Dialogue and Conversational Agents

Regina Barzilay
MIT

December, 2005

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 A travel dialog: Communicator

S1 Hello. You’ve reached the [Communicator.] Tell me your name

U2 Hi I’d like to fly to Seattle Tuesday morning
S3 Travelling to Seattle on Tuesday, August 11th in the morning.
U4 Yes.
S5 Your full name?
U6 John Doe

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Call routing: ATT HMIHY

S1 How may I help you?

U2 can you tell me how much it is to tokyo?
S3 You want to know the cost of a call?
U4 yes that’s right
S5 Please hold on for rate information

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 A tutorial dialogue: ITSPOKE

S1 What force acts on all objects within close proximity to earth?

U2 uh the force of gravity
S3 Fine. Besides the Earth’s gravitational force, are there any other forces?
U4 no

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Outline

• Analyzing Human Conversations

• Architecture of Dialogue Systems

– ASR
– NLU
– Generation
– Dialogue Manager

• Statistical model for the NLU component

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Analyzing Human Conversations

• Human data is used to inform design of

conversational systems
scheduling assistant
cross-language information access
...

• Computational questions:
– how to represent structural information in
dialogue?
– how to compute this representation?

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Speech Acts

• Austin (1962): An utterance is a kind of action

• Clear case: performatives

– I name this ship the Titanic
– I bet you five dollars it will snow tomorrow

• Austin’s idea: not just verbs

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 One utterance — three acts

• Locutionary act: the utterance of a sentence with a

particular meaning

• Illocutionary act: the act of asking, answering,

promising, etc., in uttering a sentence

• Perlocutionary act: the (often intentional)

production of certain effects upon the thoughts,
feelings, or actions of addressee in uttering a
sentence

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Example

You can’t do that!

• Locutionary force:
– Imperative

• Illocutionary force:
– Protesting

• Perlocutionary act:
– Intent to annoy addressee
– Intent to stop addresses from doing something

CuuDuongThanCong.com https://fb.com/tailieudientucntt
Five classes of Speech Acts (Searle, 1975)
• Assertives: committing the speaker to something’s being
the case (suggesting, putting forward, swearing, boasting)

• Directives: attempts by the speaker to get the addressee to

do something (asking, ordering, requesting)

• Commisives: committing the speaker to future course of

action (promising, planning, vowing, betting, opposing)

• Expressives: expressing the psychological state of the

speaker about a state of affairs (thanking, apologizing,
welcoming, deploring)

• Declarations: bringing about a different state of the world

via the utterance (I resign; You’re fired)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Dialogue Acts
• An act with associated structural information related to its
dialogue function

• Multiple classification schemes have been developed in

the past

• These schemes combine ideas from Searle, Austin and

others, but details may change from one domain to
another

• Verbmobil task
– Two-party scheduling dialogues
– Speakers were asked to plan a meeting at some future
date

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 DAMSL: forward looking func.

STATEMENT a claim made by the speaker

INFO-REQUEST a question by the speaker
CHECK a question for confirming information
INFLUENCE-ON-ADDRESSEE (Searle’s directives)
OPEN-OPTION a weak suggestion or listing of options
ACTION-DIRECTIVES on actual command
INFLUENCE-ON-SPEAKER (Austin’s commissives)
OFFER speaker offers to do something
COMMIT speaker is committed to doing something

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 DAMSL: backward looking func.

