Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität...

28
Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze

Transcript of Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität...

Page 1: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Information-State Dialogue Modelling in Several Versions

HS Dialogmanagement, SS 2002

Universität Saarbrücken

Michael Götze

Page 2: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Overview• GoDis

• Poesio & Traum Theory (PTT) and EDIS

• MIDAS

• SRI Autoroute Demonstrator

Question Under Discussion (QUDs)

Grounding & Obligations, Compositional DRT, ...

DRT, First-order-theorem-prover

Conversational Game Theorie

Page 3: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

What were the hopes connected to TRINDI ???

• easier implementation of non-trivial dialogue theories / rapid prototyping

• easier portation of systems to new domains• comparability of dialogue theories• make theories benefit from each other

(by having them implemented in one framework)

• etc. ...

Page 4: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Today

• Introduction• TRINDI: the „theory-neutral“ part – the framework• PTT and EDIS• Conversational Game Theory (CGT) and the

SRI Autoroute Demonstrator• Comparison: PTT/EDIS vs. SRI Autoroute Demonstrator• Conclusions?

Page 5: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.
Page 6: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

TRINDI: the „theory-neutral“ part

Information State (IS)

- the state of the dialogue -

its CONDITIONS can be checkedcan be changed with OPERATIONS

Update Moduleupdates the IS on the

basis of the input

Control Module

Selection Moduleselects the next system action

Page 7: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Questions for implementations

• What does the IS look like?

• How does updating work?

• How is the next action selected?

• How is all this controlled?

Page 8: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

PTT (Poesio & Traum Theory) and EDIS

• Agents perform Dialogue Acts (DAs)• Effects of DAs update IS• focus on the GROUNDING process (IS: private, grounded,

not grounded)• focus on social effects: OBLIGATIONS (to act) &

COMMITMENTS (to propositions) (vs. intentions & beliefs)

• orientation towards INCREMENTAL processing (below the utterance level)

• exploration of accessibility conditions of pragmatic processes: REFERENCE, SCOPING ( CDRT)

Page 9: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

IS in the PTT

• a new contribution results in a new discourse unit (DU) in the IS, containing obligations (OBL), the discourse history (DH), social commitments (SCP), conditions (COND) and its ID.

• UDUS: list of (still) ungrounded DUs• GND: grounded information• PDU: the previous DU• CDU: the current DU• INT: intentions

Page 10: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

act: ID: 2, accept(DP, ID2)

effect: accomplished via rule resolution

act: ID: 2, assert(DP, PROP)

effect: push(SCP, scp(DP, PROP)

effect: push(COND, accept(o(DP),ID)

scp(o(DP),PROP))

act: ID: 2, ack(DP, DU1)

effect: peRec(w,Gnd,w.pdu.tognd)

effect: remove(DU1,UDUS)

act: ID: 1, assert(DP, PROP)

effect: push(COND, accept(o(DP),ID)

scp(o(DP),PROP))

ID:2, agree(DP, ID2)

effect: push(SCP, scp(DP,P(ID2)))

act: ID: 2, check(DP, PROP)

effect: push(OBL, address(o(DP), ID)

effect: push(COND, agree(o(DP),ID)

scp(DP,PROP))

ID:2, answer(DP, ID2, ID3)

effect: push(SCP, ans(DP,Q(ID3), P(ID2)))

act: ID: 2, direct(DP, Act)

effect: push(OBL, address(o(DP), ID)

effect: push(COND, accept(o(DP),ID)

obl(o(DP),Act))

act: ID: 2, info_request(DP, Q)

effect: push(OBL, address(o(DP), ID))

Page 11: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

GND: OBL: [understandingAct(W,DU3), address(C,CA2)]DH: [CA3: C2, acknowledge(C,DU2), CA2: C2, info_request(W,?helpform)]SCP: []COND: []

UDUS: [DU3]

PDU: TOGND: OBL: [address(C,CA2]DH: [CA2: C2, info_request(W,?helpform)]SCP: []COND: []

ID: DU2

CDU: TOGND: OBL: [address(W,CA6] DH: [CA6: C2, direct(C,giveroute(W)),

CA5: C2, answer(C,CA2,CA4), CA4: C2, assert(C,want(C,route))]

SCP: [scp(C,want(C,route))]COND: [accept(W,CA6) obl(W,giveroute(W))]

ID: DU3

INT: [info_request(W,?start), giveroute(W), accept(W, CA6), acknowledge(W, DU3) ]

W: How can I help?

C: A route please.

