<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-0612">
  <Title>Exemplar-Based Sense Modulation</Title>
  <Section position="3" start_page="0" end_page="89" type="intro">
    <SectionTitle>
2. Motivations
</SectionTitle>
    <Paragraph position="0"> Elsewhere (Rais-Ghasem, 1998), the first author has reviewed existing computational lexicon models and showed that, despite their differences, they all subscribe to the same meaning theory, namely sense enumeration (Seidenberg et al., 1982; Simpson and Burgess, 1985). Fundamental to this theory are two assumptions: 1) the possibility of listing or enumerating all possible meanings for each word, and 2) the existence of a selection process in which one of these meanings is selected for a given word.</Paragraph>
    <Paragraph position="1"> One of the main disadvantages of such enumerative lexicons is their inability to account for a phenomenon generally known as semantic flexibility (see Barclay et al., 1974; Barsalou, 1993; Greenspan, 1986). In short, semantic flexibility concerns changes that context causes in representation of concepts in memory. Many psycholinguistic findings have shown that context seems to highlight or obscure certain properties of a single concept as it appears in different contexts. For example, in an early experiment, Barclay et al. (1974) demonstrated how the interpretations of familiar, unambiguous words vary with context. For instance, they argued that the choice of attributes for piano is affected by the verb selection in The man (l~ed) (tuned) (smashed) (sat on) (photographed) the piano. They then provided evidence that the prior acquisition of a sentence like The man lifted the piano (vs. The man tuned the piano) influences the effectiveness of cues like &amp;quot;something heavy&amp;quot; (vs. &amp;quot;'something with a nice sound&amp;quot;) in recall. They concluded that context can affect the encoding of concepts in memory. Similarly, Barsalou (1982) reported that subjects verify contextually relevant properties significantly faster than contextually irrelevant properties Witney et al. (1985) also report similar results. Their work is particularly interesting since they distinguished between two levels of property activation: 1) the functional level (useful in activities such as sentence comprehension and classification), and 2) the semantic access level (corresponding to the information that is actually accessed upon  seeing a word). They used intervals of 0, 300 and 600 ms.-and found that all properties of a word were initially activated (accessed), regardless of the context. However, the contextually irrelevant ('low-dominant' in their terminology) properties would die off rapidly, and hence play a negligible role in the overall sentence comprehension.</Paragraph>
    <Paragraph position="2"> Greenspan (1986) studied the effect of sentential context on concrete nouns. He examined the presence of central and peripheral properties of a noun in different contexts. For a given a noun, Greenspan presented his subjects with a pair of sentences where in each sentence a different type of properties was emphasized.</Paragraph>
    <Paragraph position="3"> For example, consider the noun basket. Being a container is a central property for basket whereas being made of straw is a peripheral one. Each of the following sentences focuses on one of these properties.</Paragraph>
    <Paragraph position="4"> The butler placed the letter in the basket.</Paragraph>
    <Paragraph position="5"> (Container) Sally took several days to weave the basket.</Paragraph>
    <Paragraph position="6"> (Straw) Later he examined subjects' recall in various memory tasks and found that the central properties were activated regardless of the context, but peripheral properties were activated only if they were emphasized by the sentence.</Paragraph>
    <Paragraph position="7"> He further showed that the emphasized central properties were more activated than unemphasized central properties. He concluded that the interpretation of a concrete noun is a function of both the sentential context and the noun's central properties (Ibid.).</Paragraph>
    <Paragraph position="8"> Anderson et al. (1976) demonstrated that general terms are contextually instantiated to more specific terms. For example, they used fish in contexts like the following sentences and hypothesized that it was likely to be instantiated respectively to goldfish, trout and shark.</Paragraph>
    <Paragraph position="9"> The grocer stared at the fish in the bowl.</Paragraph>
    <Paragraph position="10"> The grocer stared at the fish in the stream.</Paragraph>
    <Paragraph position="11"> The fish attacked the swimmer.</Paragraph>
    <Paragraph position="12"> They examined their hypothesis in a number of experiments and found that an expected instantiated term was a better cue for the recall of a sentence then the general term itself, even though the general term had appeared in the sentence and the particular term had not.</Paragraph>
