File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/82/c82-1010_metho.xml
Size: 22,022 bytes
Last Modified: 2025-10-06 14:11:23
<?xml version="1.0" standalone="yes"?> <Paper uid="C82-1010"> <Title>A FORMALISM FOR THE STRUCTURAL ANALYSIS OF DIALOGUES</Title> <Section position="3" start_page="61" end_page="61" type="metho"> <SectionTitle> 62 H. COELHO </SectionTitle> <Paragraph position="0"> Let S be a set of conversational states or configurations. A conversational state s is a sequence of at least two related contribution acts. For exemple, cll stands for the first contribution regarding the first conversational state.</Paragraph> <Paragraph position="1"> By a dialogue of length n we mean a member of the set (P x C) of sequences of n contribution acts; an~by a dialogue we mean a member of the set</Paragraph> <Paragraph position="3"> of dialogues of any length. Each member of a dialogue is of the form KS, <p,c)) (s ~S,p ~ P,c ~ C) wich we identify with the triple <p,s,c>.</Paragraph> <Paragraph position="4"> For example, T = {<pl,l,cll> , <p2,l,c21> ,~p1,2,c32> , (p2,2,c42> } is a dialogue of length 4, with 2 participants, pl and p2, 2 conversational states and 4 contributions.</Paragraph> <Paragraph position="5"> Wecall E = PxSxC the set of events, and any triple <p,s,c) an event. A dialogue is a sequence of events, grouped into units, and governed by rules. The conversational units are the invariant structures of dialogue: sub-dialogues, exchanges, monologues and contributions. A sub-dialdgue (dialogue course or segment) is any sequential subset of a dialogue, Course = ~ x:(3tPST) (xht=O and xUt&T)} An exchange is a set of two consecutive events, concerning the same conversational state and two different participants. A pre-defined exchange, conducted by the program, is called an exchange pattern. A monologue is a sequence of at least two consecutive events, concerning the same conversational state and the same participant. Mon ~U((p}xc) n A contribution of a participant to a dialogue is a sequence of his contributions. The semantics of contributions covers the following types: requests (statements, questions, and commands), answers and remarks (eg. agreement).</Paragraph> <Paragraph position="6"> The underlying structures of the situations occurring in a certain problem world determine the organization of dialogue and its systems. A grammar of dialogue is a set of rules of dialogue. Rules of dialogue state how participants understand coherent dialogues, and specify the membership of the set of legal dialogues K, such that K ~ T, where T is the set of dialogues of any length. Rules of dial9gue define the class of coherent dialogue and their attached models. They contain the way contributions are put together. A model is a system of dialogue defined as the triple <P,C,K> . The core of any model is the set R of rules defining K.</Paragraph> </Section> <Section position="4" start_page="61" end_page="61" type="metho"> <SectionTitle> ANALYSIS OF DIALOGUE </SectionTitle> <Paragraph position="0"> We consider the dialogue occuring between the program TUGA Ill and its users in the library world. Such dialogues are evolving dual processes, goal-and-rule-orie_n ted for sharing information between the participants. They are dual because there are only two participantes at a given time. They are goal oriented because they are carried on to satisfy, for example, the following objectives: i) to satisfy users straightforward request concerning the document co~ lection and the classification system, ii) to ask users about the library world (eg. the author of a paper), for conversational.guidance purpose, and iii) to present the user with proposed data (eg, the document classification), enabling him to choose from or modify it.</Paragraph> <Paragraph position="1"> They are rule-oriented because the conversational units are governed by a grammar of dialogueswhich determines the roles played by the program--the librarian and the library's secretary, and by the user. The grammar also allows the inclusion of several courses inside a dialogue, such as topic shifts for simple question-A FORMALISM FOR THE STRUCTURAL ANALYSIS OF DIALOGUES 63 -answering purposes.</Paragraph> <Paragraph position="2"> Consider a dialogue between the program and one of its users, defined as a sequen tial organization of single exchanges. The conversational description is represented in the diagram of the following figure.</Paragraph> <Paragraph position="4"> Conversational description The diagram shows the history of that dialogue. The interaction proceeds in separa te discrete steps from one to another of a finite number of configurations. The~ are three sucessive conversational states through time, each one corresponding to a single exchange, question (q) - answer (a) pair, conducted under user initiative. This history is described as follows.</Paragraph> <Paragraph position="6"> is a dialogue of length 6, with 2 participants, the user (u) and the program (p), with 3 conversational states and 6 contributions (questions and answers). Note that 'q32', for example, means that the third contribution is a question, referring to the second conversational state.</Paragraph> <Paragraph position="7"> Consider now the following segment of a nested dialogue, concerning the classifica tion of a document: (I) u-Classifique &quot;Explorations in Cognition&quot;~ (2) p-~explorations in cognitibn&quot; e desconhecido~ (3) Quer ainda classifica-lo? (4) u-Esta bem.