autoroute dialogue demonstrator

20
URL: http://www.cam.sri.com/tr/crc073/paper.ps.Z Trindi Deliverable; 1998

Upload: nottingham

Post on 21-Nov-2023

0 views

Category:

Documents


0 download

TRANSCRIPT

UR

L: h

ttp://

ww

w.c

am.s

ri.c

om/tr

/crc

073/

pape

r.ps

.ZT

rind

i Del

iver

able

; 199

8

Autoroute Dialogue Demonstrator �Ian LewinSRI InternationalCambridge Computer Science Research Centre,23 Millers Yard,Mill Lane,Cambridge,United Kingdom CB2 [email protected] 9, 19981 OverviewThe Autoroute demonstrator possesses a dialogue capability based upon thenotion of \Conversational Game" [Pow79] [Hou86] [CII+97]. Our work herehas been both theoretical - seeking a clearer picture of what the centralnotions are and how they can be re-applied - and implementational. Theresult is an improved dialogue manager whose recon�guration is designed tofollow from re-application of a much clearer theoretical analysis to a domain.�This report is extracted fromRecon�gurable Architectures for Spoken Dialogue Under-standing, prepared by SRI International Cambridge Computer Science Research Centrefor the UK Defence Evaluation and Research Agency, Malvern under the terms of Con-tract CSM3/002 for the Manager, Systems Engineering and Information Group, DERAMalvern 1

2 IntroductionOur work on building a recon�gurable dialogue architecture has been basedon developing a clear theoretical picture of dialogue which can, and has been,implemented cleanly so that a recon�guration of the software can followrelatively easily from a re-application of the theory to a domain.Our theory has used and developed the notion of conversational game andmove [Pow79] [Hou86] [CII+97]. We originally chose this notion partly for itsempirical basis. Its move categories appeared su�cient to describe most ofour Wizard-of-Oz spoken route enquiry corpus [MB92]. It also includes rule-governed sub-structures. Our development includes particularly the provisionof a clear declarative and, we believe, intuitive semantics. We thereby gleanthe usual bene�ts of separating linguistic description from computationalprocessing. We are also particularly interested in system recon�gurability.Through formal de�nition of what our dialogue acts mean - de�nitions whichre ect an intuitive picture - we provide a clear framework for re-use of ourdialogue manager. We believe our semantics may also help provide the basisfor a �rmer empirical foundation for the original theory.In the following sections we �rst describe the theoretical basis for ourdialogue manager and give an illustrative example. Section 4 gives a moreformal account of the central notions. Section 5 illustrates how the notionsare applied in the current Demonstrator system. Finally, Section 6 discussesthe issue of dialogue management recon�gurability. Conclusions are drawnin Section 7.3 A Conversational Game Dialogue Manager3.1 Theoretical OverviewIn general, dialogues result when rational agents, planning to satisfy theirown goals, decide to invoke the help of other rational agents. Starting aconversation is one way amongst others of, for example, �nding somethingout or getting something done. There are usually other ways of achievingone's objectives (to �nd something out one might consult an encyclopaedia)but, for a variety of reasons, an agent may prefer to engage other agents ina joint enterprise.Dialogues generally consist of many individual exchanges between the

participants. In the Autoroute dialogues, there are generally exchanges con-cerning the origin, destination and time of the user's planned journey. Theinternal structure of these exchanges is the central topic of conversationalgame theory. Given the rational agency model, one can also expect the over-all dialogue structure to re ect any planned structures designed to satisfythe agents' non-linguistic objectives. In this way, familiar claims concerningthe determination of dialogue structure by task structure (cf. [GS86]) maybe accommodated but this is not a central feature of the theory.Conversational Game Theory claims that there are conversational rulesgoverning exchanges. These rules are a linguistic resource for potential con-versants. They are known and exploited by conversants. Agents may bejudged by their adherence or otherwise. For example, asking a question isa conversational move that opens one particular type of exchange (a \qw"game). Once this move is made, subsequent utterances, or the lack of them,are liable to interpretation in the light of this game. There is, for instance,a social obligation to help complete an open game. Complaints that thereare no rules of conversation since virtually anything can follow anything (cf[Lev83]) therefore simply miss the point that although one can do anythingwhat one does do is liable to judgment.Previous accounts of conversational game theory have remained some-what vague on what games and moves actually are other than to say thatthey are functionally de�ned and are related to goal satisfaction. There arediscursive accounts of some particular moves and games, e.g.Check: A checking move requests the partner to con�rm infor-mation that the checker has some reason to believe but is notentirely sure about . . . [typically] . . . something which the part-ner has tried to convey explicitly or something which the checkerbelieves was meant to be inferred.These permit others to reproduce expert dialogue codings reasonably well[CII+97]. However, the underlying basis for the expert coding remains un-clear and, overall, the theory has remained an abstract architecture withinwhich speci�c proposals can be made.Our development de�nes the e�ect of a game in the public establishmentof \commitments". An agent who makes an assertion is generally committedto its truth. A promise commits one to its execution. Commitments do notimply the agent is in any particular mental state. The agent may or may not

