Spoken Dialogue Systems Julia Hirschberg CS 4706 2272021

Spoken Dialogue Systems Julia Hirschberg CS 4706 2/27/2021 1

Today • Basic Conversational Agents – – ASR NLU Generation Dialogue Manager • Dialogue Manager Design – Finite State – Frame-based – Initiative: User, System, Mixed • Information-State – – Dialogue-Act Detection Dialogue-Act Generation Evaluation Utility-based conversational agents • MDP, POMDP 2/27/2021 2

Conversational Agents • AKA: – Interactive Voice Response Systems – Dialogue Systems – Spoken Dialogue Systems • Applications: – – – 2/27/2021 Travel arrangements (Amtrak, United airlines) Telephone call routing Tutoring Communicating with robots Anything with limited screen/keyboard 3

A travel dialog: Communicator 2/27/2021 4

Call routing: ATT HMIHY 2/27/2021 5

A tutorial dialogue: ITSPOKE 2/27/2021 6

Conversational Structure • Telephone conversations – – Stage 1: Enter a conversation Stage 2: Identification Stage 3: Establish joint willingness to converse Stage 4: First topic is raised, usually by caller 2/27/2021 7

Why is this customer confused? • Customer: (rings) • Operator: Directory Enquiries, for which town please? • Customer: Could you give me the phone number of um: Mrs. um: Smithson? • Operator: Yes, which town is this at please? • Customer: Huddleston. • Operator: Yes. And the name again? • Customer: Mrs. Smithson 2/27/2021 8

Why is this customer confused? • A: And, what day in May did you want to travel? • C: OK, uh, I need to be there for a meeting that’s from the 12 th to the 15 th. • Note that client did not answer question. • Meaning of client’s sentence: – Meeting • Start-of-meeting: 12 th • End-of-meeting: 15 th – Doesn’t say anything about flying!!!!! • How does agent infer client is informing him/her of travel dates? 2/27/2021 9

Will this client be confused? A: … there’s 3 non-stops today. – True if in fact 7 non-stops today. – But agent means: 3 and only 3. – How can client infer that agent means: • only 3 2/27/2021 10

Grice: conversational implicature • Implicature means a particular class of licensed inferences. • Grice (1975) proposed that what enables hearers to draw correct inferences is: • Cooperative Principle – This is a tacit agreement by speakers and listeners to cooperate in communication 2/27/2021 11

4 Gricean Maxims • Relevance: Be relevant • Quantity: Do not make your contribution more or less informative than required • Quality: try to make your contribution one that is true (don’t say things that are false or for which you lack adequate evidence) • Manner: Avoid ambiguity and obscurity; be brief and orderly 2/27/2021 12

Relevance • A: Is Regina here? • B: Her car is outside. • Implication: yes – Hearer thinks: why would he mention the car? It must be relevant. How could it be relevant? It could since if her car is here she is probably here. • Client: I need to be there for a meeting that’s from the 12 th to the 15 th – Hearer thinks: Speaker is following maxims, would only have mentioned meeting if it was relevant. How could meeting be relevant? If client meant me to understand that he had to depart in time for the mtg. 2/27/2021 13

Quantity • A: How much money do you have on you? • B: I have 5 dollars – Implication: not 6 dollars • Similarly, 3 non stops can’t mean 7 non-stops (hearer thinks: – if speaker meant 7 non-stops she would have said 7 non-stops • A: Did you do the reading for today’s class? • B: I intended to – Implication: No – B’s answer would be true if B intended to do the reading AND did the reading, but would then violate maxim 2/27/2021 14

Dialogue System Architecture 2/27/2021 15

Speech recognition • Input: acoustic waveform • Output: string of words – Basic components: • a recognizer for phones, small sound units like [k] or [ae]. • a pronunciation dictionary like cat = [k ae t] • a grammar telling us what words are likely to follow what words • A search algorithm to find the best string of words 2/27/2021 16

Natural Language Understanding • Or “NLU” • Or “Computational semantics” • There are many ways to represent the meaning of sentences • For speech dialogue systems, most common is “Frame and slot semantics”. 2/27/2021 17

An example of a frame • Show me morning flights from Boston to SF on Tuesday. SHOW: FLIGHTS: ORIGIN: CITY: Boston DATE: Tuesday TIME: morning DEST: CITY: San Francisco 2/27/2021 18

Semantics for a sentence LIST FLIGHTS ORIGIN Show me flights from Boston DESTINATION DEPARTDATE to San Francisco on Tuesday DEPARTTIME morning 2/27/2021 20

