<?xml version="1.0" standalone="yes"?>
<Paper uid="A88-1017">
  <Title>Structure from Anarchy: Meta Level Representation of Expert System Propositions for Natural Language Interfaces. 1</Title>
  <Section position="5" start_page="122" end_page="125" type="metho">
    <SectionTitle>
3 The Structure in More Detail
</SectionTitle>
    <Paragraph position="0"> To translate user input into facts and goals of an underlying expert system, a structure that is able to provide a foundation for the translation is necessary. This structure must provide the meaning of the expert system propositions, relationships between them and supply a means of mapping semantics of words and phrases into those propositions. It is also desirable that such a structure he general, and hence to some extent transportable from one system to another.</Paragraph>
    <Paragraph position="1">  Transfer of possession category 6 Our structure consists of a group of hierarchies formed from classes of verbs. We have analyzed over 90 verbs most common to our domain and classified them into 13 categories 7. These categories can be used in any domain that requires the verbs belonging to them, because they are derived from general properties of the verbs s, thus allowing for a degree of transportability. Each verb category is organized hierarchically where each node of a hierarchy is derived from the meanings of one or more verbs. A number of selectional restrictions is attached to each node indicating constraints on the the figure, * stanch for wild card, and . means that the feature is inherited flora the patent node.</Paragraph>
    <Paragraph position="2">  necessary.</Paragraph>
    <Paragraph position="3"> SThe.~ camgoriea am beusd on works m lingumics, e. 8. \[Osso~d 791 and an Roller's ~. Fix * mo~ d~cription of tim caleSoriej see \[Datakovsky Moeldler eLM. 87\] agent, patient, object and modifier of an input sentence (not all four restrictions are specified for every category). The hierarchies group propositions of an expert system by topic. The leaves of the hierarchies contain either expert system facts or pointers to other hierarchies, thus forming a connectexl forest. The top level nodes of the hierarchies provide general classes into which a group of propositions of an expert system might fall. At the lower levels of the hierarchies the propositions are separated into more specific subclasses of a given parent node, thus further specifying their meanings. At the lowest level, each node points to only one proposition thus uniquely def'ming it within its class. For example, figure 3 shows the partial hierarchy for the Transfer of Possession category. The top level node of the hierarchy is derived from the properties of the verbs of the general class of Transfer of Possession. Verbs from that class have pointers to this node and all the propositions that deal with transfer of possession can be accommodated by this node and the nodes below it. The selectional restrictions on this node indicate that the transfer is initiated by either a human or an organization and that the beneficiary of the transfer, the object being transferred, as well as any modifiers can be unspecified until some lower level. The two nodes at the next level further divide the class of transfer of possession verbs and predicates into those dealing with physical object transfers and non physical object transfers. The \[-\] in the selectional restrictions indicate that the feature is inherited from the parent node. The restrictions on the two nodes also further specify that the object being transferred must be concrete in order to take the Plays Obj link and abstract in order to take the Non Phys Obj link. At the next level, the concept of physical object transfers (as embodied by the Plays Obj node) is further specified. In this example only one of its children, the Money node is shown 9. Again, verbs dealing specifically with money transfers may point directly to this node. The restriction on the object of the transfer must be monetary in order for this node to be chosen during parsing. This node is further subdivided into Donation and Income, where the distinction is made based on the recipient of the transfer, since donations are normally given to organizations, and income to people. Next, Income can come in two forms, Taxable and Non Taxable, as indicated by the selectional restrictions of the objects  of the transfer, and finally, the bottom level of the hierarchy contains expert system propositions. The propositions (?dependent is gross_income ?income) and (?dependent is amount_of_support ?support) belong to a general class of Transfer of Possession, and a more specific class Income, indicating that both propositions describe a type of income that is generally transferred from one party to another.</Paragraph>
    <Paragraph position="4"> However, because one deals with taxable income and the other with non taxable income, these propositions are further subdivided into subclasses at the next level.</Paragraph>
    <Paragraph position="5"> This kind of gradual division of propositions into subclasses not only provides a means for mapping user input into facts and goals of an expert system, but also allows the system m answer questions about relationships between the propositions, often without any infercncing. In addition, it allows the system to make meta level inferences it could not make without the structure. In the next section we present a brief description of the parsing algorithm and illustrate it with an example.</Paragraph>
    <Section position="1" start_page="123" end_page="124" type="sub_section">
