Spoken Dialogue Systems

1
(Spoken) Dialogue Systems
2
Dialogue structure

• Dialogue Acts (shallow view)
• Austin actions, Searle speech acts, Dialogue Acts
• DAMSL (Core and Allen, 1997)
• Predicting Dialogue Acts (Stolke et al. 2000)
• Automated discourse analysis
• Grosz Sidner Theory
• Dialogue Managers
• Finite state (McTear 1998)
• Form-based (Bohus and Rudnicky 2003)
• Plan-based (Larsson Traum, 2000 Rich et al.,
  2001)
• Probabilistic (Horvitz Paek, 1999)

3
Dialogue strategy

• Dialogue as Markov Decision Process
• State space, cost function, parameters
• Reinforcement Learning
• Learn entire dialogue structure (Levin,
  Pieraccini, Eckert, 2000)
• Learn initiative and confirmation strategy
  (Scheffler Young, 2002 Singh, Litman, Kearns,
  Walker, 2002)
• Issues
• State space design
• Simulated user
• Cost function
• No user adaptation

4
Dialogue strategy (continued)

• Other approaches
• Utility based (Paek Horvitz, 2004)
• Minimize support cost
• Dempster-Shafer theory of evidence (Chu-Carroll,
  2000)
• Value of Information (Horvitz Paek, 1999)

5
Errors in interaction

• Channel Errors
• Detection (Litman, Hirschberg, Swerts, 2000)
• Recovery
• Modality switch (Oviatt et al., 2000)
• Speak-and-spell (Filisko Seneff, 2004)
• Task errors
• Detection Recovery
• System mistakes (Horvitz Paek, 1999)
• User mistakes (Garland, Lesh, Rich, 2003)

6
Issues in SDS

• Interacting information sources
• Speech repairs (Heeman Allen, 1999)
• Dialogue act tagging (Stolcke et al., 2000)
• Speak and spell (Filisko Seneff, 2004)
• Portability
• Hub-based architectures (Allen et al., 2000
  Pellom, Ward, Pradhan, 2000)
• Task independent components/models
• PARADISE framework (Walker et al., 2000 Walker
  et al., 2001)
• Dialogue managers (Bohus Rudnicky, 2003)
• Monologue trained components (speech repairs)

7
Issues in SDS (continued)

• Computational issues
• Limits the complexity of models used
• Buying time
• Semi-supervised offline analysis of system
  behavior
• Standards
• Tools

8
Theoretical models for NLP
9
Maximum Entropy

• Introduction
• Constraints feature functions
• Assume nothing else principle
• Generalized Iterative Scaling algorithm
• Joint and conditional distributions

10
Maximum Entropy (continued)

• Advantages
• Combine multiple knowledge sources
• Local
• Word prefix, suffix, capitalization (POS -
  (Ratnaparkhi, 1996))
• Word POS, POS class, suffix (WSD - (Chao Dyer,
  2002))
• Token prefix, suffix, capitalization,
  abbreviation (Sentence Boundary - (Reynar
  Ratnaparkhi, 1997))
• Global
• N-grams (Rosenfeld, 1997)
• Word window
• Document title (Pakhomov, 2002)
• Structurally related words (Chao Dyer, 2002)
• Sentence length, conventional lexicon (Och Ney,
  2002)
• Combine dependent knowledge sources

11
Maximum Entropy (continued)

• Advantages
• Add additional knowledge sources
• Implicit smoothing
• Disadvantages
• Computational
• Expected value at each iteration
• Normalizing constant
• Overfitting
• Feature selection
• Cutoffs
• Basic Feature Selection (Berger et al., 1996)

12
Bayesian Belief Networks

• Take advantage of independencies
• Network structure design
• Hand coded
• DA tagging (Keizer, Akker, Nijholt, 2002)
• Dialogue modeling (Wai, Meng, Pieraccini, 2001
  Horvitz Paek, 1999)
• Argument generation (Zukerman, McConachy, Korb,
  1998)
• Human sentence processing (Narayanan Jurafsky,
  1998)
• External resources
• WordNet Selectional Restriction, WSD, QA
• Adjacency pairs (Galley, McKeown, Hirschberg,
  Shriberg, 2004)
• Learned from data

13
BBN (continued)

• Parameters
• Hand-coded
• Explain away effect (noisy-OR nodes)
• Learned from data
• Computational issues
• Restrict network size
• Focusing mechanism (Zukerman et al., 1998)

14
Hidden Markov Models

• Introduction
• Extensions
• Nth order HMMs
• Time duration HMMs, empty observation

15
HMMs (continued)

• State space design
• Classic view
• POS tags (Brants, 2000)
• Word classes (Law Chan, 1996)
• Sentence words (Vogel, Ney, Tillmann, 1996)
• Topic (Barzilay Lee, 2004)
• Structural states
• Text Chunking (Skut Brants, 1998)
• Named Entity Recognition (Bikel et al., 1999
  Zhou Su, 2002)

16
HMMs (continued)

• Transition and observation probabilities
• Hand coded
• Deterministic for structural states
• Learned from data
• Back-off smoothing deleted interpolation
• Smoothed counts
• Use sub-models
• Unknown observations
• Word suffix (Skut Brants, 1998)
• Special unknown observation (Bikel et al., 1999)
• Decoding
• Viterbi versus Beam-search

