IAAA / PSTALN Dialog systems - Benoit Favre last generated on January 20, 2020 - page du TP
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
IAAA / PSTALN
Dialog systems
Benoit Favre
Aix-Marseille Université, LIS/CNRS
last generated on January 20, 2020
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 1 / 35
What is a dialog system?
Definition :
▶ Input/output with natural language
▶ Free use of language
▶ Reproduces human agent behavior
▶ Reply (or not) in natural language
▶ Uni/multimodal
Spoken Dialog System (SDS)
▶ Interactive system with spoken language
▶ Required when using an acoustic communication channel only (phone)
▶ Can free other modalities (hands free)
Difficulty
▶ No control of inputs
▶ Contextualize information
▶ Automatic speech recognition → transcript errors
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 2 / 35
Examples
Online discussion
▶ Eliza, virtual therapy
⋆ https://www.eclecticenergies.com/ego/eliza
▶ Mitsuku (best chatbot at Loebner price 2013)
⋆ http://www.mitsuku.com/
Automated voice services
▶ “To erase a message, say erase..."
Customer care
▶ 1013 (in France): describe freely your problem
▶ Air Travel Information System (ATIS)
Assistants
▶ Clippy
▶ SIRI
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 3 / 35
Learning with an intelligent agent Replaces teacher in MOOC Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 4 / 35
Interaction with a robot
Caroline Lyon, Chrystopher L. Nehaniv, Joe
Saunders, Interactive Language Learning by
Robots: The Transition from Babbling to Nao (http://www.aldebaran-
Word Forms, PLoS One, 2012 robotics.com)
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 5 / 35
Problems
What’s a dialog?
▶ Study human behavior
How to understand a sentence?
▶ Rule-based and template-based system
▶ Robust concept detection
What strategies to make a successful dialog?
▶ Enforce local coherency
▶ Finite state machine
▶ Explore possible futures
How to formulate an answer
▶ Language and speech synthesis
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 6 / 35
Corpus-based study
Human-human dialogues
▶ Language in the wild
▶ Want to build a program that replicates human behavior
▶ Difficult to ignore non-verbal interactions
▶ Humans behave differently in front of a system
Wizard of oz (WoZ)
▶ Replace the system by a human operator
▶ Make the user believe that it is a real system (simulate errors / wait
time...)
▶ Collect users dialog strategies
Human-machine dialog
▶ Collect interactions with an existing system
▶ Use these data to evaluate / improve the system
Simulation
▶ From a user model, simulate human input
▶ Detect infinite loops
▶ Estimate resolution time
▶ Train system parameters
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 7 / 35Human dialogs
Sequence of spoken turns
▶ Between two or more people
Who speaks after whom?
▶ Interruptions
▶ Finish someone else’s sentence
▶ Overlap
Establish of a common ground
▶ Use of a common vocabulary
▶ Acquiescence, reformulation, convergence
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 8 / 35Speech acts
Theory (Austin & Searl) * Meaning can be expressed in terms of actions
instead of concepts (declarative logical form)
Examples:
▶ “I apologize"
▶ “Can you do this?"
Types of dialog acts
▶ verdictifs ou actes juridiques (acquitter, condamner, décréter)
▶ exercitifs (dégrader, commander, ordonner, pardonner, léguer)
▶ promissifs (promettre, faire vu de, garantir, parier, jurer de)
▶ comportatifs (sexcuser, remercier, déplorer, critiquer)
▶ expositifs (affirmer, nier, postuler, remarquer)
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 9 / 35Cooperation principle (Grice)
For a conversation to be successful, speakers must cooperate:
▶ Quantity: give the right amount of information
▶ Quality: tell the truth
▶ Relevance: say important things
▶ Manner: be clear, brief and structured
Example (Mitkov - Computational Linguistics)
▶ Quantity: Marie ate some chocolate → Marie did not eat all the
chocolate
▶ Quality: (about an invoice) It costs an arm → It costs a lot
▶ Relevance: A: Can I watch TV? B: It’s bath time.
▶ Manner: Are you ready? vs Are you ready or are you not ready?