STATEMENT speaker’s response to previous proposal

ACCEPT accepting the proposal
ACCEPT-PART accepting some part of the proposal
MAYBE neither accepting nor rejecting the proposal
REJECT-PART rejecting some part of the proposal
REJECT rejecting the proposal
HOLD putting off response
ANSWER answering a question

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Example

assert I need to travel in May

infor-req, ack And, what day in May did you want to travel?
assert, answer OK uh I need to be there for a meeting that’s from the 1st
info-req, ack And you are flying into what city?
assert, answer Seattle

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Vermobil Dialogue Acts

THANK Thanks
GREET Hello Dan
GREET It’s me again
INIT I wanted to make an appointment with you
REQ-COMMENT How does that look?
SUGGEST June 13th through 17th
REJECT No, Friday I’m booked all day
REQ-SUGGEST What is a good day of the week for you?
GIVE-REASON Because I have meetings all afternoon

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Vermobil Dialogue Acts

Hello, Mrs. Klein, we should arrange an appointment for the meeting

GREET, INTRODUCE-NAME, INIT-DATE, SUGGEST-SUPPORT-DATE
Well, I suggest in January, between the 15th and the 19th
UPTAKE, SUGGEST-SUPPORT-DATE, REQUEST-SUPPORT-DATE
Oh, that is really inconvenient
UPTAKE, REJECT-DATE
...
Very good, that suits me too, I can make it
ACCEPT-DATE,ACCEPT-DATE,ACCEPT-DATE

CuuDuongThanCong.com https://fb.com/tailieudientucntt
Automatic Interpretation of Dialogue Acts

• Task: automatic identification of dialogue acts

– Given an utterance, decide whether it is a QUESTION,
STATEMENT, SUGGEST, or ACK

• Recognizing illocutionary force will be crucial to building

a dialogue agent

• Perhaps we can just look at the form of the utterance to

decide?

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Can we just use the surface syntactic
form?

• YES-NO-Q’s have auxiliary-before-subject syntax?

– Will breakfast be served on USAir 1555?

• STATEMENTs
– I don’t care about lunch

• COMMANDs have imperative syntax:

– Show me flights from Boston to NY on Monday night

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Surface form is a weak predictor

Can I have your coffee? Question Request

I want your coffee Declarative Request
Give me your coffee Imperative Request

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Dialogue Act Ambiguity

• Can you give me a list of the flights from Atlanta to

Boston?
This looks like an INFO-REQUEST
If so, the answer is “YES”
But really it’s a DIRECTIVE or REQUEST, a polite form
of:
Please give me a list of the flights

What looks like a QUESTION can be a REQUEST

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Indirect speech acts

• Utterances which use a surface statement to ask a

question

• Utterances which use a surface question to issue a request

CuuDuongThanCong.com https://fb.com/tailieudientucntt
Sequence modeling for DA interpretation
• Words and Collocations
Please or would you good cue for REQUEST
Are you . . . good cue for INFO-REQUEST

• Prosody
Rising pitch is a good cue for INFO-REQUEST
Loudness/stress can help distinguish
yeah/AGREEMENT from yeah/BACKCHANNEL

• Conversational Structure
– yeah following a proposal is probably AGREEMENT;
yeah following an INFORM probably a
BACKCHANNEL

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 HMM model for DA interpretation

• A dialogue is an HMM

• The hidden states are the dialogue acts

• The observation sequences are sentences

– Each observation is one sentence

• The observation likelihood model is a word N-gram

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 HMMs for DA interpretation

• Goal of HMM model:

to compute labeling of dialogue acts
D = d1 , d2 , . . . , dn that is most probable given
evidence E
� P (E|D)P (D)
D = argmaxD P (D|E) = argmaxD
P (E)
= argmaxD P (E|D)P (D)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 HMMs for DA interpretation

D� = argmaxD P (E|D)P (D)

• Let W be word sequence in sentence and F be prosodic

feature sequence

• Simplifying independence assumption:

P (E|D) = P (F |D)P (W |D)

• (What are the implications of this?)