    <Paragraph position="13"> To the best of our knowledge, none of the existing computational lexicon models have adequately dealt with semantic flexibility.</Paragraph>
    <Paragraph position="14"> Having subscribed to the sense enumeration theory, the effects of context in these models are limited to selecting of one sense. Any changes in the characteristics of a selected sense either has to come in the form of a new sense or is ignored.</Paragraph>
    <Paragraph position="15"> This requires the ability to foresee any context that a word can appear in and define a separate sense for it. Obviously, this is impractical, if not impossible (see Corriveau, 1995, chapter 2).</Paragraph>
    <Paragraph position="16"> Thus, any lexicon model must support a combination of sense generation and sense selection.</Paragraph>
    <Paragraph position="17"> Of the various computational lexicon models, this issue is specially problematic for symbolic lexicons (e.g., Dyer, 1982; Wilensky, 1978) and lexicons based on semantic networks (Lange &amp; Dyer, 1989; Waltz &amp; Pollack, 1985). This is due to the fact that word senses in these models are coded as discrete entries.</Paragraph>
    <Paragraph position="18"> Owing to their adopted continuous representation, connectionist models, however, are potentially capable of dealing with contextual effects at a finer level. In fact McClelland and Kawamoto (1986) reported an unintended yet interesting result. They had presented their model with The ball broke the vase. Although throughout the training ball was always associated with the microfeature soft, in the output it was associated with the microfeature hard. They attributed this result to the fact that breakers in their experiment were all hard and the model had shaded the meaning of ball accordingly.</Paragraph>
    <Paragraph position="19"> However, the connectionist approach also has some disadvantages. First of all, connectionist lexicons presuppose a set of universal and fixed microfeatures and demand that every sense be characterized in terms of such microfeatures in advance. This constitutes a serious problem for any real world application.</Paragraph>
    <Paragraph position="20"> But what is even more important is the difficulty to separate patterns of contextual effects from the representation of a word sense. For instance, consider breakers in McClelland and Kawamoto (1986). It is impossible to examine this category of objects by itself. Thus we cannot determine 1) what qualifies an object to be a breaker or 2) what else can be a breaker? We believe separating patterns of contextual effects from internal representations of context is important. Firstly, such patterns can be thought of as ad hoc categories, categories built by people to achieve goals (Barsalou, 1983). For instance, the breakers category can be instrumental in achieving the goal of &amp;quot;breaking a window&amp;quot;. Secondly, from a learning point of view, such patterns can be very useful. Rais-Ghasem (1998) has shown how a concept can evolve (i.e., acquires new properties) from such  patterns as it appears in various contexts. Also,  Rais-Ghasem (Ibid.) has employed such patterns to implement a metaphor understanding system that interprets metaphors as class inclusion assertions (see Gluksberg &amp; Keysar, 1990).</Paragraph>
    <Paragraph position="21"> 3. A Lexicon for Sense Modulation  In this section, we propose a lexical model that not only selects senses for input words, but also contextually modulates the selected senses.</Paragraph>
    <Paragraph position="22"> Examples used in this section are from a prototype implementation of this model for sentential contexts.</Paragraph>
    <Section position="1" start_page="86" end_page="87" type="sub_section">
      <SectionTitle>
3.1 Two-Tiered Word Senses
</SectionTitle>
      <Paragraph position="0"> Cruse (1986) specifies two ways in which context affects the semantic contribution of a word: sense selection and sense modulation.</Paragraph>
      <Paragraph position="1"> Sense selection happens in cases of lexical ambiguity where one sense is chosen among a number of contrastive senses (see Pustejovsky, 1995) associated with a word. In sense modulation, the semantic characteristics of a selected sense are modulated according to its surrounding context. Cruse describes two types of sense modulation: a) highlighting or backgrounding, and b) promotion or demotion of features. In the former, context underlines certain aspects (i.e., properties) of the concept selected for a given word while dimming others, as it happens for piano in the following examples: The musician played the piano.</Paragraph>
      <Paragraph position="2"> The musician broke the piano.</Paragraph>
      <Paragraph position="3"> The musician lifted the piano.</Paragraph>
      <Paragraph position="4"> Context can also promote, or demote, certain aspects of a word's associated concept. For instance, teacher, by definition, is gender neutral. However, in a context like the following, the feature gender is promoted for teacher.</Paragraph>