believe what he asserted. He may or may not desire what he promised. Hehas however made a publicmove and laid himself open to judgement. Gamesare joint enterprises. Commitments are similarly joint. If one agent makes anassertion and another does not dissent then, generally, both are committed toits truth. Each can legitimately complain if the other subsequently behavesinconsistently with the public position. One might say, \You didn't objectwhen I said ..." or \I thought we agreed that ...".Conversational games result in public commitments through the makingof conversational moves. Asking a question raises an issue of the truth ofsome proposition. Making a request raises the prospect of a public promise.Answering a question o�ers a proposition to settle an issue raised. Agreeingto a request establishes a promise to act. Conversational moves are realizedby utterances. An utterance may realize a move simply in virtue of form (e.g.\yes" may signal an acknowledgment) or in virtue of its truth-conditionalmeaning. That \My destination is London" answers \Where are you going?"whereas \I enjoy ice-cream" does not is clearly a function of the meaningsof the three utterances. Unlike [Hou86], we do not construe moves as speechacts. Following the arguments of [Lev83], there is no force of replying yetreplying is a paradigm move.Within a game, there is a set of propositions under discussion. The set isinitialized by the opening move. The function of the remaining conversationalmoves is to alter this set. Once the game ends, any remaining propositionsare conversationally agreed.3.2 Illustrative ExampleTop-level dialogue management - the rational agent - consists of a simpleplanning-and-execution agent. The agent plans a sequence of actions, in-cluding the playing of conversational games, using a traditional AI planningalgorithm. Once a plan is found, the agent attempts to execute it. Somevery simple re-planning procedures can be invoked during plan execution ifthe plan fails or if some later steps of the original plan become redundantduring execution of earlier steps.The games themselves are played by traversing the recursive transitionnetworks which de�ne them. Traversal is a hill-climbing search with back-tracking. That is, whenever a choice is required, the set of possible nextmoves is sorted according to an evaluation function and the highest scoringmove is chosen. There is a prede�ned set of scoring functions and associated

weights. For example, one scoring function predicts the expected length of agame following a given move. Another calculates expected accuracy of un-derstanding. If a search path runs into a dead-end, the system backtracks tothe last open choice point and picks the next highest scoring move. All pos-sible analyses are therefore tried eventually. The system may also backtrackinto re-analysis of its own utterances.On the user's turn, the CLE is called to determine the syntax and seman-tics of the utterance. A move recognition process uses these plus the currentgame state (current node and propositions under discussion) to determinethe possible move analyses. Then, the preference mechanism is called tochoose between them.On the system's turn, the set of next possible moves is calculated from thegame de�nitions and the preference mechanism is called to choose betweenthem. Then, a move realization process determines an utterance meaningfor the chosen move and the CLE is again called to generate a string fromthe meaning. (Most calls are cached in advance to save processing time butsome meanings such as the contents of checking moves cannot be predicted).We illustrate our system simulating a game from our corpus using thenetworks shown below. (Although not illustrated, pardon games may connectany non-terminal node to itself). The extracted dialogue is from autoroutedialogue d4 4 [MB92].w: What time would you like to make your journeyu: Pardonw: What time would you like to make your journeyu: Five o'clock in the eveningw: Setting o� at �ve P Mu: Yes(System Trace)1 w: What time do you want to travel2 u: Pardon3 w: Please state a departure time, for example \At four P M"4 u: Five o'clock in the evening5 w: Is the departure time seventeen hundred hour6 u: Yes