      <SectionTitle>
3.1 Parsing Algorithm: Overview and
</SectionTitle>
      <Paragraph position="0"> Example.</Paragraph>
      <Paragraph position="1"> During parsing, an appropriate hierarchy is selected according to the definition of the verb in the system's dictionary, where each verb can point to any level in a hierarchy, and a selectional restriction based algorithm is used to traverse the hierarchy with the nouns of the sentence guiding the parser down the hierarchy, until an expert system proposition is reached. The information for this algorithm is encoded into each hierarchy, with the restrictions on the arguments of the verbs based on noun featm'es derived from Roget's thesaurus. The system is currently being implemented in Common lisp on a Symbolics Lisp Machine. It uses an ATN parser which has been modified to call the semantics at various points before deciding which a~ to take next. Syntax and semantk:s run in parallel, with syntax providing a deep struclxn, e of a sentence, and semantics supplying infommtion for modifier attachment. Although the verb hierarchies are the primary source of facts, some facts are derived directly from the noun features.</Paragraph>
      <Paragraph position="2"> As an example of how the natural language interface derives both Wopositions and goals from Yes~No questions posed by the user consider the question Can i claim my son who earns a salary of $2000?. A trace of the system execution of this sentence is shown in appendix I. The trace shows the nodes of the different hierarchies considered by the algorithm and where the interaction between syntax and semantics occurs. It also shows all the predicates derived by the system and a complete syntactic parse. In yes/no questions the goal is generally indicated by the main verb. The syntactic parser identifies claim as the main verb of the sentence. The verb claim is defined in the system's dictionary as Classification &lt;+&gt; Dependency 1deg, indicating that the verb belongs to the general category of Classification and a more specific subnode of that category, Dependency. The &lt;+&gt; indicates that the syntactic subject of the sentence is the semantic agent. Based on the definition of the verb the algorithm enters the ClassOeu:ation hierarchy at the Dependency node, as demonstrated in stam~ents 1 and 2 of the system trace, thus limiting the choice of propositions that this input can map into to the general category of ClassO~cation and the subclass Dependency (see figure 4). Since only one proposition, (?user can_claim ?dependent), falls into this classification, it is derived as the goal, indicating that the user wants to know whether he can or can not claim a dependent (the variables of the proposition will later be instantiated with the appropriate values).</Paragraph>
      <Paragraph position="3"> The additional information in the relative clause states that the dependent earns a salary of $2000, or (?dependent is gross_income ?income). To derive this additional information, the system selects a hierarchy based on the meaning of the verb of the relative clause. The verb to earn is defined in the dictionary as Transfer of possession &lt;+&gt;, so the algorithm enters the Transfer of Possession hierarchy (shown in figure 3). The choice of propositions that this input can map into is now limited to those in the general class of Transfer of Possession. Next, because of the feature concrete of the object (two thousand dollars) of the sentence the algorithm selects Phys Obj as the next node to consider. Based on the feature monetary of the word dollars the Money node is selected next. The Income node is chosen because the recipient of the money has the feature human, and finally, because salary is defined as payment~earned, the node Tax is selected, since earned payments are generally taxable. Finally (?dependent is gross_income ?income) is added to the working memory. The variables ?dependent and tdegA/lhough there are other meaninp of the verb, this is the most fiequmtly used mcemng in the tax domain, so the system tries this caliph/fma.</Paragraph>
      <Paragraph position="4">  ?income are later instantiated with son and $2000 respectively. The derivation of this predicate can be seen in statement 5-13 of the system trace in appendix I.</Paragraph>
      <Paragraph position="5"> Propositions can also be derived from certain noun phrases. In this example, the phrase my son indicates the existence of a child-parent relationship. The system then checks for agreement between the head pronoun I and the possessive my and once this agreement is verifies maps the representation of this relationship into the proposition (?dependent is son_of ?user), as shown in statement 4 of the trace.</Paragraph>
