<?xml version="1.0" standalone="yes"?>
<Paper uid="J84-2002">
  <Title>The Pragmatics of Referring and the IViodality of Communication 1</Title>
  <Section position="4" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 Previous Research
</SectionTitle>
    <Paragraph position="0"> Four traditions of research bear on the problems at hand: empirical work on differences between spoken and written language, discourse analysis, psychological studies of referential communication, and computational linguistics studies of discourse that have been based on observation of actual communicative interaction.</Paragraph>
    <Paragraph position="1"> The subject of oral/written language comparisons has received much attention from researchers: Anthropologists have traditionally studied characteristics of &amp;quot;oral&amp;quot; and &amp;quot;literate&amp;quot; cultures; human-factors researchers have investigated the opportunities that particular modalities afford for effective communication; and educational psychologists, using empirical and anthropological methods, have sought answers to children's reading and writing problems in the study of oral/written language differences.</Paragraph>
    <Paragraph position="2"> Rubin (1980) discusses methodological weaknesses in many oral/written studies - weaknesses that stem from a simplistic division of language experiences into &amp;quot;oral&amp;quot; and &amp;quot;written&amp;quot;. Instead, she classifies language experiences in terms of their characteristic values on several dimensions such as: the use of voice or print, the ability of &amp;quot;speaker&amp;quot; and &amp;quot;hearer&amp;quot; to interact, their spatial and/or temporal commonality, their mutual involvement in the discourse, and the concreteness of the referents.</Paragraph>
    <Paragraph position="3"> Face-to-face conversations about physically present objects are seen to lie at one extreme within this &amp;quot;communication space&amp;quot; (with &amp;quot;positive&amp;quot; values on the above dimensions), whereas written text is at the opposite. Other language experiences that differ along these dimensions include communication by telephone, keyboard, audiotape, picturephone, writing, etc. 2 Rubin reports that many studies comparing language experiences present conclusions about oral/written language differences even though the language experiences differ from one another along multiple dimensions. In such cases it is not clear if the observed differences result, for example, from the presence of voice, the ability to interact, or both.</Paragraph>
    <Paragraph position="4"> There is evidence that at least some quantitative linguistic and efficacy results are primarily determined by the presence of voice in the communication modality. A series of studies by Chapanis and colleagues (Chapanis et al. 1972, Chapanis et al. 1977) compared problem-solving effectiveness among teams communicating in face-toface, voice only, written, keyboard, and other communication modalities. Dependent measures included problem solution time, number of words, sentences, utterances, etc. Results indicate that problems are solved twice as fast in vocal modalities as they are in written ones, even though communicators use twice as many words when speaking. 3 Although motivating the development of speech-understanding systems, these results unfortunately tell us little about how the processing of spoken utterances differs from the processing of written ones.</Paragraph>
    <Paragraph position="5"> Other research has compared the syntax of spoken and written discourse. The primary findings are: Written language is syntactically more integrated than spoken, employing nominalizations, participles, complements, relative clauses, etc. (Chafe 1982); and spoken language exhibits regular patterns of false starts and hesitations (Hindle 1983, Kroch and Hindle 1982). The former results can help a system designer to determine which syntactic constructs to emphasize in a grammar for parsing. The latter results are more useful to computational z This approach essentially characterizes language situations as multidimensional vectors whose components, describing the above dimensions, are binary values (e.g., +/- voice). Thus, it is assumed that neighboring modalities afford equal communicative possibilities in all dimensions in which they are the same. This is obviously untrue for the dimension of interaction.</Paragraph>
    <Paragraph position="6"> 3 Thompson (1980) has confirmed Chapanis et al.'s results for face-to-face and keyboard modalities.</Paragraph>
    <Paragraph position="7"> Computational Linguistics Volume 10, Number 3-4, July-December 1984 99 Philip R. Cohen The Pragmatics of Referring and the Modality of Communication linguistics. Not only are regularities in nongrammatical speech identified, but a class of &amp;quot;editing&amp;quot; rules is provided that can make such utterances parsable.</Paragraph>
    <Paragraph position="8"> Current work on relaxing grammar rules and on parsing ill-formed input (Hayes and Mouradian 1981, Kwasny and Sondheimer 1981, Weischedel and Black 1980) is in much the same spirit.</Paragraph>
    <Paragraph position="9"> The purposes at hand require analyses of the pragmatic and discourse structure of actual dialogues. Grosz (1977) and Bruce (1981) (among others) have shown how such discourse analyses can have direct implications for algorithm design. In their work, transcripts of dialogues were collected and analyzed, leading to the development of algorithms for speech-understanding systems (Walker 1978, Woods et al. 1976). 4 Grosz' analyses indicate that anaphoric reference in task-oriented dialogues is constrained by the hierarchical structure of the physical task. A parallel structuring of &amp;quot;focus spaces&amp;quot; was proposed as a mechanism to constrain the search for co-referents, and became the mainstay of the discourse component of two systems (Robinson et al.</Paragraph>
