<?xml version="1.0" standalone="yes"?>
<Paper uid="H86-1009">
  <Title>Model-based Analysis of Messages about Equipment</Title>
  <Section position="4" start_page="73" end_page="74" type="metho">
    <SectionTitle>
DURING NORMAL START CYCLE OF 1A GAS TURBINE, APPROX 90 SEC
AFTER CLUTCH ENGAGEMENT, LOW LUBE OIL AND FAIL TO ENGAGE
ALARM WERE RECEIVED ON THE ACC. (ALL CONDITIONS WERE
NORMAL INITIALLY). SAC WAS REMOVED AND METAL CHUNKS
FOUND IN OIL PAN. LUBE OIL PUMP WAS REMOVED AND WAS FOUND
TO BE SEIZED. DRIVEN GEAR WAS SHEARED ON PUMP SHAFT.
</SectionTitle>
    <Paragraph position="0"> :: ~Substantially better rates have been cited for strongly expectation-based parsers, which are considered successful if they locate all the expected items within an input text.</Paragraph>
    <Paragraph position="1">  gear teeth, without specific names). While not raising any intrinsic difficulties, a domain of this size clearly provides a more rigorous test of our ability to acquire and organize domain knowledge than did many earlier &amp;quot;toy&amp;quot; domains.</Paragraph>
    <Paragraph position="2"> Another unusual aspect is the nature of the domain information. Scripts, for example, encode essentially procedural information (how to perform complex actions). The information for our domain, in contrast, is primarily structural (part-whole relationships, interconnections, etc.) and to a lesser degree functional. This difference is reflected in differences in the way the information is used - in particular, in-the analysis of noun phrases, as we shall see below. Our domain information bears greater resemblance to that used in some equipment simulation packages (e.g., STEAMER \[Honan 1984\]) and diagnosis packages \[Cantone 1983\] than it does to that conventionally seen in natural language systems. The domain knowledge plays a role in many phases of the language processing task: in the recovery of implicit operands and intersentential relations, in the analysis of noun-phrase reference, and in the determination of syntactic and semantic structure. In particular, we shall consider below its role in the processing of compound nominals, which appear frequently in such technical domains. There have been several prior studies of the processing of such compounds. The work both of Brachman \[1978\] and of McDonald and Hayes-Roth \[McDonald 1978\] emphasized the use of search procedures within semantic networks to identify the wide variety of implicit relations possible with compound nominals. We have also used network search techniques, although of a more directed sort. However, their work cited isolated examples from a variety of areas to show the generality of their approach, while we have been concerned with achieving detailed and thorough coverage within a narrower domain. Finin \[1980, 1986\] has sought to develop, within a sublanguage, general semantic categories for the relations and consituents involved in compounds. Although there are some similarities to our classification efforts, he also has aimed at providing a relatively broad and loose set of constraints. In contrast, the detailed knowledge in our equipment model -- provided for several purposes, of which noun phrase interpretation is only one -make possible much tighter constraints in our system.</Paragraph>
  </Section>
  <Section position="5" start_page="74" end_page="75" type="metho">
    <SectionTitle>
3. System overview
</SectionTitle>
    <Paragraph position="0"> The PROTEUS system has three major components: a syntactic analyzer, a semantic analyzer, and a discourse analyzer. The syntactic analyzer parses the input and regularizes the clausal syntactic structure. The semantic analyzer converts this to a &amp;quot;logical form&amp;quot; specifying states and actions with reference to specific components of the equipment. The discourse component establishes temporal and causal links between these states and actions.</Paragraph>
    <Paragraph position="1"> Initial implementations have been completed of the syntactic and semantic components, so that we are able to generate semantic representations of individual sentences. The discourse component is still under development, and so will not be discussed further here.</Paragraph>
    <Paragraph position="2"> The syntactic analyzer uses an augmented-context-free grammar and an. active chart parser. The grammar is generally based on linguistic string theory and the Linguistic String Project English Grammar \[Sager 1981\] and includes extensions to handle the various sentence fragment forms found in these messages \[Marsh 1983\]; it is written in a modified form of the Restriction Language used by the NYU Linguistic String Parser \[Sager 1975\]. Syntactic regularization maps the various forms of clauses (active, passive, relative, reduced relative, fragmentary) into a canonical form (verb operandi, operand2...) The regularization is performed by a set of interpretation rules which are associated with the individual productions and which build the regularized syntactic structure compositionally. 2 z The parser and syntactic regularization procedures were developed by Jean Mark Gawron. The regtllarization procedures were modeled after those developed for a GPSG parser \[Gawron 1982\], although the generated structures are quite different.</Paragraph>
    <Paragraph position="3">  The semantic analysis component consists of two parts: clause semantics and noun phrase semantics. The clause semantics maps a clause (a verb plus operands which include syntactic case labels) into a predicate auith arguments representing a state or action. Each verb and operand belongs to one or more semantic classes. Clause semantics relys on a set of pattern-action rules to perform the translation, with one pattern for each valid combination of verb and operand classes. Noun phrase semantics maps a noun phrase into the identifier of the equipment component specified by that phrase. Noun phrase semantics depends heavily on the equipment model, and so will be discussed further in a later section.</Paragraph>
