<?xml version="1.0" standalone="yes"?>
<Paper uid="W90-0104">
  <Title>A New Model for Lexical Choice for Open-Class Words</Title>
  <Section position="3" start_page="0" end_page="26" type="metho">
    <SectionTitle>
2. Lexical Choice as Classification
</SectionTitle>
    <Paragraph position="0"> The two major approaches (to date) for lexical choice have been discrimination nets and structure mapping systems. Both of these approaches can be regarded as classification/matching architectures, where a classifier is given an object or event, and is asked to find an appropriate lexical unit that fits that object or event.</Paragraph>
    <Paragraph position="1"> Discrimination nets (e.g., Goldman 1975; Pustejovsky and Nirenburg 1987) are basically decision trees. They are typically used as high-speed 'compiled' classifiers that select the most specific lexical unit that subsumes  the target object or event. For instance, looking at some of Goldman's examples, the event !ngest(John,Milk027), which can be represented in KL-ONE as (Ingest with VR's actor:John and theme:Milk027), has as its most specific subsuming lexical unit (Ingest with VR theme:Liquid), and thus is lexically realized as &amp;quot;drink&amp;quot;. Similarly, the action Ingest(BearO36,Fish802), which'can be represented in KL-ONE as (Ingest with VR's actor:Bear036 and theme:Fish802), has (Ingest with VR's agent:Nonhuman-animal and theme:Solid) as its most specific subsumer in a taxonomy of German lexical units, and thus is realized, in German, as &amp;quot;fressen&amp;quot;.</Paragraph>
    <Paragraph position="2"> Structure-mapping systems (e.g., Jacobs 1987; Iordanskaja et al. 1988; note that different terminology is used in different papers) take as input a semantic structure that needs to be communicated to the user, search for pieces of the input structure that are equivalent to lexical units, and then replace the matched structure by the corresponding lexical unit. The matching and subsdtution process continues until the semantic structure has been completely reformulated in terms of lexical units. For example, the structure (Human (:sex male) (:age-status adult) (:wealth high)) might be mapped into the structure (&amp;quot;man&amp;quot; (:attribute &amp;quot;rich&amp;quot;)), and hence lexitally realiTcd as &amp;quot;rich man&amp;quot;. In KL-ONE terms, the matching process can be considered to be a search for a class definition that uses only classes and role VR's that can be realized as lexical units; e.g., the above example essentially redefines the class (Human with role VR's Sex:Male, Age-status:Adult, Wealth:High) as the equivalent class (&amp;quot;man&amp;quot; with VR &amp;quot;rich&amp;quot;), where the lexical unit &amp;quot;man&amp;quot; represents the class (Human with role VR's Sex:Male, Age-status:Adult), and the lexical unit &amp;quot;rich&amp;quot; is equivalent to the role VR Wealth:High. Recently, the machine translation group at CMU has proposed an alternative lexical choice system that is based on a variant of nearest neighbor classification (Center for Machine Translation 1989; Nirenburg et al.</Paragraph>
    <Paragraph position="3"> 1987). In the CMU system, both objects and lexical units are treated as points or regions in a feature space, and the classifier works by choosing the lexical unit that is closest to the target object, using a fairly complex distance (matching) metric (collocation constraints are also taken into consideration). For example, the object (Human with VR's Sex:Male and Age:13) would be judged closest to the lexical unit (Human with VR's Sex:Male and Age:range(2,15)), and thus would be realized as &amp;quot;boy'.</Paragraph>
    <Paragraph position="4"> All of the above classification-based lexical-ehoice architectures 2 suffer from two basic flaws: * they do not allow a clean separation to be made between what the system knows, and what it wishes to communicate; * they do not provide a clean mechanism for allowing the lexical choice process to be biased towards preferred lexical units.</Paragraph>
    <Paragraph position="5"> These failures may lead classification-based systems to choose inappropriate lexical units that carry unwanted conversational implicatures (Grice 1975), and therefore mislead the user.</Paragraph>
    <Paragraph position="6"> 2.1. One Input vs Two Inputs Classification-based systems take as their input a single set of attributes about the object/event being lexicalized, and use this set of attributes to select a matching classification. However, lexical choice systems should look at two input sets of attributes: the set of object/event attributes that are relevant and need to be conveyed to the user, and the set of attributes that constitute the system's total knowledge of the object/event being lexicalized.</Paragraph>