?��� �6 � �6�& %�� ?��6� �� �6n nnn n�� �� nn n n n- -- -PARDON: Unrecog PardonQW: Qw-restrict ReplyQw CNFAckSilentAck CNF: Cnf Reply-yesReply-noReply-modFigure 1: Game De�nitionsThe rational agent chooses a QW game concerning journey-time. Net-work traversal begins and move evaluation prefers a qw move. The CLErealizes it as 1. Propositions concerning journey-time become under discus-sion. The user responds with 2 which the CLE analyzes as meaning pardon.Move recognition suggests one analysis: a pardon move. Since pardon is nota valid next move in QW, backtracking occurs. One possible path includes anested pardon game at the initial QW node and, since 1 can be re-analyzedas an unrecognized move, this path proves successful. Once more at theinitial node, move evaluation now prefers a qw restrictmove, realized by 3.Again, propositions concerning journey-time become under discussion. Next,4 is analyzed as meaning dbroute starttime(17:0) and move recognitiono�ers a reply analysis because the meaning is a subset of what was underdiscussion. (Replies containing extra information are still replies. ) Aftertraversing the reply-arc, a start-time of 5pm is under discussion. Move evalu-ation now suggests con�rming dbroute starttime(17:0). The CLE realizesthe opening move as 5. Move evaluation suggests that 6 is a reply-yesmove.The nested game therefore closes and dbroute starttime(17:0) becomesagreed because con�rmation games add negotiated propositions to publiccommitments. They also delete their contents from the propositions underdiscussion. Consequently, in our QW game no propositions remain underdiscussion. Move evaluation now prefers QW completion without acknowl-edgment. With nothing left unnegotiated, completion causes no further com-mitment updates. Control returns to the rational agent. The conversants arecommitted to one new proposition.

4 A Formal Account4.1 Rational AgencyEach conversational agent is a planner characterized by a goal g describinga preferred state of the world, a set of sentences Init describing the way theworld currently is, a set of action designators A which includes a set of con-versational games G, and a set of action postulates P describing how actionsa alter the world in terms of pre- and post- conditions. These conditions,pre(a) and post(a), are also sets of sentences. We write s j= Q meaning thatthe conjunction of all the sentences in Q describe the state denoted by s.The actions postulates take the form:si j= pre(a)! (execute(si; a; si+1)! si+1 j= post(a))(If a's precondition is true in a situation si then, if a is executed in siresulting in si+1 then a's postcondition is true in si+1.)A plan is a sequence of actions which will result in a state of the worldin which the agent's goal is true. Formally, let [HjT ] denote a sequence ofactions whose �rst member is H and whose others form a sequence T , then,where s0 j= Init[] is a plan for g in s0 $ s0 j= g[HjT ] is a plan for g in s0 $ 9s execute(s0;H; s) & T is a plan for g in sIn order to execute a plan, each action a in the plan must be associatedwith a procedure prog(a) for executing it.4.2 Conversational Games4.2.1 SyntaxThere are �nite sets of game types G and move types M and an in�nite setof propositions P . g : h ; pi denotes a game g of type 2 G and contentp 2 Powerset(P ). m : h�; qi denotes a move of type � 2 M and contentq 2 Powerset(P ). Game content is the issue raised by the game, namely thepropositions brought under discussion by the opening move. Move content isgenerally the meaning or truth conditional content of the utterance realizingit. There is a function Grammar from game types to grammars. Grammarsare recursive transition networks whose arcs are labelled with conversational