      <Paragraph position="6">  The mapping of natural language into propositions of the expert system as demonstrated above is possible because of the classification of propositions and descriptions of their meanings provided by the hierarchies. Note that the hierarchies are used to def'me semantics of words of the natural language e.g. the verb to earn is directly related to the meta level structure, or the Transfer of Possession hierarchy. The strncture given by the hierarchies also provides a description of the propositions and gives similarities and differences between them. For example, both propositions (?dependent is gross_income ?income) and (?depeadeat is amount_of..support ?support) would have the general properties of the class Income, with unique features of their particular subclasses Tax and Non Tax. This unique classification allows for the mapping of the input in the above example into the aplxopriate proposition. It also allows the system to answer questions about the differences twaween the two propositions, as shown in the next section. Another benefit of this representation is that it provides the system with a way of dealing with input sentences like My son earns $2000, that do not completely specify a particular proposition. The sentence indicates that the desired proposition is in the class Income, and the system can proceed to specify the appropriate subclass by posing questions to the user without any additional inferencing on the part of the expert system. This particular capability of the algorithm will be discussed in greater detail in future work.</Paragraph>
    </Section>
    <Section position="2" start_page="124" end_page="125" type="sub_section">
      <SectionTitle>
3.2 Other Questions that can be Answered
</SectionTitle>
      <Paragraph position="0"> from the Hierarchies The hierarchies allow the system to handle a number of questions that could not be previously handled by the expert system, and answer other questions without invoking the inference process. In particular, these include questions that deal with relationships between facts and comparisons between sessions, as well as questions requiring general information.</Paragraph>
      <Paragraph position="1"> User: My daughter receives a stipend of $5000, while my son gets a salary of $2000.</Paragraph>
      <Paragraph position="2">  AS an example of questions that can be answered without invoking the inference process, consider the hypothetical example in figure 4 where the user tells the expert that his daughter receives a stipend of $5000, which translates into the proposition (daughter is amount_of..support 5000), since stipend is defined in the dictionary as payment.given. The fact that his son has a salary of $2000 translates into the proposition (son is gross income 2000). To answer the WHY question the system could check where the derivation paths for the two sets of inputs diverged, and the difference between the two subclasses would constitutes the answer. In this example the paths diverge at the Income node of the Transfer of possession hierarchy, thus the answer can be supplied by simply examining the hierarchy.</Paragraph>
      <Paragraph position="3">  The question in the first example required both a comparison between two derivation paths as well as the knowledge of the differences between two propositions. As a second example consider the question What kinds of family relationships are recognized by the tax code? This question is about general properties of the tax code and could not be handled by the expert system without the natural language interface, even though all the necessary information was already available in the system. To answer this question it is enough to search the hierarchies for a Relationship node with a child node that describes family relationships. Such a parent-child pair is found in the Possession hierarchy (see figure 6). The answer returned would consist of all the children found under this pair.</Paragraph>
      <Paragraph position="5"> The question handling algorithm is currently under design. To process WH questions the system must ftrst be able to deter\[nine whether it can be answered from the hierm~hies, or whether the inference engine of the expert system should be invoked. Many of the necessary clues that indicate the question type have been identified, however there is still some more work to be done on this, as well as on the implementation of the module. It is clear, however, that the hierarchies give the system the ability to handle many more types of questions than the expert system alone could handle, and in many instances allow questions to be answered without invoking the inference process of the expert system.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="125" end_page="126" type="metho">
    <SectionTitle>
4 Comparison with Previous Work: NLIs
</SectionTitle>
    <Paragraph position="0"> to Expert Systems and Other Work in Semantics There has been some effort to construct natural language interfaces to expert systems, namely Prospector \[Duda eL al. 79\] and Xcaiibur \[Cartxmell eLal. 83; Carbonell and Hayes 84\]. Prospector is one of the fast expert systems to communicate with its users in natural language. During the consultation the user simply describes what has been discovered at a given site by using patterns, built with the help of the Lifer \[Hendrix et. al. 78\] system, of the form &amp;quot;There is &lt;deposit&gt;', &amp;quot;There may be &lt;deposit&gt;&amp;quot;, etc. There is not much published information that describes Prospector's natural language module. We can only hypothesize that a very simple and limited set of sentences is accepted by the system based on sample system sessions.</Paragraph>
    <Paragraph position="1"> Xcalibur's interaction with the user greatly resembles that of a natural lang~mge interface to a data base system. Unlike systems such as Mycin, Xcalibur does not do most of the asking. It is not responsible for solving the user's problem, but rather the user has to know what he wants and query accordingly. Most expert systems are designed to solve a user's problem, and this property must be reflected in the interface. Xcalibur does not seem to be suitable as an interface for such systems because it is designed to retrieve information rather than solve a problem.</Paragraph>