    <Paragraph position="7"> A lexieal choice system that looks only at the system's domain knowledge about the object/event, and ignores the set of relevant attributes, may choose inappropriate lexical items that carry unwanted relevance conversational implicatures. In particular, a system that simply selects the most specific lexical unit that subsumes the object/event (as many discrimination net sys- null tems do) may mislead the user by choosing lexical units that are too specific. For example, consider the following exchange: 1) A: &amp;quot;Is Terry a woman?&amp;quot; 2a) B: &amp;quot;No, Terry is a man&amp;quot; 2b) B: &amp;quot;No, Terry is a bachelor&amp;quot;  B's communicative goal is simply to inform A that Terry has the attributes {Human, Age-status:Adult, Sex:Male}, so utterance (2a) is an appropriate response. A lexical choice system that simply selected the most specific lexical unit that subsumed Terry would generate utterance (2b), however. Utterance (2b) is inappropriate, and would probably lead A to infer the (incorrect) conversational implicature that B thought that Terry's marital status was relevant to the conversation.</Paragraph>
    <Paragraph position="8"> A lexical choice system that looks only at the attributes being communicated, and ignores the system's  general domain knowledge about the object/event, may also make inappropriate lexical choices that lead to unwanted conversational implicatures. For example, suppose A wished to communicate to B that XNET was a Network with the attributes {Data-rate:lOMbit/sec, Circuit-type:Packet-switched}. Consider three possible lexicalizations:</Paragraph>
    <Paragraph position="10"> ing the user has some domain knowledge about Ethernets). Utterance (3a), however, would be generated by a system that simply chose the most specific lexieal unit that subsumed {Network, Data-rate:lOMbit/sec, Circuit-type:Packet-switched}. 3 This utterance fails to fulfill the communicative goal of informing the reader that the network has the attributes {Datarate:lOMbitlsec, Circuit-type:Packet-switched}, and is therefore unacceptable. Utterance (3b) would be generated by a structure-mapping system that chose a lexical unit according to the above strategy, and then added expficit modifiers to communicate attributes that were not impfied by the lexical class. 4 Tlds utterance successfully communicates the relevant information, but it also implicates, to the knowledgeable hearer, that XNET is not an Ethernet m because if it was, the knowledgeable hearer would reason, then the speaker would have used utterance (3e).</Paragraph>
    <Section position="1" start_page="25" end_page="26" type="sub_section">
      <SectionTitle>
2.2. Preferred Lexical Units
</SectionTitle>
      <Paragraph position="0"> Certain lexical units, in particular those that represent basic-level classes (Rosch 1978), are preferred and should be chosen whenever possible. Cruse (1977) and 3 Another possibility is choosing the most general lexical unit that is subsumed by the attributes being communicated. However, this cmnot be done by a ~/stma that ignores the object and only 1oo1~ at the attributes being communicated, because such a system would not know which le~ical units accurately described the object. For example, if there were two classes Ethernet and Applenet that had the attributes (Network, Data-rate:lOMbitlsec, Circuit.type:Packet-switched}, the system could only decide whether to generate &amp;quot;Ethemet&amp;quot; or &amp;quot;Applenet&amp;quot; by detexmining which of these classes subsumed the object being described (e.g., &amp;quot;Etbemet&amp;quot; should be used to describe XNET). See also exampie 5, where the most appropriate leexical unit that informs the hearer that F/do has the attributes {An/ma/, Breathes:Air} is &amp;quot;dog', not &amp;quot;mmnmal&amp;quot; or &amp;quot;animal'.</Paragraph>
      <Paragraph position="1"> ,t In this example, the 'lexical choice' system is assumed to capable of forming a complete NP. In general, it is often difficult to separate the task of selecting a single word from the task of forming a complete phrase.</Paragraph>
      <Paragraph position="2"> others have suggested that the failure to use a basic-level class in an utterance will conversationally implicate that the basic-level class could not have been used.</Paragraph>