move and game types. That is, Grammar( ) = hs; t; r; i; fi, where 2 G, sis a set of states labelled with the player who has the next turn , t � (M[G)is a set of arc labels, r � s� t� s is the transition relation, i is the initialgame state and f � s is the set of �nal game states. A sequence ofutterances u0; u1; : : : ; un realizes g : h ; pi if there is a sequence of movesm0 : h�0; q0i,m1 : h�1; q1i,. . . , mn : h�n; qni such that each ui realizes mi, themove type sequence �0; �1; : : : ; �n labels a path through Grammar( ), andq0 = p.4.2.2 SemanticsA model for a dialogue consists of hI1; I2; P; U;Cn; Pd;Mei where I1 inter-prets move and game types M [ G; I2 interprets game types G; P is the setof propositions; U is a set of undertakings; Cn : U ! Powerset(P ) is the setof possible commitments; Pd = Powerset(P ) is the possible sets of proposi-tions under discussion; Me = Powerset(P ) is the set of possible utterancemeanings. xZ is a variable x over the set Z.I2 and [[ ]]2 de�ne the \commitment" value of games and moves. I1 and[[ ]]1 de�ne the \propositions under discussion" value of games and moves.Each move m : h�; qi determines a function Pd! Pd. q determines whatto update Pd with and � determines how to update it.I1(�) : Pd �Me! Pd[[m : h�; qi ]]1 = �xPd:[I1(�)(x; q)]Games (occurring as moves within games) are similar:I1( ) : Pd�Me!Pd . A move sequence which makes a game determines by functional compo-sition a function ; ! Pd. ; indicates that nothing is under discussion beforethe �rst move. The function's range indicates the propositions to becomecommitments.Each game determines a function Cn! Cn. The propositions negotiatedin a game determine what to update Cn with and the game type determineshow to update it. Games may occur within games so the total commitmente�ect of a game includes any e�ects of nested games.I2( ) = Fn : Cn� Pd! Cn[[ [g:h ;pi m0; : : : ;mn] ]]2 = �xCn:[I2( )([[ m0; : : : ;mn ]]2(x); [[m0; : : : ;mn ]]1(;)))

A game sequence which makes a dialogue determines by composition afunction ; ! Cn. ; indicates the absence of commitments before the �rstgame. The function's range indicates the set of commitments entered intoduring the dialogue. Commitments of D = [[ D ]]2(;).5 The Demonstrator System5.1 Rational Agency DemonstratorWe plan in a simple version of the situation calculus, where the planningoperators are of formoperator(actions(S0,Action,S1),Preconditions,Postcondition)meaning that if Preconditions are true (they should be conditions on S0),and Action occurs then S1 is the result and Postconditions hold true in S1.A sample of operators for the autoroute domain is shown in �gure 2. Forexample, playing a qw game leads from a situation where Agent Ag2 knowswhat something is to the situation where that knowledge is mutually knownand agreed (on the scoreboard). An action of �nding a route leads from asituation where the parameters of the route are known to a situation wherethe route itself is known.One can prove that something is true in a situation by� looking it up in a database of facts� using a frame axiom and proving that it is true in some other situation� using some other axiomThe axioms are horn clauses of formphc(id, consequent, antecedent)meaning that consequent is true if antecedent is true. (id is just anidenti�er for horn clauses).A sample of axioms for the autoroute domain is shown in �gure 3. Theaxioms include a very simple frame axiom (once something is on the score-board, then it stays there) and de�nitions of what it means to know whatthe value of a journey parameter is (eg you know what the destination is ifthere is some destination named).

% Planning Operatorsoperator(actions(S0,[play,game(qw,Fn),Ag1,Ag2],S1),[channel(S0,Ag1,Ag2),knows_ref(S0,Ag2,Fn)],scoreboard(S1,Fn)).operator(actions(S0,[play,game(assume,Fn),Ag1,Ag2],S1),[channel(S0,Ag1,Ag2),knows_ref(S0,Ag2,Fn)],scoreboard(S1,Fn)).operator(actions(S0,[find_route,Ag1,autoroute],S1),[channel(S0,Ag1,autoroute),scoreboard(S0,quickshort),scoreboard(S0,x^time(x)),scoreboard(S0,x^from(x)),scoreboard(S0,x^to(x))],knows_ref(S1,Ag1,x^route(x))).operator(actions(S0,[play,game(inf,Fn),Ag1,Ag2],S1),[channel(S0,Ag1,Ag2),knows_ref(S0,Ag1,Fn)],scoreboard(S1,Fn)).operator(actions(S0,[play,game(goodbye,M),Ag1,Ag2],_S1),[channel(S0,Ag1,Ag2)],scoreboard(S1,have_bid_farewell)).Figure 2: planning operators and axioms