Generation and TTS • Generation component – Chooses concepts to express to user – Plans out how to express these concepts in words – Assigns any necessary prosody to the words • TTS component – Takes words and prosodic annotations – Synthesizes a waveform 2/27/2021 21

Generation Component • Content Planner – Decides what content to express to user • (ask a question, present an answer, etc) – Often merged with dialogue manager • Language Generation – Chooses syntactic structures and words to express meaning. – Simplest method • All words in sentence are prespecified! • “Template-based generation” • Can have variables: – What time do you want to leave CITY-ORIG? – Will you return to CITY-ORIG from CITY-DEST? 2/27/2021 22

More sophisticated language generation component • Natural Language Generation • Approach: – Dialogue manager builds representation of meaning of utterance to be expressed – Passes this to a “generator” – Generators have three components • Sentence planner • Surface realizer • Prosody assigner 2/27/2021 23

Architecture of a generator for a dialogue system (after Walker and Rambow 2002) 2/27/2021 24

HCI constraints on generation for dialogue: “Coherence” • Discourse markers and pronouns (“Coherence”): (1) Please say the date. … Please say the start time. … Please say the duration… … Please say the subject… (2) First, tell me the date. … Next, I’ll need the time it starts. … Thanks. <pause> Now, how long is it supposed to last? … Last of all, I just need a brief description 2/27/2021 25

HCI constraints on generation for dialogue: coherence (II): tapered prompts • • • Prompts which get incrementally shorter: System: Now, what’s the first company to add to your watch list? Caller: Cisco System: What’s the next company name? (Or, you can say, “Finished”) Caller: IBM System: Tell me the next company name, or say, “Finished. ” Caller: Intel System: Next one? Caller: America Online. System: Next? Caller: … 2/27/2021 26

Dialogue Manager • Controls the architecture and structure of dialogue – – Takes input from ASR/NLU components Maintains some sort of state Interfaces with Task Manager Passes output to NLG/TTS modules 2/27/2021 27

Four architectures for dialogue management • Finite State • Frame-based • Information State – Markov Decision Processes • AI Planning 2/27/2021 28

Finite-State Dialogue Management • Consider a trivial airline travel system – – Ask the user for a departure city For a destination city For a time Whether the trip is round-trip or not 2/27/2021 29

Finite State Dialogue Manager 2/27/2021 30

Finite-state Dialogue Managers • System completely controls the conversation with the user • Asks the user a series of questions • Ignores (or misinterprets) anything the user says that is not a direct answer to the system’s questions 2/27/2021 31

Dialogue Initiative • Systems that control conversation like this are system initiative or single initiative. • “Initiative”: who has control of conversation • In normal human-human dialogue, initiative shifts back and forth between participants. 2/27/2021 32

System Initiative • Systems which completely control the conversation at all times are called system initiative. • Advantages: – Simple to build – User always knows what they can say next – System always knows what user can say next • Known words: Better performance from ASR • Known topic: Better performance from NLU – Ok for VERY simple tasks (entering a credit card, or login name and password) • Disadvantage: – Too limited 2/27/2021 33

User Initiative • User directs the system • Generally, user asks a single question, system answers • System can’t ask questions back, engage in clarification dialogue, confirmation dialogue • Used for simple database queries • User asks question, system gives answer • Web search is user initiative dialogue. 2/27/2021 34

Problems with System Initiative • Real dialogue involves give and take! • In travel planning, users might want to say something that is not the direct answer to the question. • For example answering more than one question in a sentence: – Hi, I’d like to fly from Seattle Tuesday morning – I want a flight from Milwaukee to Orlando one way leaving after 5 p. m. on Wednesday. 2/27/2021 35

Single initiative + universals • We can give users a little more flexibility by adding universal commands • Universals: commands you can say anywhere • As if we augmented every state of FSA with these – Help – Start over – Correct • This describes many implemented systems • But still doesn’t allow user to say what the want to say 2/27/2021 36

Mixed Initiative • Conversational initiative can shift between system and user • Simplest kind of mixed initiative: use the structure of the frame itself to guide dialogue – Slot Question – – – What city are you leaving from? Where are you going? What day would you like to leave? What time would you like to leave? What is your preferred airline? ORIGIN DEST DEPT DATE DEPT TIME AIRLINE 2/27/2021 37

Frames are mixed-initiative • User can answer multiple questions at once. • System asks questions of user, filling any slots that user specifies • When frame is filled, do database query • If user answers 3 questions at once, system has to fill slots and not ask these questions again! • Anyhow, we avoid the strict constraints on order of the finite-state architecture. 2/27/2021 38