    <Paragraph position="10"> 1980, Walker 1978). Although this research did not directly address the problem of discovering cross-modal similarities and differences, the major finding of explicitly &amp;quot;stacked&amp;quot; topics serving to constrain co-reference was validated independently in a domain of casual, face-to-face conversation (Reichman 1981).</Paragraph>
    <Paragraph position="11"> Bruce's pragmatics component for the HWIM system was based on transcripts of human keyboard-mediated dialogues simulating interactions with a travel budget manager. Users were seen to be interacting in various &amp;quot;modes&amp;quot; (e.g., editing-a-trip mode, creating-a-trip mode, etc.). The system attempted to track the user's progress through these modes, using an ATN-based representation, and thereby to create expectations of his/her future utterances. Discourse analysis revealed that users did not follow the strict embedding of subdialogues required by the ATN model. Consequently, the pragmatics component was reorganized as a &amp;quot;demand&amp;quot; model in which the system was seen as responding to one of a set of pending goals. Although this research did not directly address issues of cross-modal similarities and differences, it did point out the promise of a goal-oriented view of language processing.</Paragraph>
    <Paragraph position="12"> Many researchers in the field of discourse analysis have tried to identify goals or intentions in dialogue. For example, Labov and Fanshel (1977) analyzed transcripts of therapy sessions by employing the vocabulary of linguistics and speech act theory. Their analyses presented rules for interpreting the intentions behind utterances of various syntactic forms - e.g., rules for when a hearer will interpret utterances as indirect requests for physical action or verbal confirmation.</Paragraph>
    <Paragraph position="13"> However, these rules were stipulated as regularities of discourse rather than as derived from underlying processes. Their findings should serve as data to be explained, rather than as a satisfying account of discourse.</Paragraph>
    <Paragraph position="14"> In research more relevant to computational linguistics, Mann et al. (Mann, Moore, and Levin 1977; Mann, Carlisle, Moore, and Levin 1977) applied traditional empirical methods to the identification of speaker intention and utterance function in dialogue. 5 Their goal was to build systems to replicate observers' scorings of transcripts. The observers, and ultimately the systems, were to identify repeated reference, requests, expressions of comprehension, topic structure, etc., in keyboard dialogues between a user and a computer operator, and in radio dialogues between Apollo astronauts and ground control. Much care was taken to develop a scoring scheme, train dialogue observers, and attain reliability Mann, Carlisle, Moore, and Levin 1977). A separate computer program was to have been built for processing each transcript. By merging the common features of these systems, an empirically-based theory and computational model were to have been developed. This work resulted in a goal-directed, &amp;quot;dialogue games&amp;quot; model of conversational interaction (Levin and Moore 1977), though it is not clear whether the model's formulation resulted from the merging of implementations.</Paragraph>
    <Paragraph position="15"> Finally, there is a huge literature of psychological studies of referential communication. I will not survey it here (but see Dickson 1981 for recent papers and Asher 1979 for an extensive review), but mention only two themes of relevance to this study. First, such work has shown that, in spoken interaction, noun phrase length tends to decrease as subsequent references to an object are made. However, in non-interactive spoken modalities (Krauss &amp; Weinheimer 1966), the decrease for subsequent references is lessened. These results indicate that efficiency in referential communication is a function of user feedback.</Paragraph>
    <Paragraph position="16"> The development of the component skills involved in referring is a second theme in this literature. In order to test Piaget's &amp;quot;egocentrism&amp;quot; hypotheses, a typical question asked is whether children take their listener's &amp;quot;perspective&amp;quot; into account when planning their referring expressions. 6 Another question raised is whether children of certain ages can adequately make comparisons of the properties of referents and non-referents in order to formulate an adequate referring expression. This line of 4However, neither corpus incorporated true spoken interaction. The SRI dialogues that were analyzed in depth were taken from a mixed communication mode in which one, an &amp;quot;expert&amp;quot;, typed instructions to a third party, who spoke them to an &amp;quot;apprentice&amp;quot;, and typed the apprentice's spoken replies to the expert. The BBN &amp;quot;incremental simulation&amp;quot; dialogues involved only keyboard communication.</Paragraph>
    <Paragraph position="17"> 5Similar approaches include those of Dore et al. (1978), and Sinclair and Coulthard (1975).</Paragraph>
    <Paragraph position="18"> Shatz and Gelman (1973) showed they can do so (though not necessarily accurately (Asher 1979)) at a much earlier age than had been supposed.</Paragraph>
    <Paragraph position="19">  work is more relevant to the present concerns of characterizing the act of identification, but the subskills examined are still too coarse for our needs.</Paragraph>
    <Paragraph position="20"> Previous research has thus provided many lessons, among them: * the need to compare (at least initially) modalities that are minimally different, * the need for repeatable methods for characterizing linguistic behavior at the pragmatics and discourse structure level, * the need to assess the adequacy of our theories, and * the need for couching explanations in computational terms.</Paragraph>
    <Paragraph position="21"> The present study addresses each of these needs.</Paragraph>
  </Section>
class="xml-element"></Paper>