      <Paragraph position="3"> For example, consider the following utterances:  4) A: &amp;quot;I want to flood room 16 with carbon dioxide&amp;quot; 5a) B: &amp;quot;Wail there is an animal in the room&amp;quot; 5b) B: &amp;quot;Wail there is a dog in the room&amp;quot; 5c) B: &amp;quot;Wait, there is a Pekingese in the room&amp;quot;  Assume the object in question is Fido, and A's communicative goal is simply to inform B that Fido has the attributes {An/ma/, Breathes:Air}, and hence would be adversely affected if the room was flooded with carbon dioxide. Utterances (5a), (5b), and (5c) all fulfill this communicative goal (assuming that Breathes:Air is a default attribute of An/ma/), but utterance (5b) is preferred because Dog is a basic-level class. Utterance (5a) is odd because the use of the superordinate class Animal impficates, according to Cruse's hypothesis, that the animal in question is not a Dog, Cat, or other commonly known type of animal (or at least the speaker does not know that the animal is a member of one of these species); utterance (5c) is odd because the use of the subordinate class Pekingese implicates that it is somehow relevant that the animal is a Pekingese and not some other kind of dog. If both of these impficatures are incorrect, the speaker should choose the lexical unit Dog if he wishes to avoid misleading the hearer.</Paragraph>
      <Paragraph position="4"> It should be pointed out that the strategy of simply always picking a basic-level class that subsumes the object/event will not work, because it ignores the system's communicative goals. For instance, a system that followed the basic-level strategy would, in the situation of example 3, generate utterance (3a) or (3b).</Paragraph>
      <Paragraph position="5"> Both of these are inappropriate and implicate, to the  knowledgeable user, that utt~ce (3c) could not have been used, i.e., that XNET is not an Ethernet.</Paragraph>
      <Paragraph position="6"> 3. Lexical Choice as Constraint Satisfaction  The above problems can be avoided by regarding lexical choice as a constraint-satisfaction task instead of a classification task. More precisely, the task of choosing an appropriate open-class lexical unit should be formalized as follows: Input.</Paragraph>
      <Paragraph position="7">  * Entity: a taxonomy class that represents the system's knowledge of the object or event being lexicalized.</Paragraph>
      <Paragraph position="8"> * To-Communicate: a set of predicates (attributes) that represent the relevant information about the object that needs to be communicated to the user.</Paragraph>
      <Paragraph position="9">  Output: A lexical unit Lex that is a member of the knowledge-base taxonomy, and that satisfies the follow- null ing constraints: * Accurate: Lex must be a truthful description of Entity. Formally, Lex must subsume Entity.</Paragraph>
      <Paragraph position="10"> * Valid: The use of Lex in an utterance must inform the  user that the predicates in To-Communicate hold for Entity. Formally, every predicate in To-Communicate must either be inferrable from the definition of Lex (e.g., subsume Lex), or be a default attribute that is associated with Lex.</Paragraph>
      <Paragraph position="11"> * Preferred: Lex must be a maximal element of the set of accurate and valid lexical units under certain lexical preference rules (Section 4).</Paragraph>
      <Paragraph position="12"> In other words, the lexical choice system is given two inputs, which represent the system's knowledge of the object or event, and the relevant information about that object or event that needs to be communicated to the user; and is expected to produce as its output a maximal lexical unit (under the lexical preference rules) that is truthful and conveys the relevant information.</Paragraph>
      <Paragraph position="13"> The constraint-based system makes appropriate lexical choices in each of the previous examples: * Entity = Terry, To-Communicate = {Human, Sex:Male} (example 2). Both Man and Bachelor are accurate and valid lexical units. Man is chosen, because it is basic-level and therefore preferred.</Paragraph>
      <Paragraph position="15"> (example 3). Ethernet is chosen, because it is the only accurate and valid lexical unit.</Paragraph>
      <Paragraph position="16"> * Entity = Fido, To-Communicate = {Animal, Breathes:Air} (example 5). Accurate and valid lexical units include Animal, Mammal, Dog, and Pekingese. Dog is chosen, because it is basic-level.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="26" end_page="28" type="metho">
    <SectionTitle>