Integration in a dialogue system:
▶ The user follows them if he has something to gain
▶ Should the system follow these principles?
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 10 / 35Dialog acts (DA)
Dialog acts: a specialization of speech acts
Subdivision of a speaking turn into “intentions"
▶ Question
⋆ Closed / open questions
▶ Declaration
⋆ Short answers
▶ Dysfluancies
⋆ Repetitions
⋆ Wrong pronunciation
▶ Interruption
▶ Filled pauses
▶ MRDA / DASL: specification of more than 160 types
Task-oriented dialogue act categories
▶ Greetings
▶ Opening
▶ Negotiation
▶ Closing
▶ Good-bye
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 11 / 35Anatomy of a dialog system
syntactic
words tree
concepts,
question Automatic Syntactic Semantic
relations
transcription analysis analysis
Dialog
manager
Speech Lexical Syntactic logical
answer
synthesis generation generation representation
words, primitive
prosody syntax
Comprehension
▶ Dialogue acts
▶ Grammar
▶ Concepts
Dialog management
▶ Matching
▶ Finite state machines
▶ Exploration
Generation
▶ Templates
▶ Statistical generation
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 12 / 35DA classification
Model domain dialog acts (politeness, commands, information, ...)
Corpus: (15651 instances, 66 classes)
▶ “je n’ai plus de tonalité sur ma ligne" → interruption_ligne
▶ “j’ai un problème suite euh déménagement" → mise_en_service
▶ “euh téléphone illimité en panne" → internet_voip
▶ “le clignotant reste toujours allumé" → messagerie_vocale
Classifier
1 Extract word n-gramms
⋆ je je_n'ai je_n'ai_plus n'ai_plus_de plus_de_tonalité...
2 Use them as features of a classifier (résultats mlcomp.org)
⋆ 0.207 SMO_weka_nominal
⋆ 0.216 boostexter
⋆ 0.232 sgd-logistic-stepsize0.3-iter5
⋆ 0.272 liblinear-s6-B1
3 Most relevant features
⋆ “plus de tonalité"
⋆ “sur ma ligne"
⋆ “une autre demande"
⋆ “ai un problème"
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 13 / 35Grammars
Extension of classic syntagmatic grammars
▶ Integrate domain semantics into grammar
Requête
Action Voyage
Je voudrais Moyen Départ Arrivée Temps
un train de Paris à Marseille dans l'après-midi
Grammar example:
Request -> $Action $Travel
Travel -> $Mean $Departure $Arrival $Time
Action -> I would like | I would have to | have you got
Mean -> a train | a flight | a plane
Departure -> from $City | starting from $City
Arrival -> to $City | arriving at $Ville
Weather -> in the afternoon | today | tomorrow | at $Hour | the $Date
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 14 / 35Concept detection
Example: from Media corpus (WoZ, hotel reservation)
▶ Detect triplets (modality, type, value) linked to a task
▶ heu [+:reponse:oui oui] [+:connect:opposition mais] j’aimerai
d’abord savoir si le [+:hotel:Ibis ibis] il y a un
[?:service:jacuzzi jacuzzi] et une [?:service:piscine piscine] si
ils acceptent [?:service:animaux les chiens]
Formulate as sequence prediction
▶ BIO formalism (begin-inside-outside)
▶ Model type HMM/RNN/CRF
▶ Features on words, pos-tags...
... ...
occasion O
de O
la B-evenement
fête I-evenement
du I-evenement
cheval I-evenement
donc B-connect
euh O
à B-loc-relative
proximité I-loc-relative
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 15 / 35Dialogue management
Modèle de Modèle de
l'utilisateur la tâche Réponse
Analyse
sémantique
Modèle de Modèle Commandes
discours du monde
Entrées Gestionnaire de dialogue Sorties
Task model: expected requests, possible commands
User model: do not tell the user what he already knows, predict the
following sentence (objectives, knowledge, interests)
Discourse model: conversation history (pronouns resolution), dialogue
states, what to do at a given time
World model: general knowledge for understanding
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 16 / 35ELIZA: the psychanalist EMACS: alt-x-doctor Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 17 / 35
ELIZA: How does it work?