    <Paragraph position="4"> (The division between the two parts of semantic analysis is not quite so neat as the foregoing would suggest. Some noun phrases are nominalizations representing states or actions; these are processed by clause semantics. In many noun phrases, some modifiers identify the object and the remainder describe its state. For example, in &amp;quot;broken hub ring gear&amp;quot;, hub and ring identify the gear, broken describes its state. We return to this problem in our description of noun phrase semantics below.) Our long-term objective is to dynamically schedule among the three analysis components (syntax, semantics, and discourse), as is done in some blackboard models. For program development, however, we have found it better to use a sequential organization (first syntax, then semantics, then discourse). In order to have syntactic choices influenced by semantics and discourse, and semantic choices influenced by discourse, each component may generate multiple analyses, some of which are rejected by later stages. Sometime these multiple analyses are transmitted explicitly, as a list of alternatives. More often, however, they are transmitted using a representation neutral with respect to particular features. The output of syntactic analysis is neutral with respect to quantifier scope. It is also neutral with respect to the distribution of modifiers in conjoined noun phrases (for example, in &amp;quot;filter change and adjustment of pressure regulator,&amp;quot; whether filter modifies adjustment and of pressure regulator modifies change). Furthermore, it does not assign structure to prenominal adjectives and nouns (so for example, in the phrase &amp;quot;low lube oil pressure alarm&amp;quot; it does not decide whether low modifies lube, oil, pressure, or alarm).</Paragraph>
    <Paragraph position="5"> This system development has been conducted in close Cooperation with a group at the System Development Corp., Paoli, PA. Their system, PUNDIT \[Palmer 1986\], is written in PROLOG but has many points of commonality with PROTEUS in terms of overall structure, grammar, and semantic representation. They are involved in future development of several areas, including semantic representation, time analysis, and anaphora resolution, for both the</Paragraph>
  </Section>
  <Section position="6" start_page="75" end_page="77" type="metho">
    <SectionTitle>
PUNDIT and PROTEUS systems.
4. The equipment model
</SectionTitle>
    <Paragraph position="0"> The equipment model currently serves three functions within our system: object identification. The noun phrases in the message are matched against the model (by a procedure outlined in the next section) in order to identify the objects referenced in the message. This is important both for syntactic disambiguation and as a prelude to applying domain-specific inferences.</Paragraph>
    <Paragraph position="1"> identification of intersentential relations. The identification of these relations (temporal, causal, and others) is important both for disambiguation (of adjuncts and anaphoric references, in particular) and for establishing the meaning of the message as a whole. Much of the information needed for this process - information on the structure of the equipment and the function of its components - is recorded in the equipment model.</Paragraph>
    <Paragraph position="2"> display of equipment structure and status. In order to provide some feedback to indicate whether the text was correctly understood, our system displays a structural diagram of the equipment at several levels of detail. Objects mentioned in the text, and changes in  equipment status described in the message, can be shown on the display. The information for generating these displays (positions, shapes, etc.) is stored with the equipment model. The messages refer to relatively low-level components, such as individual gears within the air compressor. We therefore had to constuct a relatively detailed model of the equipment involved. Our model has been developed through a study of the Navy manuals for this equipment.</Paragraph>
    <Paragraph position="3"> The model is basically organized as two hierarchies: a type-instance hierarchy and a part-whole hierarchy. The leaves of the part-whole hierarchy are called basic parts; the internal nodes (composite objects) are called systems. We record for each system the primary medium which it provides, conveys, or transforms; in our starting air system, the three media are compressed air, lubricating oil, and mechanical rotation. We have organized our part-whole hierarchy in part along functional lines (rather than purely on physical proximity), grouping together parts which are connected together and operate on the same medium.</Paragraph>
    <Paragraph position="4"> Since some parts are identified by their physical location, we provide a location field in both basic part and system nodes. Both types of nodes also have a function field, which indicates the effect of this part on the media or other parts. Nodes of specific types may have additional fields; for example, some mechanical components have a speed field.</Paragraph>
    <Paragraph position="5"> All of the fields just mentioned record permanent characteristics of the parts. In addition, each node has an operational-status field, which holds information about a part which is reported in a message.</Paragraph>
    <Paragraph position="6"> The model contains a lot of information about equipment structure which is specific to a particular piece of equipment. Some information, however, is more general: for example, that gears have teeth,or that impellors have blades. It would be most uneconomic to have a separate instance of tooth for each gear in the model. Instead we create an instance of the teeth for a specific gear when it is referenced in the text. Such very-low-level objects, which are instantiated dynamically as needed, are called components.</Paragraph>
