Humancomputer dialogue processing: where are we

1
Human-computer dialogue processing where are we?

• Yorick Wilks
• Oxford Internet Institute
• and
• University of Sheffield
• www.dcs.shef.ac.uk/yorick
• Tartu, February 2005

2
Main points of talk

• Data-driven performance systems versus
  AI/linguistic theoretical systems, separate since
  1970s
• This distinction has mapped, till now, on domain
  versus general systems
• Both will require modeling of intentional
  phenomena like belief, possibly speech acts,but
  not in the old way.
• It is unclear whether speech technology alone can
  accomplish this
• Some Sheffield empirical work in dialogue
  technology
• The COMPANIONS development environment.

3
Problems of dialogue theory

• It can easily be vacuous dialogues are systems
  of turn-taking
• Speech act analysis has led to a rationalist
  tradition implausibly deep levels of reasoning,
  and no data gathering
• There are apparently competing theories but
  little or no evaluation of dialogue systems
  (compare Information Extraction--Hobbs)

4
Three important historical systems

• Colbys PARRY (Stanford, 1971)
• Groszs partitioned network focus system (SRI,
  1979)
• Perrault, Cohen, Allens speech act system
  (Toronto, 1979)

5
Colbys PARRY

• Perhaps best ever performance, many users,
  robust, but not a normal subject (I.e. a
  paranoid)
• primitive individual models, some control of
  dialogue process but it had lots to say!
• Primitive simulation of intentionality
• not syntax but fast pattern matching
• much better than ELIZA

6
PARRY conversation

• Have you been hospitalized before?
• THIS IS THE FIRST TIME
• How long have you been there?
• ABOUT TWO WEEKS
• Any headaches?
• MY HEALTH IS FINE
• Are you have memory difficulties?

7
More PARRY

• NO
• Then can you tell me who is the President of the
  US?
• NIXON IS PRESIDENT
• And what day is today?
• TODAY IS THURSDAY--CAN WE GET ON WITH THE
  INTERVIEW?

8
How is this achieved in PARRY?

• WHAT IS YOUR MAIN PROBLEM _____________________
  __________
• WHAT BE YOU MAIN PROBLEM
• BE YOU MAIN PROBLEM
• WHAT BE MAIN PROBLEM
• WHAT BE YOU PROBLEM ?
• WHAT BE YOU MAIN

9
Groszs robot domain model

• Primitive notion of focus based on a structure of
  partitioned semantic nets (Hendrix)
• Unlike PARRY, has knowledge of a complex domain
• no real performance at all

10
Grosz (IJCAI 1979)

• (Explicit focus)
• S1 The lid is attached to the container with
  1/2 bolts.
• R1 Where are the BOLTS?
• (Implicit focus)
• S1 Attach the lid to the container.
• R1 Where at the BOLTS?

11
Perrault, Cohen, Allen system

• Based on speech act reasoning
• user must have one of two goals, meeting or
  catching a train
• Passenger/User Do you know when the Windsor
  train arrives?
• This is labelled as a REQUEST not a
  REQUEST-INFORM (Y/N) because the system knows the
  user knows it knows!

12
Perrault et al.

• System has domain knowledge and reasoning power
• but virtually no performance
• was the first to assign speech act labels to
  dialogue items
• has a simple rigid model of nested belief

13
Fixed nested beliefs passengers view of
systems view of passengers beliefs.
passenger
system
passenger
14
Perrault et al.

• They had the appropriate notion of extreme
  ellipsis and zero anaphor
• Passenger The 3.15 train to Windsor?
• System Gate 10
• Note NOT It leaves from Gate 10

15
1970s division of approaches to machine
conversation.

• Domain-dependent systems with coded world
  knowledge and some parsing and reasoning
• VERSUS
• Wide shallow systems with little knowledge and
  high performance
• Published AI academic systems all in first group
• Only the second group performed at all

16
Moving academic systems towards performance

• Best is Traums TRAINS system--descendant of
  Allens work (Toronto-Rochester tradition)
• uses corpora and retains reasoning
• moved to the movies in California!