Multiple frames • flights, hotels, rental cars • Flight legs: Each flight can have multiple legs, which might need to be discussed separately • Presenting the flights (If there are multiple flights meeting users constraints) – It has slots like 1 ST_FLIGHT or 2 ND_FLIGHT so user can ask “how much is the second one” • General route information: – Which airlines fly from Boston to San Francisco • Airfare practices: – Do I have to stay over Saturday to get a decent airfare? 2/27/2021 39

Multiple Frames • Need to be able to switch from frame to frame • Based on what user says. • Disambiguate which slot of which frame an input is supposed to fill, then switch dialogue control to that frame. • Main implementation: production rules – Different types of inputs cause different productions to fire – Each of which can flexibly fill in different frames – Can also switch control to different frame 2/27/2021 40

Defining Mixed Initiative • Mixed Initiative could mean – User can arbitrarily take or give up initiative in various ways • This is really only possible in very complex plan-based dialogue systems • No commercial implementations • Important research area – Something simpler and quite specific which we will define in the next few slides 2/27/2021 41

True Mixed Initiative 2/27/2021 42

How mixed initiative is usually defined • First we need to define two other factors • Open prompts vs. directive prompts • Restrictive versus non-restrictive grammar 2/27/2021 43

Open vs. Directive Prompts • Open prompt – System gives user very few constraints – User can respond how they please: – “How may I help you? ” “How may I direct your call? ” • Directive prompt – Explicit instructs user how to respond – “Say yes if you accept the call; otherwise, say no” 2/27/2021 44

Restrictive vs. Non-restrictive grammars • Restrictive grammar – Language model which strongly constrains the ASR system, based on dialogue state • Non-restrictive grammar – Open language model which is not restricted to a particular dialogue state 2/27/2021 45

Definition of Mixed Initiative Grammar Open Prompt Restrictive Doesn’t make sense System Initiative Non-restrictive User Initiative 2/27/2021 Directive Prompt Mixed Initiative 46

Voice. XML • • Voice e. Xtensible Markup Language An XML-based dialogue design language Makes use of ASR and TTS Deals well with simple, frame-based mixed initiative dialogue. • Most common in commercial world (too limited for research systems) • But useful to get a handle on the concepts. 2/27/2021 47

Voice XML • Each dialogue is a <form>. (Form is the Voice. XML word for frame) • Each <form> generally consists of a sequence of <field>s, with other commands 2/27/2021 48

Sample vxml doc <form> <field name="transporttype"> <prompt> Please choose airline, hotel, or rental car. </prompt> <grammar type="application/x=nuance-gsl"> [airline hotel "rental car"] </grammar> </field> <block> <prompt> You have chosen <value expr="transporttype">. </prompt> </block> </form> 2/27/2021 49

Voice. XML interpreter • • Walks through a VXML form in document order Iteratively selecting each item If multiple fields, visit each one in order. Special commands for events 2/27/2021 50

Another vxml doc (1) <noinput> I'm sorry, I didn't hear you. <reprompt/> </noinput> - “noinput” means silence exceeds a timeout threshold <nomatch> I'm sorry, I didn't understand that. <reprompt/> </nomatch> - “nomatch” means confidence value for utterance is too low - notice “reprompt” command 2/27/2021 51

Another vxml doc (2) <form> <block> Welcome to the air travel consultant. </block> <field name="origin"> <prompt> Which city do you want to leave from? </prompt> <grammar type="application/x=nuance-gsl"> [(san francisco) denver (new york) barcelona] </grammar> <filled> <prompt> OK, from <value expr="origin"> </prompt> </filled> </field> - “filled” tag is executed by interpreter as soon as field filled by user 2/27/2021 52

Another vxml doc (3) <field name="destination"> <prompt> And which city do you want to go to? <grammar type="application/x=nuance-gsl"> [(san francisco) denver (new york) barcelona] </grammar> <filled> <prompt> OK, to <value expr="destination"> </filled> </field> <field name="departdate" type="date"> <prompt> And what date do you want to leave? <filled> <prompt> OK, on <value expr="departdate"> </filled> </field> 2/27/2021 </prompt> 53

Another vxml doc (4) <block> <prompt> OK, I have you are departing from <value expr="origin”> to <value expr="destination”> on <value expr="departdate"> </prompt> send the info to book a flight. . . </block> </form> 2/27/2021 54

Summary: Voice. XML • • Voice e. Xtensible Markup Language An XML-based dialogue design language Makes use of ASR and TTS Deals well with simple, frame-based mixed initiative dialogue. • Most common in commercial world (too limited for research systems) • But useful to get a handle on the concepts. 2/27/2021 55