D� = argmaxD P (F |D)P (W |D)P (D)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 HMMs for DA interpretation

D� = argmaxD P (F |D)P (W |D)P (D)

• P (D): probability of sequence of dialogue acts

• P (F |D): probability of prosodic sequence given one

dialogue act

• P (W |D): probability of word string in a sentence given

dialogue act

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Estimating P (D)
Markov assumption: each dialogue act depends only on
previous N (N = 3)
N

P (D) = P (di |di−1 , . . . , di−N +1 )

i=2

opening 0.23
0.76
0.99
suggest constraint
0.63
0.18
0.36 0.19
accept reject
0.18 0.22

0.46
closing

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Estimating P (W |D)

• Each dialogue act has different words

– Questions have are you, do you, etc.
N

P (W |D) = P (wi |wi−1 , . . . , wi−N +1 , di )

i=2

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Estimating P (F |D)

• A classifier (decision tree) trained on simple

acoustically-based prosodic features:
– Average energy at different places in utterance
– Various duration measures

• Prosody allows us to distinguish between various DAs:

– Statement
– Yes-No-Question
– Declarative-Question

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Estimating P (F |D)

• Classifier give posterior p(D|F )

• We need p(F |d) to fit into HMM

p(F |d)p(d)
p(d|F ) =
p(F )

• Rearranging terms to get a likelihood:

p(F |d) P (d|F )
=
p(F ) p(d)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Final HMM equation for DA interpretation

�
D = argmaxD P (F |D)P (W |D)P (D)

M N N
P (di |F )
P (di |di−1 , . . . , di−M +1 ) P (wi |wi−1 , . . . , wi−N +1 , di )
p(di )
i=2 i=2 i=2

• We can use Viterbi decoding to find D �

• In real dialogue systems, obviously can’t use future

dialogue acts, so predict up to current act

• In rescoring passes (for example for labeling

human-human dialogues for meeting summarization), can
use future info.

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Outline

• Analyzing Human Conversations

• Architecture of Dialogue Systems

– ASR
– NLU
– Generation
– Dialogue Manager

• Statistical model for the NLU component

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Dialogue System Architecture

Speech Natural Language

Recognition Understanding

Dialogue Task
Manager Manager
Text−to−Speech Natural Language
Synthesis Generation

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Automatic Speech Recognition engine

• Based on a standard ASR engine

– Maps speech to words

• Has specific characteristics for dialogue

– Language model could depend on where we are in the
dialogue
– Could make use of the fact that we are talking to the
same human over time (speaker adaptation)
– Confidence values (we want to know if the system
misunderstood the human)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 LM for Dialogue Systems

• Language models for dialogue are often based on

hand-written Context-Free or finite-state grammars rather
than N-grams

• We can have LM specific to a dialogue state

– If system just asked “What city are you departing
from?”
– LM should predict:
� City names only
� FSA: (I want to (leave|depart)) (from) [CITYNAME]
� N-grams trained on answers to “Cityname”
questions from labeled data

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Natural Language Understanding
There are many ways to represent the meaning of sentences
For speech dialogue systems, most common is “Frame and slot
semantics”
Show me morning flights from Boston to NY on Toesday

SHOW:

FLIGHTS:

ORIGIN

CITY: Boston

DATE: Tuesday

TIME: morning

DEST

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 How to generate this semantics
• Design a semantic grammar for a domain
LIST � show me | I want | can I see |
DEPARTTIME� (after|before|around) HOUR | morning | evening
HOUR � one | . . . | twelve | (am|pm)
FLIGHTS � a (flight) — flights
ORIGIN � from CITY
DESTINATION � to CITY
CITY � Boston | San Francisco

• Use a parser to map a sentence into a semantic

representation (we will see an example of statistical
mapping later in the lecture)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
Generation and Text-to-Speech Synthesis

• Generation component
– Chooses concepts to express to user
– Plans how to express these concepts in words
– Assigns any necessary prosody to the words

• TTS component
– Takes words and prosodic annotations
– Synthesizes a waveform

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Generation Component

• Chooses syntactic structures and words to express

semantic predicates (provided by dialogue manager)

• Typically implemented using template-based method

– all concepts are associated with corresponding
templates
– each template has variables instantiated during the
generation process
What time do you want to leave CITY-ORIG?
– LM scores are used to select among alternatives

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Dialogue Manager

• Takes input from ASR/NLU components

• Maintains some sort of state

• Interfaces with Task Manager

• Passes output to NLG/TTS modules

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Dialogue Manager
• Finite State
– Single-initiative: system completely controls the
conversation with the user
– Implementation: cascade of FSAs

• Frame-based
– Mixed-initiatives:
� system asks questions of user, filling any slots that user specifies
� when frame is filled, do database query

– Implementation: production rules that switch control

among various frames

• Planning Agents (next time)

• Markov Decision Processes (next time)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Finite State Dialogue Manager
What city are you leaving from?