17
(No Transcript)
18
TRAINS CORPUSDavid Traum (Rochester)

• utt1 s hello can I help you
• utt2 u yeah I want t- I want to determine the
  maximum number of boxcars of oranges by
  seven a.m. tomorrow morning
• utt3 so hm
• so I guess all the boxcars will have to go
  through oran- through Corning because
  thats where the orange juice orange
  factory is
• utt4 so from Corning to Bath how far is
  that
• utt5 s two hours
• utt6 u and its gonna take us also an hour to
  load boxcars right
• utt7 s right
• utt8 u okay so hm so every
  trip will take at least three hours
  then
• utt9 um
• utt10 s right we can unload any amount of
  cargo onto a train in one hour
• utt11 so we can so a maximum of
  three boxcars in an hour

19
AI people join the performersthe Loebner
Competition

• Well designed competition over the last ten
  years---essentially a modified Turing test.
• Domain dependent entries till 1997, then domain
  free.
• The game is for journalists to separate and rank
  human and machine partners
• Search Loebner on web for full transcripts

20
Loebner 1994

• NEW JUDGE
• 150
  Judge 1
• Do you think the country is headed generally in a
  more conservative or more liberal direction right
  now?
• 151
  Terminal
• Didnt you say would you discribe that old texas
  billionaire as a conservative?? Answer me one
  question why are liberals portrayed as donkees?
  Are you a caring person? Youre not really
  talking about me, are you?
• 152
  Judge 1
• Ive got the feeling Ive heard these answers
  before. Is our nation swinging leftward or
  rightward right now?
• 
  Terminal
• We are at the beginning of a strong and
  wonderful economic recovery, and we can thank the
  liberals!

21
Loebner 1997

• PROGRAM101804Hi, my name is Catherine, whats
  yours?
• JUDGE04101832My name is Elliott.
• PROGRAM101834Well, its nice to meet you,
  Elliott.
• PROGRAM101848Did you see that story on CNN
  last night about the lesbian
• PROGRAM101904couple who came out at a White
  House party on Sunday?
• JUDGE04101955Yes, I did. I think it may all
  be a publicity stunt for Ellen.
• PROGRAM102001Ellen Degeneres was one of them
  she was kissing her lover
• PROGRAM102011in front of Bill Clinton. Do you
  think he was very
• PROGRAM102025embarrassed?
• JUDGE04102100I think he want the gay vote. He
  will do anything to get it and keep it.
• JUDGE04102115He does not have a conservative
  base.
• PROGRAM102121
• PROGRAM102121Are you still a Friend of Bill
  Clinton?

22
We won in 1997!

• The CONVERSE program pretended to be Catherine, a
  34-year old British journalist living in New York
  (to explain any usage problems).
• She could talk about 80 appropriate topics
• (80domain free???) clothes, shopping, food,
  abortion, fitness, sex.etc.

23
Moving performance towards theory

• The Loebner Prize Competition
  www.acm.org/-loebner/loebner-prize.html

24
The CONVERSE prototype 1997

• Push-me-pull-you architecture
• strong driving top-down scripts (80) in a
  re-enterable network with complex output
  functions
• bottom-up parsing of user input adapted from
  statistical prose parser
• minimal models of individuals
• contained Wordnet and Collins PNs
• some learning from past Loebners BNC

25
CONVERSE architecture
26
Restarting Sheffield dialogue work in 2001

• an empirical corpus-based stochastic dialogue
  grammar that maps directly to dialogue acts and
  uses IE to match concepts with templates.
• A better virtual machine for script-like objects
  encapsulating both the domain and conversational
  strategy (cf. PARRY and Grosz) to maintain the
  push-pull approach.
• Need for resources to build belief system
  representations and quasi-linguistic models of
  dialogue structure, scripts etc.
• A model of speakers, incrementally reaching
  VIEWGEN style ascription of belief procedures to
  give dialogue act reasoning functionality

27
(No Transcript)
28
VIEWGENa belief model that computes agents
states

• Not a static nested belief structure like that of
  Perrault and Allen.
• Computes other agents RELEVANT states at time of
  need
• Topic restricted search for relevant information
• Can represent and maintain conflicting agent
  attitudes

29
VIEWGEN as a knowledge basis for
reference/anaphora resolutionprocedures

• Not just pronouns but grounding of descriptive
  phrases in a knowledge basis
• Reconsider finding the ground of
  that old Texas billionaire as
  Ross Perot, against a background of what the
  hearer may assume the speaker knows when he says
  that.