Information-State and Dialogue Acts • If we want a dialogue system to be more than just formfilling • Needs to: – Decide when the user has asked a question, made a proposal, rejected a suggestion – Ground a user’s utterance, ask clarification questions, suggestion plans • Suggests: – Conversational agent needs sophisticated models of interpretation and generation • In terms of speech acts and grounding • Needs more sophisticated representation of dialogue context than just a list of slots 2/27/2021 56

Information-state architecture • • Information state Dialogue act interpreter Dialogue act generator Set of update rules – Update dialogue state as acts are interpreted – Generate dialogue acts • Control structure to select which update rules to apply 2/27/2021 57

Information-state 2/27/2021 58

Dialogue acts • Also called “conversational moves” • An act with (internal) structure related specifically to its dialogue function • Incorporates ideas of grounding • Incorporates other dialogue and conversational functions that Austin and Searle didn’t seem interested in 2/27/2021 59

Verbmobil task • Two-party scheduling dialogues • Speakers were asked to plan a meeting at some future date • Data used to design conversational agents which would help with this task • (cross-language, translating, scheduling assistant) 2/27/2021 60

Verbmobil Dialogue Acts THANK GREET INTRODUCE BYE REQUEST-COMMENT SUGGEST REJECT ACCEPT REQUEST-SUGGEST INIT GIVE_REASON FEEDBACK DELIBERATE CONFIRM CLARIFY 2/27/2021 thanks Hello Dan It’s me again Allright, bye How does that look? June 13 th through 17 th No, Friday I’m booked all day Saturday sounds fine What is a good day of the week for you? I wanted to make an appointment with you Because I have meetings all afternoon Okay Let me check my calendar here Okay, that would be wonderful Okay, do you mean Tuesday the 23 rd? 61

Automatic Interpretation of Dialogue Acts • How do we automatically identify 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? 2/27/2021 62

Can we just use the surface syntactic form? • YES-NO-Q’s have auxiliary-before-subject syntax: – Will breakfast be served on USAir 1557? • STATEMENTs have declarative syntax: – I don’t care about lunch • COMMAND’s have imperative syntax: – Show me flights from Milwaukee to Orlando on Thursday night 2/27/2021 63

Surface form != speech act type Locutionary Force Illocutionary Force Can I have the rest of your sandwich? Question Request I want the rest of your sandwich Declarative Request Give me your sandwich! Imperative Request 2/27/2021 64

Dialogue act disambiguation is hard! Who’s on First? Abbott: Well, Costello, I'm going to New York with you. Bucky Harris the Yankee's manager gave me a job as coach for as long as you're on the team. Costello: Look Abbott, if you're the coach, you must know all the players. Abbott: I certainly do. Costello: Well you know I've never met the guys. So you'll have to tell me their names, and then I'll know who's playing on the team. Abbott: Oh, I'll tell you their names, but you know it seems to me they give these ball players now-a-days very peculiar names. Costello: You mean funny names? Abbott: Strange names, pet names. . . like Dizzy Dean. . . Costello: His brother Daffy Abbott: Daffy Dean. . . Costello: And their French cousin. Abbott: French? Costello: Goofe' Abbott: Goofe' Dean. Well, let's see, we have on the bags, Who's on first, What's on second, I Don't Know is on third. . . Costello: That's what I want to find out. Abbott: I say Who's on first, What's on second, I Don't Know's on third. 2/27/2021 65

Dialogue act ambiguity • Who’s on first? – INFO-REQUEST – or – STATEMENT 2/27/2021 66

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 2/27/2021 67

Dialogue Act ambiguity • Similarly, what looks like a STATEMENT can be a QUESTION: Us OPEN- I was wanting to make some arrangements OPTION for a trip that I’m going to be taking uh to LA uh beginnning of the week after next A HOLD OK uh let me pull up your profile and I’ll be g right with you here. [pause] A CHECK And you said you wanted to travel next week? g Us 2/27/2021 ACCEP Uh yes. 68 T

Indirect speech acts • Utterances which use a surface statement to ask a question • Utterances which use a surface question to issue a request 2/27/2021 69

DA interpretation as statistical classification • Lots of clues in each sentence that can tell us which DA it is: • 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 2/27/2021 70