Page 12: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Updating in PTT

1. Create a new DU and push it on top of UDUs.

2. Perform updates on the basis of backwards grounding acts.

3. If any other type of act is observed, record it in the dialogue history in CDU and apply the update rules for this kind of act.

4. Apply update rules to all parts of the IS which contain newly added acts.

Page 13: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Selection in PTT

• intentions lead to actions

• for choosing dialoge acts following factors are taken into account:– obligations– potential obligations (arising from COND)– insufficiently understood dialogue acts– intentions to perform complex acts

Page 14: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Controlling in PTT

• ????

Page 15: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Example(1) W: How can I help?(2) C: A route please.(3) W: Where would you like to start?(4) C: Malvern.(5) W: Great Malvern?(6) C: Yes.(7) W: Where would you like to go?(8) C: Edwinstowe.(9) W: Edwinstowe in Nottingham?(10) C: Yes.(11) W: When do you want to leave?(12) C: 6 pm.(13) W: Leaving at 6 pm?(14) C: Yes.(15) W: Do you want the quickest or the shortest route?(16) C: Quickest.(17) W: Please wait while your route is calculated.

Page 16: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Summary PTT

• ...

Page 17: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Conversational Game Theory (CGT)

• Power ´79, Houghton ´86, Carletta et al. ´97

• RATIONAL AGENTS plan to satisfy their GOALS by undertaking ACTIONS

• dialogues consist of exchanges between agents: CONVERSATIONAL GAMES, with mutually known and understood CONVERSATIONAL RULES

Page 18: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

CGT – agents with „split personality“

• Rational Agent– plans & executes conversational games (as

atomic actions)

• Game Player– plays the conversational games

Page 19: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Conversational Games

QW Game

qw

qw-r

rw ack

cnfRyes|Rno|Rmod

INF Game inf ack

PARDON Game unrec pdn

INTERRUPT Game unimp INF game

HELLO Game hello hello

Page 20: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

IS in CGT

Rational Agent:(plans & executes conversational

games (as atomic actions))

Game Player:(plays the conversational games)

PLAN: stack(actions)

SCOREBOARD: set(propn)

AGENDA: stack(possible_parses)

CURRTOKEN: index of current token

ALLTOKEN: stack(set(propn))

Page 21: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Updating in CGT

Rational Agent:- makePlan - generates a plan

- dropAction - if top goal is satisfied, remove it!

- undertakeAction - if top goal is not satisfied, generate an agenda item

Game Player:- 5 rules steering the network traversing

Page 22: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Selection in CGT

Rational Agent:- only one goal in the implementation: giving a route

Game Player:- 4 rules steering the selection of next moves

Page 23: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Controlling in CGT• Control algorithm:

1. Call the update module. (dialogue monitoring)2. Call the register-utterance module. (dialogue contribution generator)3. Repeat

• Update algorithm:1. Are there any update rules whose preconditions are fulfilled in the

current IS?2. If so, take the first one and execute the updates specified in the efects

of the rule and repeat.3. If not, stop.

• Register-utterance algorithm:1. If it is the system‘s turn to say something

a) call the selection moduleb) call the generator

2. If it is the user‘s turn to say something, call the input module.

Page 24: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Control Preferences: Updating & Selection

• Playing the game =? Parsing What kind of parsing???

• incremental & parallel parsing:– each possible parse is stored on the agenda

– which is the preferred one???• confidence (confirmation moves and games win)

• informativity (unrestricted questions win)

• shorter games! (simple acknowledgments win vs. confirmations)

• the one the user picks out

Page 25: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Summary CGT

• Game-based theory

• division of labour between rational agent and the game player

• monitoring vs. contributing

• control preferences

Page 26: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

What were the hopes connected to TRINDI ???

• easier implementation of non-trivial dialogue theories / rapid prototyping

• easier portation of systems to new domains• comparability of dialogue theories• make theories benefit from each other

(by having them implemented in one framework)

• etc. ...

Page 27: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

PTT & EDIS CGT

IS grounded vs. ungrounded information

rational agent vs. game player

UPDATE -‚dialogue act rules‘ - rules on a recursive-transition- network

SELECTION

CONTROL -incremental below the utterance level- serial?

- incremental & parallel

Page 28: Information-State Dialogue Modelling in Several Versions HS Dialogmanagement, SS 2002 Universität Saarbrücken Michael Götze.

Conclusion?

• easier implementation of non-trivial dialogue theories / rapid prototyping

• easier portation of systems to new domains• comparability of dialogue theories• make theories benefit from each other

(by having them implemented in one framework) ???

• etc. ...