30
(No Transcript)
31
cause(deleted(x,sub-directory) not(happy(david)))
deleted(system,sub-directory)
system
belief
deleted(system,sub_directory)
simon
belief
system
goal
Candidate Planning Actions for system
I have just deleted the subdirectory
Click here for Plan recognition
Perform Action
Belief representation of the first turn of an
exchange.
32
CONVERSE VIEWGEN

• for ViewGen
• Ballim, A.and Wilks, Y. 1991
  Artificial Believers the ascription of belief
  Erlbaum, Hillsdale NJ
• for CONVERSE and contemporary paradigms (esp. in
  industry)
• Proceedings of Bellagio Workshops on
  Human-Machine Conversation, 97,98,00

33
Meanwhile, the world had moved on and empiricism
had reached dialogue processing

• Dialogue act-to-utterance learning using machine
  learning over n-grams and preceding dialogue acts
  (Cf. Samuels et al. 1998)
• Speech act sequence statistics from Verbmobil
  (Cf. Maier Reithinger)
• Longman BNC dialogue ngrams (50 all domains)
• Tagged Switchboard, Verbmobil, BNC and domain
  corpora

34
And especially, Steve Young 2002

• ASR now a mature statistically based technology
• In other areas of dialogue, statistical methods
  less mature
• A complete dialogue system can be seen as a
  Partially Observable Markov process
• Subcomponents can be observed in turn with
  intermediate variables

35
Youngs statistical modules

• Speech understanding
• Semantic decoding
• Dialogue act detection
• Dialogue management and control
• Speech generation
• I.e. roughly same as everyone elses!!

36
(No Transcript)
37
Strategy not like Jelineks MT strategy of 1989!

• Which was non/anti-linguistic with no
  intermediate representations hypothesised
• Young assumes rougly the same intermediate
  objects as we do but in very simplified forms.
• The aim to to obtain training data for all of
  them so the whole process becomes a single
  throughput Markov model.

38
Young concedes this model may only be for simple
domains

• His domain is a pizza ordering system
• A typical DialogueActSemantics could therefore
  be
• Purchase_Requestqty2toppingpepperoni

39
Speech Understanding

• Classic mapping of waveforms Y to dialogue acts
  A, given system beliefs B to seek
• Argmax P(AuY,Bs) which expands to
• Argmax P(YAu,Bs)P(AuBs)
• Then the intermediate wordsequence is introduced,
  giving
• P(YAu,Bs)? P(Y,WAu,Bs)
• w
• This is a redescription of what empirical NLP
  people had been doing for 5 years but no mention
  of them!

40
This breaks into two parts

• A classic two stage matching of sounds to a word
  lattice (I.e. ASR)
• Followed by the mapping of the Dialogue acts to
  the words, given the system beliefs (I.e. system
  state in general)
• the dialogue state is crucial for handling
  ambiguity and identifying underspecified dialogue
  acts

41
Introduction of concepts as well as DAs

•  it is convenient to introduce another
  intermediate representation.C represents the set
  of semantic concepts encoded within W 
• So P(AuW,Bs)?P(AuC,Bs)P(CW,Bs)
• c
• I.e.training for concepts given wordsstates
• And Dialogue acts given conceptsstates.
• Experience suggsts both can be trained directly
  on the words!

42
(No Transcript)
43
Semantic decoding

• Means adding one  concept  per main word input
  (!) by means of a  semantic template grammar 
  such as
• PIZZA-QTYTOPPING pizza(s)
• A parse tree is an option if recursive syntactic
  structure is required, over simple rule matching,
  where concepts label states and concept bigram
  probabilities are computed.
• Parse tree gives Hidden Understanding Model using
  lexical and concept bigrams.
• Or learn the simple grammar templates above
  directly

44
Dialogue act detection

• Range of methods (including MDL) for solving
  argmax P(AuC,Bs)
• Interesting that this is not thought to depend on
  the WORDS (as most people do it)
• No reference to linguistic methods for this
  (since Samuels et al 98).

