain.</Paragraph>
  </Section>
class="xml-element"></Paper>