% 2. Planning Domain Axioms% Frame Axiom: If p is true in S0 and S1 is the situation% following S0, then p is true S1 toophc(scoreboard,scoreboard(S1,Fn),and(earliest(S1,S0),scoreboard(S0,Fn))).% Defn of `earliest'% earliest(X,Y) is true if Y is the situation from which X% derives. Examples: earliest(X,X), earliest(s(s(X)),X).phc(earliest,earliest(S1,S1),var(S1)).phc(earliest,earliest(S1,S2),and(nonvar(S1),and(S1 =.. [s,S0],earliest(S0,S2)))).% Defns of knowing WHAT something is in terms of% knowing THAT something isphc(knows_ref,knows_ref(x^to(x)),exists([I,X],and(dbroute_endplace(I,X),\+ term_contains_skolems(X)))).phc(knows_ref,knows_ref(x^time(x)),exists([I,X],and(dbroute_starttime(I,X),\+ term_contains_skolems(X)))).phc(knows_ref,knows_ref(x^time(x)),exists([I,X],and(dbroute_endtime(I,X),\+ term_contains_skolems(X)))).phc(channel,channel(_,'the CLARE system',autoroute),true). Figure 3: planning axioms

5.2 Conversational Games DemonstratorIn �gure 4, the game de�nitions for the example system are shown (ie thefunction Grammar discussed earlier). So G = f qw, cnf, inf, assume, hello,goodbye, apology, pardon g, and M = f qw, qw restrict, rw, ack, eta, cnf,ryes, rno, rmod, inf, assume, hello, goodbye, end, apology, unrecognizable,pardon, abandon g . The networks are de�ned by clauses of the followingform: arc(game(qw),s0/p1,move(qw):copy,s1/p2)In this example, an arc in the qw network is denoted which leads fromnode s0 (where it is p1's turn to move) to node s1 (where it is p2's turn). Thearc is labelled with \move(qw):copy". \copy" is I1(move), ie the semanticsof this move are that the propositions under discussion are simply copiedfrom one game node to the next. Other examples include rno which is areply-no move in a con�rmation game which simply empties the current setof propositions under discussion and rmod which is a reply-modi�ed movewhich replaces the current propositions under discussion with the proposi-tional content of the move itself.5.3 Preference Module DemonstratorIn �gure 5, the preference functions for the example system are declaredand weights are assigned. In move recognition, only one function is de-�ned dm move ana (its weight v(move move ana) = 1). Furthermore, thefunction scores every move with value 1 (f(m) = 1 for all moves m).Consequently, there is no interesting way of choosing between possible moveanalyses in recognition. In move generation however, three functions are de-�ned and dm pv confidence, which reports how con�dent we are in whatwe believe are the current propositions under discussion, is assigned a higherweight than the other functions. Therefore the agent is likely to prefer accu-racy over speed.In �gure 6, the preference functions themselves are de�ned. The scoringfunction for expected game length, for example, scores the eta move veryhighly and cnf and pardon games very low. The scoring function whichanticipates how much information we may glean from a game scores ordinaryquestion moves (such as \Where do you want to go?") higher than more

% Conversational Game Definitionsarc(game(qw),s0/p1,move(qw):copy,s1/p2).arc(game(qw),s0/p1,move(qw_restrict):copy,s1/p2).arc(game(qw),s1/p2,move(rw):replace,s2/p1).arc(game(qw),s2/p1,move(ack):copy,s3/p2).arc(game(qw),s2/p1,move(eta):copy,s3/p2).arc(game(qw),s2/p1,game(cnf):delete,s2/p1).arc(game(cnf),s0/p1,move(cnf):copy,s1/p2).arc(game(cnf),s1/p2,move(ryes):copy,s2/p1).arc(game(cnf),s1/p2,move(rno):empty,s3/p1).arc(game(cnf),s1/p2,move(rmod):replace,s0/p1).arc(game(inf),s0/p1,move(inf):copy,s1/p2).arc(game(assume),s0/p1,move(assume):replace,s1/p2).arc(game(hello),s0/p1,move(hello):copy,s1/p2).arc(game(hello),s1/p2,move(hello):add,s2/p1).arc(game(goodbye),s0/p1,move(goodbye),s1/p2).arc(game(goodbye),s1/_,move(end),s2/p1).arc(game(apology),s0/p1,move(apology),s1/p2).arc(game(pardon),s0/p1,move(unrecognizable):copy,s1/p2).arc(game(pardon),s1/p2,move(pardon):copy,s2/p1).arc(game(B),A,game(pardon):copy,A) :- \+ B = pardon.arc(game(_),A,move(abandon):none,nowhere).final_state(game(qw),s3/p2,add).final_state(game(cnf),s2/p1,add).final_state(game(cnf),s3/p1,add).final_state(game(cnf),s5/p1,add).final_state(game(inf),s1/p2,add).final_state(game(assume),s1/p2,add).final_state(game(hello),s2/p1,add).final_state(game(goodbye),s2/p1,add).final_state(game(apology),s1/p2,add).final_state(game(pardon),s2/p1,id).Figure 4: Game definitions