      <Paragraph position="5"> The teacher stroked his beard.</Paragraph>
      <Paragraph position="6"> Similarly, certain aspects can be demoted, as is solidity, a property usually associated with butter, in this context: Mary poured the butter into the jar.</Paragraph>
      <Paragraph position="7"> To account for both sense selection and sense modulation, we propose to structure word senses into two tiers: sense-concept tier and sense-view tier (see Figure 1). While the sense-concept tier specifies the core meaning of a word in a given context, the sense-view tier specifies how this meaning is to be viewed in that context. It is our contention that this simple structure is capable of accounting for various types of sense modulation.</Paragraph>
      <Paragraph position="8">  Here is an example of the word sense generated by the implemented prototype (see section 4) for piano in The musician moved the piano.</Paragraph>
      <Paragraph position="9">  As shown, this word sense consists of two parts: a sense-concept and a sense-view. Since concepts to a large extent are conventionalized, sense-concepts are specified only by means of the concepts they represent (concept PIANO in this case). Sense-views, however, are specified as a set of properties (e.g., WEIGHT) and their possible values (e.g., HEAVY). Each property is also accompanied with a number between 0 and 1 indicating the property's weight (or centrality) in a given sense-view. The implemented system relies on the Mikrokosmos ontology (Mahesh and Nirenburg, 1995) to specify properties for sense-concepts and sense-views j.</Paragraph>
      <Paragraph position="10"> Now, compare the above word sense with the word sense generated for piano in a different context such as The musician played the piano.  particular ontology or set of properties. The choice of Mikrokosmos is primarily justified because of its availability and relative comprehensiveness.</Paragraph>
      <Paragraph position="11">  Note that the two word senses share the same sense-concept. They differ, however, on their sense-views. Because of the sense-view, the first word sense portrays piano as an object of moving (by highlighting properties such as weight and artifact). Conversely, the second word sense underlines a different aspect of piano, namely the fact that it is a musical Instrument.</Paragraph>
      <Paragraph position="12"> As examples of property demotion/promotion, consider the word senses respectively generated for mttsician in the above two contexts. In the first context, we get:  are demoted in this example, since they are irrelevant to the role played by musician in this context (i.e.. being a mover). This is not the case in The musician played the piano.</Paragraph>
    </Section>
    <Section position="2" start_page="87" end_page="88" type="sub_section">
      <SectionTitle>
3.2 Exemplar-Based Sense-Views
</SectionTitle>
      <Paragraph position="0"> In this section, we outline a lexicon model in xvhich sense-views are developed gradually and incrementally. That is achieved by exploiting exemplars 2. Exemplars are previously processed contexts that exemplify the usage of word senses. In other words, every exemplar consists of a number of word senses, each of the latter formed by a pair of the form (senseconcept/sense-view). null We first define the alike relationship: Definition 1: Sense-concepts SC, and SC 2 are alike if they appear in similar contexts, and they share the same thematic role.</Paragraph>
      <Paragraph position="1"> For example, according to this relationship, piano, fridge and computer are alike in the following inputs: They used a crane to move the stove.</Paragraph>
      <Paragraph position="2"> 2 Similar to semantic relations in Cruse (1986). John carried the computer.</Paragraph>
      <Paragraph position="3"> Four men moved the piano.</Paragraph>
      <Paragraph position="4"> In other words, from the view point of &amp;quot;being moved&amp;quot;, these representationally different sense-concepts can be classified together.</Paragraph>
      <Paragraph position="5"> This parallels the two different roles that Franks (1995) proposed for concepts. He distinguished between the representational and classificatory functions of concepts. While the former is used to discem instances of one concept from others, the latter specifies how an instance of a concept should be classified.</Paragraph>
      <Paragraph position="6"> Franks (Ibid.) argues that, depending on context, fake gun could be classified along with a gun, a toy, a replica, and a model.</Paragraph>
      <Paragraph position="7"> Given that all alike sense-concepts share the same sense-view, we can define a sense-view as the intersection of the properties of the sense-concepts that are connected to it. Figure 2 illustrates this situation.</Paragraph>
      <Paragraph position="8">  A sense-view is initialized to the properties of the first sense-concept to join it. As more sense-concepts join, the sense-view evolves to better reflect what is common between all those senseconcepts. For example, the following is the content of the sense-view object of moving initiated with only one exemplar: The man moved the stove.</Paragraph>