Key words → responses
▶ “BONJOUR" → “Comment vas-tu aujourd’hui.. De quoi désires-tu
discuter?"
▶ “PEUX-TU" → “Tu ne crois que je suis capable deDialog management: AIML (ALICE)
AIML (Artificial Intelligence Markup Language)
WHO IS YOUR DADDY?
Steve
Recursive rules
Memorize previous answer and topic (broad category)
Synonyms
Random
Learn new responses from user input
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 19 / 35ALICE: how to make a bot ?
Available source code
▶ Program AB: https://code.google.com/p/program-ab/
▶ Knowledge base:
https://code.google.com/p/aiml-en-us-foundation-alice/
Question/answers
▶ Yes ("Is violet a color?")
▶ No ("Are fish mammals?")
▶ Sometimes ("Is the sky blue?")
Reductions
▶ Synonyms: “Hello", “Hi there", “Howdy" → “Hi".
▶ Simplification: “I am feeling very happy right now" → “I am happy".
▶ Input segmentation: “Yes my name is Jim" → “Yes" + “My name is
Jim".
Personality
age = 15
baseballteam = Red Sox
birthday = Nov. 23, 1995
birthplace = Bethlehem, Pennsylvania
boyfriend = I am single
celebrities = Oprah, Steve Carell, John Stewart, Lady Gaga
... Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 20 / 35Voice XML
W3C standard
XML representation of a dialogue
Please choose airline, hotel, or rental car.
[airline hotel "rental car"]
You have chosen .
Can specify a grammar for the fields to be filled and the order in
which to fill them
Code evaluation ()
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 21 / 35Information retrieval-based system
Find most relevant answer for a question
▶ Exemple : stack overflow
Multi-user: ask a question to another user (chatbots)
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 22 / 35Automaton
Finite state machine to represent dialogue states
▶ Don’t list flights until all the fields are filled
Entry into a state is conditioned by an interpretation (of concepts,
recognition DA...) and dialogue history
A state corresponds to a command and/or a response from the system
Limit: locked in the structure of the dialogue
▶ Given a state and an interpretation of the user, we always go to
another state
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 23 / 35MDP
Problem: Which trajectory to follow in the dialog automaton to
minimize task completion time?
▶ We don’t know the future, we have to make assumptions
Markov Decision Process
▶ At each time t, the process (the user) is in a state s (it has an
intention)
▶ The system chooses to perform an action a (for example it asks a
question)
▶ This action randomly places the process (the user) in state s ′ ,
according to a probability Pa (s, s ′ )
▶ This change of state gives the system a gain Ra (s, s ′ )
▶ The choices of the process depend only on a and s and not on the rest
of the history (Markov property)
How to maximize the cumulative gain over the entire dialogue?
▶ We call policy a series of actions noted Pi
▶ Use dynamic programming to find the policy that maximizes
∞
∑
γt Ra (st , st+1 )
t=0
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 24 / 35POMDP
Partially Observable Markov Decision Process (POMDP)
▶ Extension of MDP the state (user intent) isn’t directly observed
▶ Distribution over states that the user can be in given what she says
▶ Observations are user sentences
MDP POMDP
s' s'
R(s,s') R(s,s')
s s
Passé Futur Passé Futur
Requires approximate inference
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 25 / 35Model comparison
Markov Chain
Markov Decision Process ? ?
? ?
Hidden Markov Chain
? ?
Partially-observable MDP
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 26 / 35Advanced elements
Initiative
▶ System initiative (= le technicien freebox qui suit un script)
▶ User initiative (Google Now, commande vocale)
▶ Mixed initiative :
⋆ “SIRI, can you change brightness?" (user initiative)
⋆ “Yes, how bright?" (system initiative)
Confirmation
▶ Explicit
⋆ “J’aimerai aller de Marseille à Barcelone"
⋆ “Votre lieu de départ est-il Marseille ?"