%% define the dialogue manager scoring functions and their weights%% dm(ana) functions score move analyses of input utterances%% dm(nm) functions score potential next moves by the system%%% Name: scoring_function/3%% 1 : + : processing stage id%% 2 : + : name of scoring function%% 3 : + : scoring function weightscoring_function(dm(ana),dm_move_ana,1).scoring_function(dm(nm),dm_pv_confidence,2).scoring_function(dm(nm),dm_exp_slot_fillers,1).scoring_function(dm(nm),dm_exp_game_length,1).Figure 5: preference function declaration and weightsrestricted versions of a question (such as \Please state the name of yourdestination").6 Recon�gurabilityWe distinguish game strategy, game knowledge and rational agency.6.1 Recon�gurable Game StrategyBy \game strategy" we mean how to make the choices required in order toplay particular instances of conversational games. \Conversational GameTheory" de�nes what a game is in terms of the possible strings of moves thatcan realize it. The theory does not (and is not intended to) determine whichmove an agent should make of all possible moves that he could make. Thetheory only determines the set of possible next moves. Similarly, in analysis,the theory does not (and is not intended to) determine which move wasrealized by an utterance of all the possible moves that it might have realized.In order to make such choices, all the possible next moves (or, analyses of thelast move) are evaluated using weighted preference functions and the movewith the best score is chosen. That is, if f(m) is the score assigned to move

%% Defines preference functions on dialogue moves%%%% Each preference measure should define a score for every move type%% defined in the conversational games and for every possible game%% state via the format:%% scoring_function_name(move_name:game_state,score)dm_move_ana(_MoveType:_,1).dm_pv_confidence([game(cnf),dbroute_type(_,_)]:_,0) :- !.dm_pv_confidence([game(cnf),dbroute_header(_)]:_,0) :- !.dm_pv_confidence([game(cnf),_]:_,1) :- !.dm_pv_confidence([move(cnf),_]:_,1) :- !.dm_pv_confidence([move(pardon),_]:_,1) :- !.dm_pv_confidence([move(inf),_]:_,1) :- !.dm_pv_confidence(_:_,0).dm_exp_slot_fillers([move(qw),_]:ds([_,Rate]),3) :- Rate >= 0.5,!.dm_exp_slot_fillers([move(qw_restrict),_]:ds([_,Rate]),2) :- Rate >= 0.5,!.dm_exp_slot_fillers([move(qw),_]:ds([_,Rate]),2) :- Rate < 0.5,!.dm_exp_slot_fillers([move(qw_restrict),_]:ds([_,Rate]),3) :- Rate < 0.5,!.dm_exp_slot_fillers([game(cnf),dbroute_type(_,_)]:_,0) :-!.dm_exp_slot_fillers([game(cnf),dbroute_header(_)]:_,0) :-!.dm_exp_slot_fillers([game(cnf),_]:_,2) :-!.dm_exp_slot_fillers(_:_,0) :- !.dm_exp_game_length([move(eta),_]:_,3) :-!.dm_exp_game_length([move(ack),_]:_,2) :-!.dm_exp_game_length([game(cnf),_]:_,0) :-!.dm_exp_game_length([game(pardon),_]:_,0) :-!.dm_exp_game_length(_:_,1) :-!.Figure 6: preference function de�nitions