4. Preferences Among Lexical Classes
</SectionTitle>
    <Paragraph position="0"> If several lexical units are accurate and valid, a set of lexical preferences rules is used to select the lexical unit the system will utter. The preference for basic-level classes was previously mentioned (Section 2.2), but it is complicated by entry-level effects (Section 4.1), Additional lexical preferences include the length\[subset preference (Section 4.2). Combined, the lexical preference rules impose a lexical preference hierarchy on the lexical units in the knowledge base. Figure 2 shows part of the lexical preference hierarchy that is associated with the knowledge base of Figure 1.</Paragraph>
    <Paragraph position="1"> 4.1. Basic-Level vs Entry-Level Preferences Hirschberg (1985) has suggested that it may be better to use Jolicoeur et al.'s (1984) notion of entry level classes instead of Rosch's basic level classes. The difference is that under the entry-level hypothesis, which category is unmarked (i.e., which category may be used without generating a conversational implicature) may depend on how atypical the object is. For example, consider:.</Paragraph>
    <Paragraph position="2"> (the speaker points to a robin)  7a) ``Look at the bird&amp;quot; 7b) &amp;quot;Look at the robin&amp;quot; (the speaker points to an ostrich) 8a) &amp;quot;Look at the bird&amp;quot; 8b) &amp;quot;Look at the ostrich&amp;quot;  Under the basic-level hypothesis, a category is either basic-level or it is not, and ff it is basic-level, then it is always the unmarked way of referring to any object that belongs to it. Therefore, under this hypothesis utterances (7a) and (8a) are both unmarked and carry no conversa6onal implicatures, since Bird is a basic-level category for most urban Americans. Under the entry-level hypothesis, in contrast, while a basic-level category is the unmarked way of referring to 'normal' members of the category, it may not be the unmarked way of referring to atypical members. Instead, a more specialized category may be the unmarked way of referring to atypical members. Thus, under the entry-level hypothesis, even ff utterance (7a) was the unmarked way of referring to robins (which are typical birds), utterance (8b) could still be the unmarked way of referring to ostriches (which are atypical birds).</Paragraph>
    <Paragraph position="3"> The lexical-choice system can allow for entry-level effects ff it allows any lexical unit to be marked as basic-level in the taxonomy, but then only considers the lowest such marked class to be a true basic-level (and hence lexically-preferred) class for an object. More precisely, ff an object has two subsumers A and B that are both marked as basic-level classes, and A subsumes B, then the system should only treat B as a lexically-preferred class for the object. For example, in Figure 1 Bird and Ostrich are both marked as basic-level. Therefore, the lexical-choiee system should treat Bird (but not Sparrow) as a lexically-preferred class for Tweety (a Sparrow), and Ostrich (but not Bird) as a lexically-preferred class for Big-Bird (an Ostrich).</Paragraph>
    <Section position="1" start_page="26" end_page="28" type="sub_section">
      <SectionTitle>
4.2. Length/Subset Preferences
</SectionTitle>
      <Paragraph position="0"> A lexical unit A is almost always preferred over a lexical unit B if A's surface form uses a subset of the words used by B's surface form (this can be considered  to be a consequence of Grice's maxim of quantity (Grice 1975)). Consider, for example,  9a) ``Don't go swimming; there is a shark in the water&amp;quot; 9b) &amp;quot;Don't go swimming; there is a tiger shark in the water&amp;quot;  According to the subset lexical, preference rule, lexical unit Shark is preferred over lexical unit Tiger-shark. Therefore, the use of utterance (9b) carries the conversational implicature that utterance (9a) could not be used, i.e., that it was relevant that the animal was a Tiger-shark and not some other kind of Shark. A hearer who heard utterance (9b) might infer, for example, that the speaker thought that tiger sharks were unusually dangerous kinds of sharks. If no such implicature was intended by the speaker, then he should use utterance (9a), not utterance (9b).</Paragraph>
      <Paragraph position="1"> A stronger version of this preference rule would be to prefer lexical unit A to lexical unit B if A's surface form used fewer open-class words than B's surface form. This would, for example, correctly predict that Dog is preferred over Great-Dane, and that Flower is preferred over Rocky-Mountain-iris. This preference is usually accurate, but it does fail in some cases. For example, it is questionable whether Porsche is preferred over Sports-car, and doubtful whether Mammal is preferred over Great-Dane.</Paragraph>