Statistical classifier model of dialogue act interpretation • Our goal is to decide for each sentence what dialogue act it is • This is a classification task (we are making a 1 -of-N classification decision for each sentence) • With N classes (= number of dialog acts). • Three probabilistic models corresponding to the 3 kinds of cues from the input sentence. – Conversational Structure: Probability of one dialogue act following another P(Answer|Question) – Words and Syntax: Probability of a sequence of words given a dialogue act: P(“do you” | Question) – Prosody: probability of prosodic features given a dialogue act : 2/27/2021 P(“rise at end of sentence” | Question) 71

An example of dialogue act detection: Correction Detection • Despite all these clever confirmation/rejection strategies, dialogue systems still make mistakes (Surprise!) • If system misrecognizes an utterance, and either – Rejects – Via confirmation, displays its misunderstanding • Then user has a chance to make a correction – Repeat themselves – Rephrasing – Saying “no” to the confirmation question. 2/27/2021 72

Corrections • Unfortunately, corrections are harder to recognize than normal sentences! – Swerts et al (2000): corrections misrecognized twice as often (in terms of WER) as non-corrections!!! – Why? • Prosody seems to be largest factor: hyperarticulation • English Example from Liz Shriberg – “NO, I am DE-PAR-TING from Jacksonville) • A German example from Bettina Braun from a talking elevator 2/27/2021 73

A Labeled dialogue (Swerts et al) 2/27/2021 74

Machine Learning and Classifiers • Given a labeled training set • We can build a classifier to label observations into classes – Decision Tree – Regression – SVM • I won’t introduce the algorithms here. • But these are at the core of NLP/computational linguistics/Speech/Dialogue • You can learn them in: – AI - CS 121/221 2/27/2021 – Machine Learning CS 229 75

Machine learning to detect user corrections • Build classifiers using features like – Lexical information (words “no”, “correction”, “I don’t”, swear words) – Prosodic features (various increases in F 0 range, pause duration, and word duration that correlation with hyperarticulation) – Length – ASR confidence – LM probability – Various dialogue features (repetition) 2/27/2021 76

Generating Dialogue Acts • Confirmation • Rejection 2/27/2021 77

Confirmation • Another reason for grounding • Errors: Speech is a pretty errorful channel – Even for humans; so they use grounding to confirm that they heard correctly • ASR is way worse than humans! • So dialogue systems need to do even more grounding and confirmation than humans 2/27/2021 78

Explicit confirmation • • S: Which city do you want to leave from? U: Baltimore S: Do you want to leave from Baltimore? U: Yes 2/27/2021 79

Explicit confirmation • U: I’d like to fly from Denver Colorado to New York City on September 21 st in the morning on United Airlines • S: Let’s see then. I have you going from Denver Colorado to New York on September 21 st. Is that correct? • U: Yes 2/27/2021 80

Implicit confirmation: display • U: I’d like to travel to Berlin • S: When do you want to travel to Berlin? • U: Hi I’d like to fly to Seattle Tuesday morning • S: Traveling to Seattle on Tuesday, August eleventh in the morning. Your name? 2/27/2021 81

Implicit vs. Explicit • Complementary strengths • Explicit: easier for users to correct systems’s mistakes (can just say “no”) • But explicit is cumbersome and long • Implicit: much more natural, quicker, simpler (if system guesses right). 2/27/2021 82

Implicit and Explicit • Early systems: all-implicit or all-explicit • Modern systems: adaptive • How to decide? – ASR system can give confidence metric. – This expresses how convinced system is of its transcription of the speech – If high confidence, use implicit confirmation – If low confidence, use explicit confirmation 2/27/2021 83

Computing confidence • Simplest: use acoustic log-likelihood of user’s utterance • More features – Prosodic: utterances with longer pauses, F 0 excursions, longer durations – Backoff: did we have to backoff in the LM? – Cost of an error: Explicit confirmation before moving money or booking flights 2/27/2021 84

Rejection • e. g. , Voice. XML “nomatch” • “I’m sorry, I didn’t understand that. ” • Reject when: – ASR confidence is low – Best interpretation is semantically ill-formed • Might have four-tiered level of confidence: – – Below confidence threshhold, reject Above threshold, explicit confirmation If even higher, implicit confirmation Even higher, no confirmation 2/27/2021 85

Dialogue System Evaluation • Key point about SLP. • Whenever we design a new algorithm or build a new application, need to evaluate it • Two kinds of evaluation – Extrinsic: embedded in some external task – Intrinsic: some sort of more local evaluation. • How to evaluate a dialogue system? • What constitutes success or failure for a dialogue system? 2/27/2021 86