      <Paragraph position="9">  This definition gradually becomes more representative of this sense-view as new alike  sense-concepts are included. After adding computer, as it appears in: The student moved the computer.</Paragraph>
      <Paragraph position="10"> we get:  continues until the sense-view reaches a certain level of stability (for more details, see Rais-Ghasem, 1998).</Paragraph>
    </Section>
    <Section position="3" start_page="88" end_page="89" type="sub_section">
      <SectionTitle>
3.3 From Words to Senses
</SectionTitle>
      <Paragraph position="0"> Having described the two-tiered word senses and our adopted exemplar-based approach to develop sense-views, we can now overview the models overall behavior.</Paragraph>
      <Paragraph position="1"> Input to the system is a context, presented as a number of input words and along with their syntactic categories and case markers (Delisle et al., 1993). For example, a sentence like The musician played the piano would be presented to the model as musician (noun, p-subj), play (verb, -), and piano (noun, p-obj). Case markers p-subj and p-obj respectively indicate &amp;quot;positional subject&amp;quot; and &amp;quot;positional object&amp;quot;. The goal is first, to select proper senses for input words, and second, to contextually modulate those senses. The model attempts to accomplish both tasks by looking for an 3 This process also meets, to some extent, another requirement, namely idiosyncrasy of word senses. For example, the property hem 3 ' is present in this sense-view only because of its strong presence in the sense-concepts added to this sense-view so far. That may not be the case for another reader and may not hold over time.</Paragraph>
      <Paragraph position="2">  between a sense-view and a concept is defined based on the degree of overlap in their properties.</Paragraph>
      <Paragraph position="3"> For instance, assume that the model already maintains the following exemplars: El. The musician played the guitar.</Paragraph>
      <Paragraph position="4"> E2. Mary played soccer.</Paragraph>
      <Paragraph position="5"> Also suppose the following input is presented to the model.</Paragraph>
      <Paragraph position="6"> John played the piano.</Paragraph>
      <Paragraph position="7"> To determine the adaptability of this input with E1 and E2, we should first find sense-views corresponding to the input words, and then measure their compatibility. Let us assume that we know that the input words John, play and piano respectively correspond to sense-views associated with musician, play and guitar in E1 and sense-views associated with Mary, play and soccer in E2.</Paragraph>
      <Paragraph position="8"> Let's begin with E2. It is possible to select a sense for the input word John which is compatible with sport-player (the sense-view associated with Mary). The same is true for input word play and the sense-view playingsport. However, finding a sense for piano which is compatible with the sense-view sport (soccer's associated sense-view) is not possible. Therefore, the input context and E2 are not adaptable.</Paragraph>
      <Paragraph position="9"> Conversely, E1 and the input context, are adaptable: John is compatible with music-player (musician's associated sense-view), play with playing-music, and piano with musicalInstrument. null Adaptability, in fact, allows the model to decide if a set of word senses can be selected or generated for the input words. Put roughly, given some input words, the resulting word senses must be reusable together in an adaptable exemplar (see Rais-Ghasem, 1998, for more details).</Paragraph>
      <Paragraph position="10"> In Definition 2, correspondence between an input word and a sense-view was defined based on the thematic roles. Since thematic roles are unknown for the input words, the implemented system relies on case markers to find corresponding sense-views for an input word.</Paragraph>
      <Paragraph position="11"> Every sense-views maintains a list of case markers (which may evolve as new members join). Here is an example of the sense-view  The markers with-pp and p-subj indicate that breakers are usually marked either as subject or with-prepositional phrase in the input. Here are examples: The thief broke the windshield with a bat.</Paragraph>
      <Paragraph position="12"> The rack smashed the window.</Paragraph>
      <Paragraph position="13"> Once an adaptable exemplar is found, the model not only knows what senses must be selected for the input words, but also by associating these senses with their corresponding sense-views, the selected senses will be modulated. For example, John in the above example will be associated with the sense-view music-player (a case of property promotion). Similarly, the word play will be disambiguated _to playing-music (as opposed to plavin,~-sport in E2).</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>