m by preference function f and v(f) is the weight assigned by the agent tothe preference function. Then, to calculate which move to choose, the agentcalculates for each move its utility U(m) de�ned by:U(m) =Xf v(f)� f(m)The agent then chooses the move with the highest utility.Each preference function represents some salient feature of a possiblemove and determines a numeric score according to how much the featureapplies to that move. For example, in the current Autoroute system, there arethree features applicable to choices of the next possible move to make. Onefeature concerns how much con�dence one has in one's current belief of whatthe current state of the game is. Con�rmation moves and games naturallyhave a higher value for this feature than other possible moves since they arelikely to increase our con�dence in that we have understood what has gonebefore. A second feature concerns how much additional information (in termsof slot �llers) one might expect to gain from a given move. An \unrestricted"question move such as \Where do you want to go?" scores more highly withregard to this feature than \Please state the name of your destination" sincethe former are more likely to be answered by more informative answers suchas \I want to go fromMalvern to Cambridge". The third feature measures thelikely impact on game length of making a given move. This feature positivelydisfavours con�rmation games and encourages simple acknowledgments.The weights on the preference functions are also numeric scores and re ectthe value of the preference function to the agent. For example, a \safety-�rst" agent will generally attach a high weight to the con�dence preferencefunction and a lesser weight to the expected game length.Di�erent game strategies are con�gured by altering the scoring functionweights. One can easily experiment with making the system strongly dispre-fer lengthy games or not care overly about accuracy. The functions them-selves can also be modi�ed. One might, for example, make the scoring sys-tem sensitive to which speech recognizer was being used or how con�dentthe speech recognizer was of its own hypothesis. (This remains unimple-mented because our recognizers [NC96] [RPP+90] rank utterance hypothesesonly relatively to each other). We make no particular claims for our moveevaluation component of course - only that our dialogue grammar enables aclear distinction between game strategy and game knowledge. Game strategy

choice could also be implemented through processing a logical representationof beliefs, desires and intentions if one desired.It may happen that no choice of weights for the existing preference func-tions can give consistently good choices when presented with various possiblemoves to make (or analyze). If so, it is probably the case that the preferencefunctions themselves do not accurately re ect the features of dialogue movesthat actually motivate move (or analysis) choice. One must then re�ne, rede-�ne or add to the existing preference functions themselves rather than simplyaltering their values.The most challenging part of such a recon�guration is undoubtedly de-ciding what the relevant features of dialogue moves are and providing analgorithm for calculating a numeric score indicating how much the featureapplies in any given instance. For example, the current \expected gamelength" preference function currently encodes a very simple ranking scheme:initiating any sub-game is likely to lengthen the game more than making anatomic move. That is, the only evidence cited to determine \expected gamelength" is the move type of the next move. Clearly, one might seek a moresophisticated measure - for example, using a corpus to calculate the averagegame length of games whose initial sequence of moves matches those of thecurrent game. Alternatively, one might decide that expected game length isnot a valid criterion at all. Perhaps, only overall dialogue length is relevantand the distinction between many games with fewer moves and fewer gameswith many moves is not important.6.2 Recon�gurable Game KnowledgeTheoretically, games and moves are a domain independent linguistic resourceand, ideally, once one knows what they are, no recon�guration of this moduleshould be necessary. Naturally in practice things are not so straightforward.First, the theory of conversational games as developed so far has tended toremain an architecture within which speci�c proposals can be made. Thereare some extant proposals (notably [KID92]) but no advocate of conversa-tional game theory would claim the existence of a comprehensive set of gamesand moves. Consequently, at least the addition of new games and moves isinevitable. Secondly, even on the \architectural view", the theory is some-what vague on what might count as an instantiation of the theory. Very littleis said on what games and moves actually are other than to say that theyare functionally de�ned. Presumably, not just any functional account will