Dialogue System Evaluation • It turns out we’ll need an evaluation metric for two reasons – 1) the normal reason: we need a metric to help us compare different implementations • can’t improve it if we don’t know where it fails • Can’t decide between two algorithms without a goodness metric – 2) a new reason: we will need a metric for “how good a dialogue went” as an input to reinforcement learning: • automatically improve our conversational agent performance via learning 2/27/2021 87

Evaluating Dialogue Systems • PARADISE framework (Walker et al ’ 00) • “Performance” of a dialogue system is affected both by what gets accomplished by the user and the dialogue agent and how it gets accomplished Maximize Task Success Minimize Costs Efficiency Measures Qualitative Measures 2/27/2021 88 Slide from Julia Hirschberg

PARADISE evaluation again: • Maximize Task Success • Minimize Costs – Efficiency Measures – Quality Measures • PARADISE (PARAdigm for Dialogue System Evaluation) 2/27/2021 89

Task Success • % of subtasks completed • Correctness of each questions/answer/error msg • Correctness of total solution – Attribute-Value matrix (AVM) – Kappa coefficient • Users’ perception of whether task was completed 2/27/2021 90

Task Success • Task goals seen as Attribute-Value Matrix ELVIS e-mail retrieval task (Walker et al ‘ 97) “Find the time and place of your meeting with Kim. ” Attribute Selection Criterion Time Place Value Kim or Meeting 10: 30 a. m. 2 D 516 • Task success can be defined by match between AVM values at end of task with “true” values for AVM 2/27/2021 91 Slide from Julia Hirschberg

Efficiency Cost • Polifroni et al. (1992), Danieli and Gerbino (1995) Hirschman and Pao (1993) • Total elapsed time in seconds or turns • Number of queries • Turn correction ration: number of system or user turns used solely to correct errors, divided by total number of turns 2/27/2021 92

Quality Cost • # of times ASR system failed to return any sentence • # of ASR rejection prompts • # of times user had to barge-in • # of time-out prompts • Inappropriateness (verbose, ambiguous) of system’s questions, answers, error messages 2/27/2021 93

Another key quality cost • “Concept accuracy” or “Concept error rate” • % of semantic concepts that the NLU component returns correctly • I want to arrive in Austin at 5: 00 – DESTCITY: Boston – Time: 5: 00 • Concept accuracy = 50% • Average this across entire dialogue • “How many of the sentences did the system understand correctly” 2/27/2021 94

PARADISE: Regress against user satisfaction 2/27/2021 95

Regressing against user satisfaction • Questionnaire to assign each dialogue a “user satisfaction rating”: this is dependent measure • Set of cost and success factors are independent measures • Use regression to train weights for each factor 2/27/2021 96

Experimental Procedures • Subjects given specified tasks • Spoken dialogues recorded • Cost factors, states, dialog acts automatically logged; ASR accuracy, barge-in hand-labeled • Users specify task solution via web page • Users complete User Satisfaction surveys • Use multiple linear regression to model User Satisfaction as a function of Task Success and Costs; test for significant predictive factors 2/27/2021 97 Slide from Julia Hirschberg

User Satisfaction: Sum of Many Measures Was the system easy to understand? (TTS Performance) Did the system understand what you said? (ASR Performance) Was it easy to find the message/plane/train you wanted? (Task Ease) Was the pace of interaction with the system appropriate? (Interaction Pace) Did you know what you could say at each point of the dialog? (User Expertise) How often was the system sluggish and slow to reply to you? (System Response Did the system work the way you expected it to in this conversation? (Expected Behavior) Do you think you'd use the system regularly in the future? (Future Use) 2/27/2021 98 Adapted from Julia Hirschberg

Performance Functions from Three Systems • • • ELVIS User Sat. =. 21* COMP +. 47 * MRS -. 15 * ET TOOT User Sat. =. 35* COMP +. 45* MRS -. 14*ET ANNIE User Sat. =. 33*COMP +. 25* MRS +. 33* Help – – 2/27/2021 COMP: User perception of task completion (task success) MRS: Mean (concept) recognition accuracy (cost) ET: Elapsed time (cost) Help: Help requests (cost) 99 Slide from Julia Hirschberg

Performance Model • Perceived task completion and mean recognition score (concept accuracy) are consistently significant predictors of User Satisfaction • Performance model useful for system development – Making predictions about system modifications – Distinguishing ‘good’ dialogues from ‘bad’ dialogues – As part of a learning model 2/27/2021 100

Now that we have a success metric • Could we use it to help drive learning? • In recent work we use this metric to help us learn an optimal policy or strategy for how the conversational agent should behave 2/27/2021 101