17
Extensions - MEMM

• HMMS
• Generative model
• Models joint distribution
• Can not handle multiple non-independent features
  of the current observation

• Maximum Entropy Markov Models
• Conditional model
• Models conditional distribution
• Can handle multiple non-independent features of
  the current observation

• Applications
• Text segmentation (McCallum, Freitag, Pereira
  2000)
• Stateless ME lt HMM lt Features HMM lt MEMM
• Restore capitalization (Chelba and Acero, 2004)

18
Extensions - CRF

• MEMM have the Label Bias Problem
• Solution Conditional Random Fields (Lafferty,
  McCallum, Pereira, 2001)
• Condition on the entire observation sequence
• Undirected graphical model

19
Extensions CRF (continued)

• Applications
• POS (Lafferty, McCallum, Pereira, 2001)
• Information Extraction (research papers) (Peng,
  McCallum 2004)
• Shallow Parsing (NP chunking) (Sha, Pereira 2003)
• Relational Markov Networks (Taskar, Abbeel,
  Koller 2002)
• Extension of CRF
• Model multiple type of entities and the
  relations between them
• Collective classification instead of individual
  classification

20
Recommendation Systems
21
Introduction

• Automatic information filtering
• Active user, Active item
• Ratings
• RS vs. Information Retrieval
• Classification
• Item-centered
• Memory-based
• Model-based
• User-centered
• Collaborative filtering (memory-based,
  model-based)
• Demographic
• Hybrid

22
Item-centered RS

• Memory-based
• Neighborhood (of items) formation
• Cosine similarity (Billsus Pazzani, 2000),
  adjusted cosine
• Correlation similarity (Pearson)
• Combining neighborhood (items) ratings
• Weighted sum (weightsimilarity)
• Linear regression (Sarwar et al., 2001)
• Model-based
• Mean-vector (Pazzani, 1999 Good et al., 1999)
• Naïve Bayes (Billsus Pazzani, 2000), Ripper
  (Good et al., 1999)
• Utility-based (Faltings et al. 2004 Jameson et
  al., 1995 Linden et al., 1997)
• Knowledge-based RS (Burke, 2001, 2002)

Content based
23
User-centered RS

• Collaborative filtering (CF) (Goldberg et al.,
  1992)
• Memory-based
• Neighborhood (of users) formation
• Cosine similarity, Pearson coefficient
• Default voting, Inverse user frequency (Breese et
  al., 1998)
• Combining neighborhood (users) ratings
• Weighted sum
• Significance weighting, rating normalization
  (Herlocker et al., 1999)
• Case amplification (Breese et al., 1998)
• Computational cost

24
User-centered RS (continued)

• Collaborative filtering (continued)
• Model-based
• BBN (Breese et al., 1998)
• Clustering users (Breese et al., 1998),
  usersitems (Ungar Foster, 1998)
• Cast as a machine learning problem (Billsus
  Pazzani, 1998)
• Demographic
• Demographic information (Krulwich, 1997)
• Users webpage (Pazzani, 1999)

25
Hybrid RS

• Complementarity between CF and content-based RS
• Complex ratings, new item problem, serendipitous
  predictions
• Hybrid RS
• Pseudo users Filterbots
• General (Sarwar et al., 1998)
• Personal (Good et al., 1999) mean-vector based
• Collaboration via content (Pazzani, 1999)
• Item-similarity (O'Sullivan et al. 2004)
• CF in knowledge-based RS (Burke, 2002)
• Latent variable models (Schein, Popescul, Ungar,
  Pennock, 2002)
• Ensemble methods (Pazzani, 1999)

26
Issues in RS

• Cold-start problem
• New user problem
• Utility-based, knowledge-based, demographic
• CF What items to rate?
• A gauge set (Goldberg et al., 2001)
• Specific items PopularityEntropy (Rashid et
  al., 2002)
• New item problem
• Content-based, hybrid

27
Issues in RS (continued)

• Evaluation
• Baseline
• Cross-validation
• Temporal issues
• Training data size
• Metrics
• Mean absolute error
• List of results
• Precision, Recall, F-measure
• Expected utility of a ranked list
• Coverage
• ROC curve

28
Issues in RS (continued)

• Interaction with the user
• Interests change over time (Billsus Pazzani,
  1998)
• Interests change from interaction with the data
  (Hirashima et al. 1997)
• Implicit interest indicators
• Web-browsing (Claypool et al. 2001)
• TV programs (O'Sullivan et al., 2004)
• Presentation media (Billsus Pazzani, 2000)
• User involvement in correcting profiles (Wærn,
  2004)
• Interaction principles in utility-based RS

29
Issues in RS (continued)

• Domain characteristics
• Item properties
• Dynamic domains
• Domains with very similar items
• Presentation media (what to present)
• Hypermedia (Mobasher, Dai, Luo, Nakagawa, 2002)
• Computational issues
• PCA based RS (Goldberg et al., 2001)
• Distribute load between server and client
• Security/Privacy