be valid. Furthermore, in an account which makes appeal to the notions ofgame and move and the possibility of nesting games, one really wants propercriteria for discerning moves from games, for discerning the di�erent types ofmove and game, and for discerning the possible recursive structures withingames. Thirdly, this vagueness also tends to infect the existing proposals.Almost the only empirical study conducted within conversational game the-ory is that of [KID92] which measures agreement between naive and expertcoders on sorting utterances into move types. The criteria speci�ed by theexperts are quite diverse and often somewhat vague so that it is not di�cultto worry about the underlying basis. One recent critic [Car96] has askedwhether the agreement statistic actually measures how well the naive coderscan learn the experts' instructions rather than re ecting the existence of anaturally occurring classi�cation within the data. (Carletta also criticizesthe calculation of the measure itself, on other grounds).One primary objective in developing our own more formal account ofconversational games was precisely to provide a clearer and �rmer basis forthe discernment of games and moves. For example, given the de�nition of agame as a function from public commitments to public commitments, it oughtto prove possible to mark up a dialogue corpus with boundaries indicating atwhat point the conversants have become committed to certain propositions.One particular merit of our proposal in this regard is that dialogue analyzersare not being asked to guess the mental states of the conversants. They arebeing asked to state at what point a third party (such as themselves) couldreasonably hold the conversants to certain propositions. It is an empiricalmatter whether this could be done or not. However, it at least appears tobe more testable than determining games and moves on the basis of their\function". Similarly, the notion of a conversational move as a functionfrom and into \propositions under discussion" is also one which may provideclearer criteria for the discernment of move types and contents than thesort of criteria employed in [KID92]. For the discernment of game structurewithin games, the relevant criterion is whether the same issue remains underdiscussion throughout and whether a proposition already under discussionremains under discussion at the close of the current game.6.3 Recon�gurable Rational AgentsIn general, rational agent re-modelling is an open research topic. We have noparticular claims to make here other than that, since it is liable to be com-

putationally intensive, it is practically useful to reason at the higher abstractlevel of games rather than utterances. For instance, when a conversationfails, it may be impossible for systems to detect which particular utteranceis responsible - it often makes sense simply to deem the whole exchange afailure and start again.Conceptually, recon�guring the rational agent simply involves changingwhat the agent knows, what his goal is or what atomic actions he is capableof playing. Practically of course the task is quite signi�cant and a generalfacility with planning notions and the use of the CLE theorem-prover wouldbe essential in order to change the structure of the agent in non-trivial ways.7 ConclusionsWe have used an empirically motivated theory of dialogue and developedfor it a formal syntax and semantics which, we believe, re ects an intu-itive understanding of the central notions. This has enabled us to constructa reasonably straightforward implementation which suitably processes ourdeclarative dialogue de�nitions. In conjunction with a speech recognizer andthe CLE natural language understanding system, and using a collected dia-logue corpus, we have been able to implement a complete dialogue system.Our declarative dialogue de�nitions provide a clear framework for dialoguemanagement recon�guration.References[Car96] J. Carletta. Assessing agreement on classi�cation tasks: the kappastatistic. Computational Linguistics, 22 number 2:249{254, 1996.[CII+97] J. Carletta, A. Isard, S. Isard, J. Kowtko, G. Doherty-Sneddon,and A. Anderson. The reliability of a dialogue structure codingscheme. Computational Linguistics, forthcoming, 1997.[GS86] B.J. Grosz and C.L. Sidner. Attention, intentions and the struc-ture of discourse. Computational Linguistics, 12:175{204, 1986.[Hou86] G. Houghton. The Production of Language in Dialogue. PhDthesis, University of Sussex, 1986.

[KID92] J.C. Kowtko, S.D. Isard, and G.M. Doherty. Conversational gameswithin dialogue. HCRC research paper RP-31, 1992.[Lev83] S.C. Levinson. Pragmatics. Cambridge University Press, 1983.[MB92] R.K. Moore and S.R. Browning. Results of an exercise to collect`genuine' spoken enquiries using woz techniques. In Proceedingsof the Institute of Acoustics 14 6, pages 613{620, 1992.[NC96] Nuance-Communications. Nuance speech recognition system, ver-sion 5, developer's manual. Technical report, Nuance Communi-cations, Menlo Park, California, 1996.[Pow79] R. Power. The organization of purposeful dialogues. Linguistics,17:107{152, 1979.[RPP+90] M.J. Russell, K.M. Ponting, S.M. Peeling, S.R. Browning, J.S.Bridle, and R.K. Moore. The arm continuous speech recognitionsystem. In Proceedings of ICASSP'90 Albuquerque, New Mexico,1990.