New Idea: Modeling a dialogue system as a probabilistic agent • A conversational agent can be characterized by: – The current knowledge of the system • A set of states S the agent can be in – a set of actions A the agent can take – A goal G, which implies • A success metric that tells us how well the agent achieved its goal • A way of using this metric to create a strategy or policy for what action to take in any particular state. 2/27/2021 102

What do we mean by actions A and policies ? • Kinds of decisions a conversational agent needs to make: – When should I ground/confirm/reject/ask for clarification on what the user just said? – When should I ask a directive prompt, when an open prompt? – When should I user, system, or mixed initiative? 2/27/2021 103

A threshold is a human-designed policy! • Could we learn what the right action is – – Rejection Explicit confirmation Implicit confirmation No confirmation • By learning a policy which, – given various information about the current state, – dynamically chooses the action which maximizes dialogue success 2/27/2021 104

Another strategy decision • Open versus directive prompts • When to do mixed initiative 2/27/2021 105

Outline • • The Linguistics of Conversation Basic Conversational Agents – – • ASR NLU Generation Dialogue Manager Design – Finite State – Frame-based – Initiative: User, System, Mixed • • Voice. XML Information-State – Dialogue-Act Detection – Dialogue-Act Generation • • Evaluation Utility-based conversational agents – MDP, POMDP 2/27/2021 106

END of TODAY’S LECTURE • THE FOLLOWING SLIDES ARE AN OPTIONAL ADVANCED DISCUSSION OF MARKOVDECISION-PROCESS DIALOGUE SYSTEMS. 2/27/2021 107

Review: Open vs. Directive Prompts • Open prompt – System gives user very few constraints – User can respond how they please: – “How may I help you? ” “How may I direct your call? ” • Directive prompt – Explicit instructs user how to respond – “Say yes if you accept the call; otherwise, say no” 2/27/2021 108

Review: Restrictive vs. Non-restrictive gramamrs • Restrictive grammar – Language model which strongly constrains the ASR system, based on dialogue state • Non-restrictive grammar – Open language model which is not restricted to a particular dialogue state 2/27/2021 109

Kinds of Initiative • How do I decide which of these initiatives to use at each point in the dialogue? Grammar Open Prompt Restrictive Doesn’t make sense System Initiative Non-restrictive User Initiative 2/27/2021 Directive Prompt Mixed Initiative 110

Modeling a dialogue system as a probabilistic agent • A conversational agent can be characterized by: – The current knowledge of the system • A set of states S the agent can be in – a set of actions A the agent can take – A goal G, which implies • A success metric that tells us how well the agent achieved its goal • A way of using this metric to create a strategy or policy for what action to take in any particular state. 2/27/2021 111

Goals are not enough • Goal: user satisfaction • OK, that’s all very well, but – Many things influence user satisfaction – We don’t know user satisfaction til after the dialogue is done – How do we know, state by state and action by action, what the agent should do? • We need a more helpful metric that can apply to each state 2/27/2021 112

Utility • A utility function – – maps a state or state sequence onto a real number describing the goodness of that state I. e. the resulting “happiness” of the agent • Principle of Maximum Expected Utility: – A rational agent should choose an action that maximizes the agent’s expected utility 2/27/2021 113

Maximum Expected Utility • Principle of Maximum Expected Utility: – A rational agent should choose an action that maximizes the agent’s expected utility • Action A has possible outcome states Resulti(A) • E: agent’s evidence about current state of world • Before doing A, agent estimates prob of each outcome – P(Resulti(A)|Do(A), E) • Thus can compute expected utility: 2/27/2021 114

Utility (Russell and Norvig) 2/27/2021 115

Markov Decision Processes • Or MDP • Characterized by: – a set of states S an agent can be in – a set of actions A the agent can take – A reward r(a, s) that the agent receives for taking an action in a state – (+ Some other things I’ll come back to (gamma, state transition probabilities)) 2/27/2021 116

A brief tutorial example • Levin et al (2000) • A Day-and-Month dialogue system • Goal: fill in a two-slot frame: – Month: November – Day: 12 th • Via the shortest possible interaction with user 2/27/2021 117

What is a state? • In principle, MDP state could include any possible information about dialogue – Complete dialogue history so far • Usually use a much more limited set – – – Values of slots in current frame Most recent question asked to user Users most recent answer ASR confidence etc 2/27/2021 118

State in the Day-and-Month example • Values of the two slots day and month. • Total: – – – 2 special initial state si and sf. 365 states with a day and month 1 state for leap year 12 states with a month but no day 31 states with a day but no month 411 total states 2/27/2021 119