Where are you going?

What date do you want to leave?

Is it a one−way trip?
YES NO
Do you want to go from What date do you want to return?
<FROM> to <TO> on <DATE>?
YES
NO Do you want to go from
<FROM> to <TO> on <DATE>?
returning on <RETURN>?
YES
[Book the flight]
NO

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Outline

• Analyzing Human Conversations

• Architecture of Dialogue Systems

– ASR
– NLU
– Generation
– Dialogue Manager

• Statistical model for the NLU component

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Statistical NLU component

• A fully statistical approach to natural language interfaces

• Task: map a sentence + context to a database query

User: Show me flights from NY to Boston, leaving tomorrow

System: [returns a list of flights]

Show: (Arrival-time)
Origin (City ”NY“)
Destination: (City ”Boston”)
Date: (November 27th, 2003)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Representation

• W=input sentence

• H=history (some representation of previous sentences)

• T=a parse tree for W

• F,S=a context-independent semantic representation for W

• M=a context-dependent representation for W (M depends

on both F, S and H)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Example
W = input sentence; H = history; T = a parse tree for W; F, S = a
context independent semantic representation for W; M = a
context-dependent semantic representation for W

User: Show me flights from Newark or New York to Atlanta, leaving

tomorrow
System: returns a list of flights
User: When do the flights that leave from Newark arrive in Atlanta
W = When do the flights that leave from Newark arrive in Atlanta
Show: (flights)
Origin (City ”NY“) or (City ”NY“)
H=
Destination: (City ”Atlanta”)
Date: (November 27th, 2003)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Example
W = input sentence; H = history; T = a parse tree for W; F, S = a
context independent semantic representation for W; M = a
context-dependent semantic representation for W

User: Show me flights from Newark or New York to Atlanta, leaving

tomorrow
System: returns a list of flights
User: When do the flights that leave from Newark arrive in Atlanta
W = When do the flights that leave from Newark arrive in Atlanta
Show: (Arrival-time)
F,S= Origin (City “Newark”)
Destination: (City ”Atlanta”)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Example
Show: (flights)
Origin (City ”NY“) or (City ”NY“)
H=
Destination: (City ”Atlanta”)
Date: (November 27th, 2003)

Show: (Arrival-time)
F,S= Origin (City “Newark”)
Destination: (City ”Atlanta”)

Show: (Arrival-time)
Origin (City “Newark”)
M=
Destination: (City ”Atlanta”)
Date: (November 27th, 2003)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 A Parse Tree

Each non-terminal has a syntactic and semantic tag,

e.g., city/npr
/top

/wh−question

/aux flight/np arrival/vp

time
/wh−head
do arrival location
When /vp−head /pp
location city
/prep /npr

in Atlanta

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Building a Probabilistic Model
• Basic goal: build a model of P (M |W, H) – probability of a
context-dependent interpretation, given a sentence and a
history
• We’ll do by building a model of P (M, W, F, T, S|H),
giving
�
P (M, W |H) = P (M, W, F, T, S|H)
F,T ,S

and

argmaxM P (M |W, H) = argmaxM P (M, W |H)

�
= argmaxM P (M, W, F, T, S|H)
F,T ,S

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Building a Probabilistic Model

Our aim is to estimate P (M, W, F, T, S|H)

• Apply Chain rule:

P (M, W, F, T , S|H) = P (F |H)P (T , W |F, H)P (S|T , W, F, H)P (M |S, T , W, F, H)