</Paragraph> <Paragraph position="8"> (5) p-Vamos classifica-lo~ (6) Por favor, de-me o titulo de uma referencia do documento em quest~o.</Paragraph> <Paragraph position="9"> (7) u-Qual ~ o tTtulo do Nilsson? (8) p-&quot;artifical intelligence&quot; (9) Vamos classifica-lo.</Paragraph> <Paragraph position="10"> (I0) Por favor, de-me o titulo de outra referencia do documento em quest~o.</Paragraph> <Paragraph position="11"> (II) u-Nao tenho nenhuma.</Paragraph> <Paragraph position="12"> (12) p-Esta bem~ (13) Ja que nao consegui obter nenhuma referencia, n~o sei dar-lhe qualquer sugestao~ Quer ainda classifica-lo? (14) u-Nao quero.</Paragraph> <Paragraph position="13"> (15) p-Esta bem~ E que mais? (Classify &quot;Explorations in Cognition&quot;~) (&quot;explorations in cognition is unknown: Do you still want to classify it?) (All right.) (Let us classify itS) Please give me the title of a reference from the document in question.) (What is the title of Nilsson?) (&quot;artificial intelligence&quot;. Let us classify it~ Please, give me the title of another reference from the document in question.) (I have none.) (All right~ As I could not get any reference, I don't known to give you any suggestion: Do you still want to classify it?) (No, I don't.) (All right~ And what else?) The conversational description, represented in the figure 2, shows the organization and levels of the dialogue. Numbers, placed on the left of contributions, are used for referencing them to the segment considered. 64 H. COELHO r .............. \] O--L ........... UJ i i , ,- 'ts.d.J ! ,, I .................. JL ........... ~ f~l u - u~ ~-ques~on L ......... J - irogrorn o-~.~,i, er r -remork Fig. 2 Conversational description The histo~ of this dialogue is described as follows, T=((u, 1 ,ql l) ,<p ,2 ,a22>,Kp ,2 ,q32> ,4~ 2 ,a4~ ,KP ,3,r53~ ,<p ,3 ,q63 >,~ ,4 ,q74>,</Paragraph> <Paragraph position="15"> is, a dialogue of length 15, with 2 participants, the user (u) and the program (p), 6 conversational states and 15 contributions (questions (q), answers(a.) and remarks (r).</Paragraph> </Section> <Section position="5" start_page="61" end_page="61" type="metho"> <SectionTitle> GRAM~R OF DIALOGUES </SectionTitle> <Paragraph position="0"> The grammar of dialogues of program TUGA is a complete and precise description of the properties of a certain class of dialogues. The properties concern the structures of the _dialogues, occurring in a libra~ world, and organized as models.</Paragraph> <Paragraph position="1"> This grammar machine~ is able to parse situations, and it is-ve~ much like the one able to parse a natural language sentence: the objects recognized, dialogue units, are characterized as structured objects asse~led out of recognizable parts according to know rules of assembly.</Paragraph> <Paragraph position="2"> A dialogue carried out ~ TUGA has ~o participants, the program (p) and its users (u), and therefore ~o mutually exclusive states, the &quot;agent&quot; and the &quot;passive par ticipante&quot;. Both participants may take the initiative during the encounter, ie. the program may be an &quot;agent&quot; or a &quot;passive participant&quot;. The &quot;agent&quot; claims the turn to speak at any given moment, and plays an active role. The &quot;passive participant&quot; does not claim the turn to speak at any given moment.</Paragraph> <Paragraph position="3"> Considering ~o states for each participant, there are four possible conversational states. However, we only consider ~o states: &quot;agent&quot;-- &quot;passive participant&quot; and &quot;passive participant&quot;-- &quot;agent&quot;. (The other ~o states represent in some sense a failure of dialogue.) The BNF specification of the grammar syntax above the discourse level, presented below, characterizes only th~ class of dialogues considered.</Paragraph> <Paragraph position="4"> Let us consider only the first few rules in order to make explicit their meaii~g.A general dialogue, 'converse', is defined as an opening following by a sub-dialogue which may be followed by a sub-dialogue or closed by user initiative. The user may also suspend t~llporarily the dialogue without affecting it. This feature justifies the existence of two kinds of opening: one for the dialogue start and the other for the re-start. A dialogue is simply a sequence of exchanges or monologue, or is followed by several models of dialogue. For example, dialogue on the classificatim of a document is handled by dialogue model 'converse5', which is defined by rules 'course'. These rules define several kinds of possible courses during the interaction. Dialogue on adding new documents is handled by dialogue model 'converse6', which is served by a set of exchange patterns :- a sequence of pre-defined program questions whose order may be altered by user. Some feature of a sentence, together with the current context, can trigger a hypothesis that an instance of some particular model or pattern is being conveyed. All this means that the program can cope with user changes of mind, single or multiple data, provided in any order, and can avoid asking questions whose answerswere provided either implicitly or explicitly at some earlier time.