Actions in MDP models of dialogue • Speech acts! – – – Ask a question Explicit confirmation Rejection Give the user some database information Tell the user their choices • Do a database query 2/27/2021 120

Actions in the Day-and-Month example • • ad: a question asking for the day am: a question asking for the month adm: a question asking for the day+month af: a final action submitting the form and terminating the dialogue 2/27/2021 121

A simple reward function • For this example, let’s use a cost function • A cost function for entire dialogue • Let – Ni=number of interactions (duration of dialogue) – Ne=number of errors in the obtained values (0 -2) – Nf=expected distance from goal • (0 for complete date, 1 if either data or month are missing, 2 if both missing) • Then (weighted) cost is: • C = wi Ni + we Ne + wf Nf 2/27/2021 122

3 possible policies Dumb Open prompt Directive prompt 2/27/2021 P 1=probability of error in open prompt P 2=probability of error in directive prompt 123

3 possible policies Strategy 3 is better than strategy 2 when improved error rate justifies longer interaction: open directive 2/27/2021 P 1=probability of error in open prompt P 2=probability of error in directive prompt 124

That was an easy optimization • Only two actions, only tiny # of policies • In general, number of actions, states, policies is quite large • So finding optimal policy * is harder • We need reinforcement leraning • Back to MDPs: 2/27/2021 125

MDP • We can think of a dialogue as a trajectory in state space • The best policy * is the one with the greatest expected reward over all trajectories • How to compute a reward for a state sequence? 2/27/2021 126

Reward for a state sequence • One common approach: discounted rewards • Cumulative reward Q of a sequence is discounted sum of utilities of individual states • Discount factor between 0 and 1 • Makes agent care more about current than future rewards; the more future a reward, the more discounted its value 2/27/2021 127

The Markov assumption • MDP assumes that state transitions are Markovian 2/27/2021 128

Expected reward for an action • Expected cumulative reward Q(s, a) for taking a particular action from a particular state can be computed by Bellman equation: • Expected cumulative reward for a given state/action pair is: – – immediate reward for current state + expected discounted utility of all possible next states s’ Weighted by probability of moving to that state s’ And assuming once there we take optimal action a’ 2/27/2021 129

What we need for Bellman equation • • A model of p(s’|s, a) Estimate of R(s, a) How to get these? If we had labeled training data – P(s’|s, a) = C(s, s’, a)/C(s, a) • If we knew the final reward for whole dialogue R(s 1, a 1, s 2, a 2, …, sn) • Given these parameters, can use value iteration algorithm to learn Q values (pushing back reward values over state sequences) and hence best policy 2/27/2021 130

Final reward • What is the final reward for whole dialogue R(s 1, a 1, s 2, a 2, …, sn)? • This is what our automatic evaluation metric PARADISE computes! • The general goodness of a whole dialogue!!!!! 2/27/2021 131

How to estimate p(s’|s, a) without labeled data • Have random conversations with real people – Carefully hand-tune small number of states and policies – Then can build a dialogue system which explores state space by generating a few hundred random conversations with real humans – Set probabilities from this corpus • Have random conversations with simulated people – Now you can have millions of conversations with simulated people – So you can have a slightly larger state space 2/27/2021 132

An example • Singh, S. , D. Litman, M. Kearns, and M. Walker. 2002. Optimizing Dialogue Management with Reinforcement Learning: Experiments with the NJFun System. Journal of AI Research. • NJFun system, people asked questions about recreational activities in New Jersey • Idea of paper: use reinforcement learning to make a small set of optimal policy decisions 2/27/2021 133

Very small # of states and acts • States: specified by values of 8 features – – – Which slot in frame is being worked on (1 -4) ASR confidence value (0 -5) How many times a current slot question had been asked Restrictive vs. non-restrictive grammar Result: 62 states • Actions: each state only 2 possible actions – Asking questions: System versus user initiative – Receiving answers: explicit versus no confirmation. 2/27/2021 134

Ran system with real users • 311 conversations • Simple binary reward function – 1 if competed task (finding museums, theater, winetasting in NJ area) – 0 if not • System learned good dialogue strategy: Roughly – Start with user initiative – Backoff to mixed or system initiative when re-asking for an attribute – Confirm only a lower confidence values 2/27/2021 135

State of the art • Only a few such systems – From (former) ATT Laboratories researchers, now dispersed – And Cambridge UK lab • Hot topics: – Partially observable MDPs (POMDPs) – We don’t REALLY know the user’s state (we only know what we THOUGHT the user said) – So need to take actions based on our BELIEF , I. e. a probability distribution over states rather than the “true state” 2/27/2021 136