⋆ “Oui"
⋆ “Votre lieu d’arrivée est-il Barcelone"
▶ Implicite
⋆ “J’aimerai aller de Marseille à Barcelone"
⋆ “Quand souhaitez-vous aller de Marseille à Barcelone ?"
Who is the user talking to?
▶ Kinect / Nao
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 27 / 35Chatbots with deep learning
Replace classic elements
1 Intent classification (what do you want?)
2 Slot filling (what are the parameters of your query?)
3 Next action prediction
Typical frameworks (i.e. RASA)
▶ Provide pretrained intents / slots
▶ Active learning for annotating data
▶ Dialog flow is pre-scripted (best control)
Data-driven approach to conversation modeling
▶ Given a conversation up to a point, can we predict what will happen
next
▶ No need for linguistic analysis, but no linguistic prior
▶ Examples:
1 Alternating language models
2 Turn retrieval
3 Machine translation (history → next turn)
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 28 / 35Alternating language model
A simplified version of the encoder-decoder (or seq2seq) framework
▶ Trained the same way as a regular word-based language model
▶ At prediction time, alternate between user input and generation
⋆ Training data needs to be in the same form
Implementations
▶ Word-by-word prediction
▶ Any language model (GPT-2...)
▶ Attention mechanisms
w1 w2 w3 w1 w2 w3
M H w1 w2 M
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 29 / 35Representation learning
Create an information retrieval system
▶ Which can retrieve the next turn given a history
▶ Encode history with a first recurrent model
▶ Encode next turn with a second recurrent model
▶ Compute a similarity between those representations (dot product)
Training objective: triplet ranking
▶ Make sure the correct association has a higher score than a randomly
selected pair
Problem: the cost of retrieving a turn
▶ Everything can be precomputed, just the dot product remains
▶ Many approaches for finding approximate nearest neighbors in a high
dimensional space (ie. locality preserving hashing)
history
M w1 w2 H w1 w2 cosine
response
M w1 w2
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 30 / 35Bi-encoder training
Maximize margin between the result of hi · ri and ni · ri
▶ hi is the history
▶ ni is a random history
▶ ri is the response
1∑
Loss = max(0, 1 − hi · ri + ni · ri ))
n i
Keras model
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 31 / 35Experiments on Datcha corpus
Corpus: Orange ATH TV
Stat Train Valid Test
Conversations 16,140 698 606
Turns 465,693 20,090 18,392
Words 7,744,262 327,979 299,340
Preprocessing
▶ Tokenization (based on penn tokenizer)
▶ A few rules to strip additional URLs, phone numbers, etc.
▶ Lower case
▶ Concatenate turns of the same participant with
▶ Separate conversations by
▶ Replace all TC[1-9] by a generic TC
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 32 / 35Datcha results
Evaluation metrics
∑
▶ Perplexity (PPL): − n1 logP(turn|history)
▶ Better-than-random (BTR): n1 |P(turn|history) > P(turn|noise)|
Results on the ATH TV test set (3 last files):
Method PPL BTR
Language model 17.52 69.39%
Information retrieval 11.85 93.91%
Parameters
▶ LM: vocab=30k, layers=2, hidden=650, sample=1024, maxlen=35,
batch=20, optim=sgd, epochs=8
▶ Bi-encoder: vocab=30k, embeddings=128 (init=w2v), hidden=256,
maxlen=64, repr=128, batch=256, optim=Nadam, epochs=100
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 33 / 35t-SNE Analysis t-SNE Projections of turn representations Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 34 / 35
Conclusion
Dialogue systems
▶ Understanding
▶ Planning
▶ Data driven
Open issues
▶ Noisy input
⋆ Speech recognition / spelling errors
▶ Unrestricted domain
⋆ Chit-chat
⋆ Rapid development of new domains
▶ Grounding
⋆ Tackle real-world objects
▶ End-to-end training
⋆ Speech-recognition → dialog management → speech synthesis
▶ Training without experiencing
⋆ Simulation
Current state of the art
▶ https://nlpprogress.com/english/dialogue.html
Benoit Favre (AMU) PSTALN: Dialog January 20, 2020 35 / 35You can also read