</Paragraph> <Paragraph position="5"> We use rules of interpretation, below the discourse level. The rules of interpreta tion deal with what the user does, eg, requests (statements, questions and commands) and answers. Other rules deal with what the program does, eg, answers, questions, and remarks (comments and agreements). Here are, for example, thne'e of these r~1~s: Rule-- If the user makes a statement, the program interprets it as a request for confirmation.</Paragraph> <Paragraph position="6"> Rule-- If the user asks a closed question (form Q-S, where S means the statement corresponding to the question) and the program responds with an existential E(yes/ /no), then the program is understood as answering the user with the tratement E-S.</Paragraph> <Paragraph position="7"> Rule--If the user issues a command, then the program interprets it as a valid request for an action A only if the following conditions hold: 66 H. COELHO the request is ended with an exclamation mark, and action A is classifying a document, generating a category, adding data items and, deleting data items.</Paragraph> <Paragraph position="8"> The first two actions also cover the general purpose of gathering information through a referent: the referenced document or the classification.</Paragraph> </Section> <Section position="6" start_page="61" end_page="61" type="metho"> <SectionTitle> DIALOGUE ORGANIZATION STRUCTURES AND LEVELS </SectionTitle> <Paragraph position="0"> Any original natural language sentence is parsed by a cascade of two DCG's. The first DCG, above the discourse level, represents the syntactic information about dialogue form (exchange patterns and dialogue models) and is closely coupled to a set of scenarios which represent the pragmatic information about the task domain (a collection of situation descriptions). The second DCG, on the sentence level, represents the syntatic information about sentence form and the general semantic information.</Paragraph> <Paragraph position="1"> Let us observe in more detail the main dialogue forms, engineered and manipulated by the first DCG in order to build up the overall skeleton of the dialogue hJstor~ An exchange pattern (eg.'dialogue') is a pre-defined exchange between the program and the user. It is defined by a name and a number, and provided by a message and the number of the following contribution. It consists of a question of an expected form, followed by a simple dialogue. The question is constructed with the value of the message (eg. a proper noun). The simple dialogue is the standard way to inter pret user contributions: the question-answer pair. As regards exchange patterns, the user contribution expected is not always an answer: the user also question the program, and by doing so a new dialogue is nested in the previous ..ne. Program questions are motivated by the content of user request. For exemple, interrogating 'the name of a new category' and 'under what categories may it be placed' are generated when the user wants to create a new category in thec~ssification system.</Paragraph> <Paragraph position="2"> The exchange pattern is called by the grammar of dialogues through its name and number. In case of non acceptance of the program question by the user, the initia tive for restarting the dialogue belongs to the user. But the new dialogue may b'e nested in the previous dialogue, as often occurs in the process of classifying new documents.</Paragraph> <Paragraph position="3"> A dialogue model (eg. 'converse') is a suite of unconstrained exchanges between the program and the user. It generates detailed expectations about the next contri bution, by having an ordering for calling exchange patterns which may be alteredby user. The user may give several answe~which need not be ordered. Also, he may modify his previous answers. The program uses the success or failure of is predic tions to determine what role the user contribution plays in the dialogue. Whene~r a dialogue model is activated, an appropriate exchange pattern is invoked,and the program poses a question to the user and interprets the user response. If a failure occurs, the program is able to come back to the same topic. For example, during the classification of a documents the user may oppose the program and request information about the location of a category in the classification system. And, the program only accepts three titles of documents,~nowD to its data base, and according to its classification method. Therefore, it goes on asking the user till it attains that limit, and skips any unknown title. But if the user gives up, the program restarts a new dialogue. These features are implemented either by using recursion or backtracking. R~cursion is the ability of a procedure to contain a procedure call to another copy of itself. The declaration of the procedure, for the dialogue model in charge of the classification process, contains a procedure call which matches the name of another copy of the same procedure declaration. Counters control the process of recursion during the program asking for references in the classification process. And, an handling device deals with contradictions arising when a new document is archived. Numbers (l and O) are assigned according to existing contradiction or otherwise, and summed over the facts. If the result is A FORMALISM FOR THE STRUCTURAL ANALYSIS OF DIALOGUES 67 non-zero, a failure aries forcing backtracking and the restarting of the process.