      <Paragraph position="2"> There are cases where the basic-level preference conflicts with (and takes precedence over) both the sub-set and the length preferences. Such conflicts are probably rare, because psychological and linguistic findings suggest that basic-level classes are almost always lexically realized with single words (Rosch 1978; Berlin et al. 1973). However, there are a small number of basic-level classes that have multi-word reali7ations, and this can lead to conflicts of the above type. Consider, for example,  10a) &amp;quot;Joe has a mach/ne&amp;quot; 10b) ``Joe has an appliance&amp;quot; 10c) &amp;quot;Joe has a washing machine&amp;quot;  Washing-machine is probably basic-level for most Americans. Therefore, utterance (10c) is preferred over utterances (10a) and (10b), despite the fact that the length preference suggests that utterances (10a) and (10b) should be preferred over utterance (10c), and the subset preference suggests that utterance (10a) should be preferred over utterance (10c).</Paragraph>
    </Section>
    <Section position="2" start_page="28" end_page="28" type="sub_section">
      <SectionTitle>
4.3. Other Lexical Preference
</SectionTitle>
      <Paragraph position="0"> There are lexical preferences that are not captured by either the basic-level preference or the subset/length preference. For example, suppose the speaker wished to refer to two animals, a horse and a cow. Consider the difference between</Paragraph>
      <Paragraph position="2"> None of the above are basic-level classes (Horse and Cow are basic-level for most urban Americans). Therefore, neither the basic-level nor the length/subset rules indicate any preferences among the above. However, it seems clear that utterance (lla) is much preferable to utterance (lib), and that utterance (lib) is probably preferable to utterance (llc). In addition, the use of utterances (llb) or (llc) seems to implicate that utterunee (lla) could not have been used.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="28" end_page="29" type="metho">
    <SectionTitle>
5. Default Attributes
</SectionTitle>
    <Paragraph position="0"> One final point is that the representation of the semantics of lexical units must include default attributes as well as definitional information. These defaults may represent domain knowledge (e.g., birds typically fly) or useful conventions that have evolved in a particular environment (e.g., most computers at Harvard's Aiken Computation Lab run the UNIX operating system).</Paragraph>
    <Paragraph position="1"> Systems that ignore default attributes may make inappropriate lexical choices, and therefore generate utteranoes that carry unwanted conversational implicatures.</Paragraph>
    <Paragraph position="2"> For example, ff To-Communicate was {Bird, Canfly:True}, and Entity was Tweety, consider the difference between 12a) &amp;quot;Tweety is a b/rd&amp;quot; 12b) ``Tweety is a bird that can fly&amp;quot; If the generation system ignored default attributes, it would have to generate something like utterance (12b).</Paragraph>
    <Paragraph position="3"> Utterance (12b) sounds odd, however, and a person who heard it might infer unwanted and unintended conversational implicatm'es, e.g., that some other bird under discussion was not able to fly. Utterance (12a) is much better, but it can only be generated by a generation system that takes into consideration the fact that Canfly:True is a default attribute of Bird.</Paragraph>
    <Paragraph position="4"> For another example, sutvose an NLG system wished to inform a user that a particular computer was a VAX that ran the UNIX operating system and the Latex text processor (i.e., To-Communicate = {VAX, Operating- null system:UNIX, Available-software:Latex}). Consider two possible utterances: 13a) &amp;quot;Hucl is a VAX that runs Latex&amp;quot; 13b) &amp;quot;Hucl is a UNIX VAX that runs Latex&amp;quot; Utterance (13a) is acceptable, and indeed expected, if the user thinks that Operating-system:UNIX is a default attribute of VAX's in the current environment (e.g., at the Aiken Computation Lab). In a different environment, where users by default associate Operatingsystem:VMS with VAX's, utterance (13a) would be misleading and unacceptable, and utterance (13b) should be generate&amp;</Paragraph>
  </Section>
class="xml-element"></Paper>