45
Dialogue management

• This is the one where it is hard to see how he
  can get non-trivial data.
• Data can be seen as a transition matrix of system
  states S against system actions As, filling the
  matrix cells with new system states.
• Model of training is reinforcement learning
  ascribed to Pierracini and then Walker.
• No evidence of what such training changes in
  practice for a non-trivial system
•  the typical system S will typically be
  intractably large and must be approiximated 
• Puzzle  the users beliefs cannot be directly
  observed and must therefore be inferred  true,
  but.

46
Questions about the Speech program for dialogue

• Is this just a description of the empirical NLP
  program of work or an attempted reduction like
• Jelineks MT without linguists
• Pollacks RAAM for (Fodors) syntactic recursion
• Can data be found to reduce all of a (fairly
  general) DM to a transition space that can be
  reward trained?
• What is the real effect of training a DM--however
  found? Reducing unused paths?
• How in principle express changes to planning and
  belief spaces?

47
Back to work in NLPLearning to tag for Dialogue
Acts initial work

• Reithinger and Klesen (1997), n-gram language
  modelling, Verbmobil corpus (75)
• Samuels et al.(1998) TBL learning on n-gram DA
  cues, Verbmobil corpus (75)
• Stolcke et al. (2000) more complex n-gram
  (including over DA sequents), more complex
  Switchboard corpus (71)

48
Work at Sheffield Starting with a naive
classifier for DA s

• Direct predictivity of DA s by n-grams as a
  preprocess to any ML algorithm.
• Get P(dn) for all 1-4 word n-grams and the DA
  set over the Switchboard corpus, and take DA
  indicated by n-gram with highest predictivity
  (threshold for probability levels)
• Do 10-foldcross validation (which lowers scores)
• Gives 63 over Switchboard but using only a
  fraction of the data Stolcke needed.
• Work by Nick Webb (in EC Amities Project)

49
Extending the pretagging with TBL

• Gives 72 (Stolckes 71) over the Switchboard
  data, but only 3 is due to TBL (rather than the
  naive classifier).
• Samuiels unable to see what TBL is doing for him.
• This is just a base for a range of more complex
  ML algorithms (e.g. WEKA).
• Figure same as Stolckes without the cross
  validation

50
Design of a Dialogue Action Manager at Sheffield

• General purpose DAM where domain dependent
  features are separated from the control
  structure.
• The domain dependent features are stored as
  Dialogue Action Forms (DAFs) which are similar to
  Augmented Transition Networks (ATNs)
• The DAFs represent general pupose Dialogue
  manoevres as well as application specific
  knowledge.
• The control mechanism is based on a basic stack
  structure where DAFs are pushed and popped during
  the course of a user session.

51
The structure of the DAM

• The control mechanism together with the DAFs
  provide a flexible means for guiding the user
  through the system goals (allowing for topic
  change and barge in where needed).
• User push is given by the ability to suspend
  and stack a new DAF at any point (for a topic or
  any user maneuver)
• System push is given by the prestacked DAFs
  correcsponding to what the system wants to show
  or elicit.

52
DAM Implementation

• The core mechanism for the DAM is a simple
  pop-push stack (slightly augmented), onto which
  structures are loaded and run.
• The structures are Dialogue Action Forms (DAFs),
  which are implemented as Augmented Transition
  Networks (ATNs).
• ATNs are transition networks that can function
  at any level from finite state machines to Turing
  Machine power.

53
DAM Implementation 2

• The stack is pre-loaded with DAFs that satisfy
  the overall goal of designing a bathroom in the
  COMIC system.
• The control structure has a preference for
  continuing to run the DAFs on the stack (system
  initiative) unless the user input is outside the
  scope of the DAF (user initiative). In which
  case, the control structure pushes onto the stack
  the new DAF that most closely matches the user
  input.