</Paragraph> <Paragraph position="4"> In the course of conversing, the program remembers the events, by storing conversational states containing the name of the participant, the number of the contribution, and the active information, such as questions or answers. This knowledge about the current dialogue history and skeleton is maintained in order to be served in any specific situation or instance to which it applies. In fact, it is the support for the organizational ability of the grammar of dialogues .The example below illustrate this ability. Consider the segment of a nested dialogue, presented in figure 2, concerning the classification of a document.This dialogue presents a sub-dialogue and an interruption due to the user. The sub-dialogue corresponds to a dialogue model, which organizes interactions concerning the implementation of the document classification method. The interruption is presented as an exchange,nested in the sub-dialogue. This conversational description shows the organization, the depth of nesting, the change of initiative, and the levels of the dialogue (one for the ordinary tdalogue; the other for the particular dialogue model; and, the last one for the interruption).</Paragraph> <Paragraph position="5"> Consider another example of a dialogue (numbers appended to the natural language contributions correspond to the conversational states).</Paragraph> <Paragraph position="6"> (I) u-Crie Prolog~ p-Prolog e conhecido~ (2) u-Quero criar uma categoria~ (3) p-Por favor, qual o home da nova categoria? u-&quot;Pattern recognition&quot;.</Paragraph> <Paragraph position="7"> p-&quot;Pattern recognition&quot; e desonhecida~ (4) Pot favor, debaixo de que categoria a pretende inserir? u-&quot;Machine vision&quot;.</Paragraph> <Paragraph position="8"> p-&quot;Machine vision&quot; ~ conhecida~ (5) A nova categoria ficou inserida no sistema de classifica9ao e recebeu o numero 2141.</Paragraph> <Paragraph position="9"> E que mais? The history of this dialogue is: (Create Prolog~)(Prolog is known~) (I want to generate a category,) Please, what is the name of the new category?) &quot;Pattern recognition&quot; is unknown~ Please, under what category do you want to insert it?) &quot;Machine vision&quot; is known~ The new category has been inserted in the classification system and received 2141 as a number.</Paragraph> <Paragraph position="10"> The program knows that the first user contribution, a command,Ku,l,cll> , opens a dialogue composed of a simple question-answer pair (conversational state I). The dialogue goes on with another user command,<u,2,c32>, which invokes a dialogue 68 H. COELHO model for creating new categories. This dialogue ~ode\] calls two exchange patterns, and the dialogue is closed with the program answer<p,2,a92>. Note that this last event has the same conversational state number (2) as the event invoking the dialogue model. This information, shared by the program and one of its users, helps the program to decide on what to do and how to proceed. It chooses its course of action by inspecting previous user decisions, through the remembering mechanism. The dialogue model, responsible for gathering information about new documents, exemplifies the use of backtracking. If the user changes his mind at any stage of the dialogue, the program backtracks to follow up the consequences of the new information. Supplied facts contradicting those already known are detected in the immediate interpretation of the user's input, and when this interpretation is complete, a failure leading to the restart of interpretation at a previous level occurs if a contradiction has been found. This mechanism overrides the repetition of unanswered questions, and skips questions by recognizing the content and form of the user's answers.</Paragraph> </Section> <Section position="7" start_page="61" end_page="61" type="metho"> <SectionTitle> SCENARIOS </SectionTitle> <Paragraph position="0"> Scenarios are sets of expectations and presumptions regarding a certain type of situation, They are of the form &quot;if <situation description> then expect <situation description)&quot;, or &quot;if <situation description~ is a satisfied then do <action description>&quot;. Situation and action descriptions are triggered by verbs. Scenarios are used in TUGA \[l~ as recognition devices for classifying and identifying situations in a dialogue, and they call the exchange patterns and organize their invocation. All the embedded knowledge embedded in the scenarios covers the ability.</Paragraph> <Paragraph position="1"> - to derive questions from relevant information or from the logical consequences of the information that is known about the questioned topic combined with general knowledge of the library world, and - to handle the user's answers TUGA is a program able to play two roles in the library world.</Paragraph> <Paragraph position="2"> It acts as a librarian and as a library's secretary.</Paragraph> <Paragraph position="3"> Possible events in the library world are grouped into the following scenarios: inside scenario C.</Paragraph> <Paragraph position="4"> Exchange patterns are classified according to their use in these scenarios, as shown in the following figure. The classification is made possible thnpugh two of their arguments: name and number.</Paragraph> <Paragraph position="5"> This taxonomy for situation recognition is made available during a dialogue. It contains the pragmatic knowledge of the task domain, and supports the p, ogram ability to converse with users in a more clever way, when put aside the gi-..~mar of dialogues.</Paragraph> </Section> class="xml-element"></Paper>