• Independence assumption:
P (M, W, F, T , S|H) = P (F )P (T , W |F )P (S|T , W, F ) × P (M |S, F, H)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Building a Probabilistic Model

P (M, W, F, T, S|H) = P (F )P (T, W |F )P (S|T, W, F )×P (M |S, F, H)

• The sentence processing model is a model of

P (T, W, F, S). Maps W to (F, S, T ) triple (a
context-independent interpretation)

• The contextual processing model goes from a (F, S, H)

triple to a final interpretation, M

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Example
Show: (flights)
Origin (City ”NY“) or (City ”NY“)
H=
Destination: (City ”Atlanta”)
Date: (November 27th, 2003)

Show: (Arrival-time)
F,S= Origin (City “Newark”)
Destination: (City ”Atlanta”)

Show: (Arrival-time)
Origin (City “Newark”)
M=
Destination: (City ”Atlanta”)
Date: (November 27th, 2003)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Building a Probabilistic Model

P (T, W, F, S) = P (F )P (T, W |F )P (S|T, W, F )

• First step: choose the frame F with probability P (F )

Show: (Arrival-time)
Origin
Destination:

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 The Sentence Processing Model

P (T, W, F, S) = P (F )P (T, W |F )P (S|T, W, F )

• Next step: generate the parse tree T and sentence W

• Method uses a probabilistic context-free grammar, where

markov processes are used to generate rules. Different
rule parameters are used for each value of F

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 The Sentence Processing Model
flight
/np

/det flight flight−constraint

/corenp /rel−clause

P(/det flight/corenp flight−constraints/rel−clause|flight/np)

= P(/det|NULL, flight/np) *P(flight/corenp|/det,flight/np)
* P(flight−constraints|relclause|flight/corenp,flight/np)
* P(STOP|flight−constraints/relclause,flight/np)

• Use maximum likelihood estimation

Count(corenp, np)
PM L (corenp|np) =
Count(np)

• Backed-off estimates generate semantic, syntactic parts of

each label separately

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 The Sentence Processing Model

• Given a frame F , and a tree T , fill in the semantic slots S

Show: (Arrival-time)
Origin
Destination:
Show: (Arrival-time)
Origin Newark
Destination: Atlanta

• Method works by considering each node of the parse tree

T, and applying probabilities P (slot-fill-action|S,node)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 The Sentence Processing Model: Search

P (T, W, F, S) = P (F )P (T, W |F )P (S|T, W, F )

• Goal: produce n high probability (F, S, T, W ) tuples

• Method:
– In first pass, produce n-best parses under a parsing model
that is independent of F
– For each tree T , for each possible frame F , create a
(W, T, F ) triple with probability P (T, W, |F ). Keep the top n
most probable triples.
– For each triple, use beam search to generate several high
probability (W, T, F, S) tuples. Keep the top n most
probable.

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 The Contextual Model
Show: (flights)
Origin (City ”NY“) or (City ”NY“)
H=
Destination: (City ”Atlanta”)
Date: (November 27th, 2003)

Show: (Arrival-time)
F,S= Origin (City “Newark”)
Destination: (City ”Atlanta”)

Show: (Arrival-time)
Origin (City “Newark”)
M=
Destination: (City ”Atlanta”)
Date: (November 27th, 2003)

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 The Contextual Model

• Only issue is whether values in H, but not in (F, S), should be

carried over to M.
Show: (Arrival-time)
Origin (City “Newark”)
M=
Destination: (City ”Atlanta”)
Date: (November 27th, 2003)

• Method uses a decision-tree model to estimate probability of

”carrying over” each slot in H which is not in F, S.

CuuDuongThanCong.com https://fb.com/tailieudientucntt
 Summary

• HMM model for DA labeling

• Architecture of Dialogue Systems

• Statistical model for the NLU component

• Next time:
– Planning Agents
– Markov Decision Processes for Dialogue
Management

CuuDuongThanCong.com https://fb.com/tailieudientucntt