54
DAM Implementation 3

• DAFs model the individual topics and
  conversational manoeuvres in the application
  domain..
• The stack structure will be preloaded with those
  DAFs which are necessary for the COMIC bathroom
  design task and the dialogue ends when the
  Goodbye DAF is popped.
• DAFs and stack interpreters together control the
  flow of the dialogue

55
Dialogue Management the stack

• Work by Catizone and Setzer (EU COMIC Project)
• Contrast with work at CSLI (Stanford) by Peters
  and Lemon.

Greeting DAF
Room measurement DAF
Style DAF

Good-bye DAF
56
Dialogue Management the augmented stack

• What happens if we match/push onto the stack a
  DAF that is already on the stack (further down)?
• In this case, we run the DAF and it is then
  either satisfied (poppable) or not.
• If it is then, on popping, we remove any
  (unrun/prestored) copies of the DAF below it on
  the stack (slight augmentation here of a simple
  stack!)

57
An augmented stack 2

• If it is not satisfied (almost certainly because
  of context dependent DAFs still to come on stack)
  the DAF traps the response and recommends waiting
  to discuss that topic and is popped. The copy of
  the same DAF lower down the stack is then
  accessed and run later at the appropriate place
  in the dialogue (this is one of a set of possible
  strategies with which we shall experiment).

58
An augmented stack 3

• The result we would like with this mechanism is
  that a relevant and partly run DAF can be forced
  to the top of the stack and all intermediate
  partly run DAFs discarded (cf. Groszs notion of
  conversational topics being closed and
  unreachable for restarting)
• .

59
DAF example
60
Learning to segment the dialogue corpora?

• Segmenting the corpora we have with a range of
  tiling-style algorithms (initially by topic)
• To segment it plausibly, hopefully into segments
  that correspond to structures for DM (Dialogue
  Action Frames in our naming)
• Being done on the annotated corpus (i.e. a corpus
  word model) and on the corpus annotated by
  Information Extraction semantic tags (a semantic
  model of the corpus)
• Repetitions of DAsemantics template patterns amy
  may a MDL packing of the corpus possible as an
  alternative segmenting method.

61
Sheffield Dialogue Research Challenges

• Will a Dialogue manager raise the DA 75ceiling
  top-down?
• Multimodal dialogue managers. Are they completely
  independent of the modality? Are they really
  language independent?
• What is a good virtual machine for running a
  dialogue engine? Do DAFs provide a robust and
  efficient mechanism for doing Dialogue
  Management?
• Will they offer any interesting discoveries on
  stack access to, and discard of, incomplete
  topics (cf. Stacks and syntax).

62
More research challenges

• Applying machine learning to transcripts so as to
  determine the content of dialogue management,
  i.e. the scope and content of candidate DAFs.
• Can the state set of DAFs and a stack be trained
  with reinforcement learning?
• Can we add a strong belief/planning component to
  this and populate it empirically?
• Fusion with QA

63
Why the dialogue task is still hard

•  Where am I  in the conversationwhat is being
  talked about now, what do they want?
• Does topic stereotopy help or are just
  Finite-State pairs enough (VoiceXML? TRINDIKIT?)?
• How to gather the beliefs/knowledge required ,
  preferably from existing sources?
• Are there distinctive procedures for managing
  conversations that can be modelled by simple
  virtual machines?
• How to learn the structures we need--assuming we
  do---how to get and annotate the data?

64
How this research is funded

• AMITIES is a EU-US cooperative R D project
  (the first in NLP) 2001-2005 to automate call
  centers.
• University of Sheffield (UK prime)
• SUNY Albany (US prime)
• Duke U. (US)
• LIMSI Paris (Fr)
• IBM (US)
• COMIC is an EU R D project ( 2001-2005) to
  model multimodal dialogue
• MaxPlanck Inst (Nijmegen) (Coordiantor)
• University of Edinburgh
• MaxPlanck Inst (Tuebingen)
• KUL Nijmegen
• University of Sheffield
• ViSoft GMBH

65
There are now four not two competing approaches

• Logic-based systems with reasoning (old and still
  unvalidated by performance)
• Extensions of speech engineering methods, machine
  learning and simple intermediate structure
• Simple handcoded finite state systems in VoiceXML
  (Chatbots and commercial systems)
• Rational hybrids based on structure and machine
  learning.