    <Paragraph position="7"> The equipment model has been implemented using flavors on a Symbolics LISP Machine. Types of objects are represented by flavors; instances of objects are represented by instances of flavors. The part-whole hierarchy and other fields are stored in instance variables. The structure display is performed by procedures associated with the flavors. The equipment model, and its use in the system, are described in more detail in \[Ksiezyk 1986\]. 5. Noun phrase analysis The syntactic analysis component analyzes the clause structure and delimits the noun phrases, but does not assign any structure to the pre-nominal modifiers. The noun phrase analyzer within the semantic component therefore has a dual role: to determine the structure of the pre-nominal modifiers and to identify the instance in the equipment model named by the noun phrase (or the set of instances, if this phrase could be applied to any of several parts). (Although there are a limited number of instances, it is not possible to record a single name for each part and then interpret noun phrases by simply looking the name up in a table. A single part can be named in many different ways -- depending in part on prior context -- so a full-fledged interpretation procedure is required. ) The noun phrase is analyzed bottom-up using a set of reduction rules. Each reduction rule combines the head of a phrase with some of its modifiers to form a larger constituent. By reference to the model, each rule also determines the set of instances which can be named by the constituent; if the set is empty, the application of the rule is rejected. Reductions are performed repeatedly until the entire phrase is reduced to a single constituent. If no such reduction is possible, the syntactic analysis is rejected; in this way noun phrase semantics can weed out some incorrect syntactic analyses.</Paragraph>
    <Paragraph position="8">  The applicable reductions are determined by the dictionary entries for the words in the noun phrase. Each word is assigned two properties, its model class and its semantic class. The model class indicates how the word can be related to some entity in the domain model. One value of model class is instance, specifying that the word names a set of instances in the model; this set is also included in the dictionary entry. Examples are &amp;quot;pump&amp;quot;, &amp;quot;shaft&amp;quot;, &amp;quot;gear&amp;quot;, etc. Larger constituents built while analyzing the noun phrase are also considered to be of type instance. One reduction rule allows us to combine two instances: instance ~- instance- instance for example, &amp;quot;LO&amp;quot; + &amp;quot;PUMP&amp;quot; - &amp;quot;LO PUMP&amp;quot;, &amp;quot;SAC&amp;quot; + (&amp;quot;LO PUMP&amp;quot;) - &amp;quot;SAC LO PUMP&amp;quot;. The set of model instances for the result consists of those instances of the second constituent which can be linked through some path in the model to some instance of the first constituent. The types of links traversed in the search are a function of the semantic class of-the first constituent; for example, &amp;quot;SAC&amp;quot; has the semantic class machinery, so we search the part/whole links, the location links, and the from/to links (which tie together components of the same system).</Paragraph>
    <Paragraph position="9"> There are several other model classes and corresponding reduction rules. The class slot-filler is used for words which are values of features of instances, but are not themselves instances (for example, &amp;quot;LUBE&amp;quot; in the phrase &amp;quot;LUBE OIL&amp;quot;). The class slot-name is used for words which correspond to feature names, such as &amp;quot;SPEED&amp;quot; in &amp;quot;HIGH SPEED ASSEMBLY&amp;quot;. The class component is used for parts which (as explained in the previous section) are not instantiated in the permanent equipment model but can be instantiated dynamically as needed.</Paragraph>
    <Paragraph position="10"> Modifiers describing the state of a part, such as &amp;quot;cracked&amp;quot; or &amp;quot;sheared&amp;quot;, are handled differently. If noun phrase semantics gets the input &amp;quot;sheared ring gear&amp;quot; it will look for an instance of ring gear with the operational-state &amp;quot;sheared&amp;quot;. Such an instance would be present if a previous sentence had mentioned that a gear was sheared. If such an instance is found, it is identified as the correct referent; noun phrase semantics has in effect done anaphora resolution. If no instance is found, noun phrase semantics returns the instances of &amp;quot;ring gear&amp;quot; and the left-over modifier &amp;quot;sheared&amp;quot;. Clause semantics (which invokes noun phrase semantics) then treats this like a clause &amp;quot;ring gear was sheared&amp;quot;; later in the processing of this sentence, this will cause &amp;quot;sheared&amp;quot; to be assigned as the operational-state of ring gear. A related technique can be used to handle some of the ambiguities in cojoined noun phrases. For example, in the sentence '.'INVESTIGATION REVEALED STRIPPED LO PUMP DRIVE AND HUB RING GEAR&amp;quot;, syntax alone cannot determine which of the modifiers &amp;quot;STRIPPED&amp;quot;, &amp;quot;LO&amp;quot;, &amp;quot;PUMP&amp;quot;, or &amp;quot;DRIVE&amp;quot; also modify &amp;quot;HUB RING GEAR&amp;quot;. So syntax marks these as possibly applicable to &amp;quot;HUB RING GEAR&amp;quot; and passes the phrase to semantics. If semantics finds that some of these modifiers cannot be integrated into the noun phrase, they will be ignored, thus implicitly resolving the syntactic ambiguity.</Paragraph>
  </Section>
class="xml-element"></Paper>