    <Section position="1" start_page="125" end_page="126" type="sub_section">
      <SectionTitle>
4.1 Other work in Semantics
</SectionTitle>
      <Paragraph position="0"> Our work draws on Palmer's \[Palmer 85\], but is different from it in several ways. Palmer's Inference-driven sexnantic analysis is specifically designed for a finite, well-defined, i.e. limited domain. The main element of her approach is a set of partially instantiated logical terms, or semantic propositions, which capture the different relationships that can occur in a given domain.</Paragraph>
      <Paragraph position="1"> Unlike Palmer's work, our interpreter deals with a complex real world domain. It also makes a greater separation between domain specific and domain independent knowledge to allow for a degree of transportability. Also, while our semantics provides a hierarchical organization, Palmer's does not.</Paragraph>
      <Paragraph position="2"> Other work that has influenced our own also includes that of Graeme Hirst \[Hirst 83\] and Steve Lytinen tLytinen 84\]. One of the main differences between our work and the work mentioned above (including Palmer's) is that our semantics imposes a structure on top of an unstructured underlying system, which is not the goal of the work mentioned above.</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="126" end_page="126" type="metho">
    <SectionTitle>
5 Possible Automation of Hierarchy
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="126" end_page="126" type="sub_section">
      <SectionTitle>
Design
</SectionTitle>
      <Paragraph position="0"> The lack of automatic construction of the hierarchies and automatic classification of propositions in them is currently a limitation in our system. If, for a given domain, a certain tree has to be extended, such extension will have to be done by hand. Also, propositions have to be hand encoded in the hierarchies. This makes transportability to other domains more difficult. After the top level categories are selected, the rest of the nodes of the hierarchies and the propositions, as well as the selectional restrictions can not be done interactively. However, we feel that the hierarchies lend themselves to automation construction by an Expert System Expert, because they are based on the linguistic properties of the verbs in the domain, as well as on the knowledge of the meanings of propositions.</Paragraph>
      <Paragraph position="1"> In the future, we would like to design a customization phase similar to that of Team \[Martin, Appelt and Pereira 83; Grosz et al. 85\] and Teli \[-Ballard 86\]. With such a customization phase, a given expert, such as an Expert Systems Expert, can spend several horn's automatically building up the necessary parse Irees for a given domain. We feel that such a module would geatly enhance the system and make it much more usable.</Paragraph>
    </Section>
  </Section>
  <Section position="8" start_page="126" end_page="126" type="metho">
    <SectionTitle>
6 Conclusions and Future Research
</SectionTitle>
    <Paragraph position="0"> In this paper we presented a slructure for expert systems, similar to a dam base schema, that facilitates construction of natural language interfaces. This structure is based on verb classification and hierarchical structuring within each categocy. The hierarchies provide a grouping of expert system propositions into classes, thus capturing the similarities and differences betweea the pmlx~itions.</Paragraph>
    <Paragraph position="1"> This grouping provides a mal~ing between user input and the propositions of the expert system, as well as a mechanism for dealing with several types of questions without additional ~pert system inferencing. The structme provides a mechanism for answering questions that could not be previously handled by the expert system. It also provides a flexible and somewhat general mapping allowing for a degree of wansportability.</Paragraph>
    <Paragraph position="2"> One of our primary goals is to complete the implementation of our ideas. Processing of statements and yes/no questions has been fully implemented and the work on petagraph parsing and handling of semanUcally incomplete input is our current focus. In the future we plan to add such features as complete WH question processing and an automatic hierarchy construction algorithm.</Paragraph>
  </Section>
  <Section position="9" start_page="126" end_page="126" type="metho">
    <SectionTitle>
7 Acknowledgments
</SectionTitle>
    <Paragraph position="0"> I would like to thank my advisor, Kathleen McKeown for all her help and guidance in this work and Robert Ensor of AT&amp;T for his helpful comments.</Paragraph>
    <Paragraph position="1"> Appendix I (process '((can I claim my son who earns a salary of twothousand donars)))  1. In Tree: CLASSIFY 2. Considering the children of DEPENDENCY 3. the proposition that was derived is ((?USER ICAN_CLAIMI ?DEPENDENT)) back to syntax...</Paragraph>
    <Paragraph position="2"> 4. the proposition derived from the noun phrase (MY SON) is (?DEPENDENT IS ISON_OFI ?USER) 5. In Tree: TRANS OF_POS 6. Considering the children of TRANS OF POS beck to syntax...</Paragraph>
    <Paragraph position="3"> 7. Considering the children of TRANS_OF__POS back to syntax...</Paragraph>
    <Paragraph position="4"> 8. Considering the children of TRANS_OF..POS 9. Considering the children of IPHYS_OBJI 10. Considering the children of MONEY 11. Considering the children of INCOME 12. Considering the children of TAX 13. the proposition that was derived is</Paragraph>
  </Section>
class="xml-element"></Paper>