Speech and Language Processing for Adaptive Training Diane Litman Professor, Computer Science Department Senior Scientist, Learning Research & Development Center University of Pittsburgh
Outline The State of the Art: A Brief Survey Adaptation to Student Uncertainty in Tutorial Dialogue: A Case Study – ITSPOKE: System and Corpora – Uncertainty Detection – System Adaptation – Experimental Evaluation Summing Up
What is Natural Language Processing? “The goal of this new field is to get computers to perform useful tasks involving human language, tasks like enabling human-machine communication, improving human-human communication, or simply doing useful processing of text or speech.” [Jurafsky and Martin 2008] Many names and facets –Speech and Language Processing –Human Language Technology –Computational Linguistics
Relevance for Adaptive Training Knowledge of Language is often needed to –trigger adaptation –personalize training, using the enormous amount of machine-readable text and audio that is now available Conversational Agents are becoming an important form of human-computer interaction
Knowledge of Language Phonetics and Phonology: speech sounds Morphology: words and their internal composition Syntax: the structuring of words into larger units Semantics: the meaning of words and larger units Pragmatics: interpretation in situational context Discourse: interpretation in context of previous utterances
Computational Models (and Associated Algorithms) State Machines Formal Rule Systems / Grammars Logic-Based Formalisms Models of Uncertainty
A Brief Survey of Applications NLP Applications to Education
A Brief Survey of Applications NLP Applications to Education Learning Language (reading, writing, speaking) Tutors Scoring
A Brief Survey of Applications NLP Applications to Education Learning Language (reading, writing, speaking) Using Language (to teach everything else) Tutors Scoring Conversational Tutors / Peers CSCL
A Brief Survey of Applications NLP Applications to Education Learning Language (reading, writing, speaking) Using Language (to teach everything else) Tutors Scoring Readability Processing Language Conversational Tutors / Peers CSCL Discourse Coding Lecture Retrieval Questioning & Answering
Outline The State of the Art: A Brief Survey Adaptation to Student Uncertainty in Tutorial Dialogue: A Case Study – ITSPOKE: System and Corpora – Uncertainty Detection – System Adaptation – Experimental Evaluation Summing Up
Tutorial Dialogue Systems Why is one-on-one tutoring so effective? “...there is something about discourse and natural language (as opposed to sophisticated pedagogical strategies) that explains the effectiveness of unaccomplished human [tutors].” [Graesser, Person et al. 2001] Currently only humans use full-fledged natural language dialogue
Spoken Tutorial Dialogue Systems Most human tutoring involves face-to-face spoken interaction, while most computer dialogue tutors are text-based Can the effectiveness of dialogue tutorial systems be further increased by using spoken interactions?
Potential Benefits of Spoken Dialogue: I Conversation provides a learning environment that promotes student activity Self-explanation correlates with learning and occurs more in speech
Potential Benefits of Spoken Dialogue: II Speech contains prosodic information, providing new sources of information about the student for adaptation A correct but uncertain student turn –ITSPOKE: How does his velocity compare to that of his keys? –STUDENT: his velocity is constant
Potential Benefits of Spoken Dialogue: III Spoken computational environments may foster social relationships that may enhance learning
Potential Benefits of Spoken Dialogue : IV Some applications inherently involve spoken language –Conversational skill training Others require hands-free interaction –e.g., NASA
Outline The State of the Art: A Brief Survey Adaptation to Student Uncertainty in Tutorial Dialogue: A Case Study – ITSPOKE: System and Corpora – Uncertainty Detection – System Adaptation – Experimental Evaluation Summing Up
Back-end is Why2-Atlas system [VanLehn, Jordan, Rose et al. 2002] Sphinx2 speech recognition and Cepstral text-to-speech
Back-end is Why2-Atlas system [VanLehn, Jordan, Rose et al. 2002] Sphinx2 speech recognition and Cepstral text-to-speech
Back-end is Why2-Atlas system [VanLehn, Jordan, Rose et al. 2002] Sphinx2 speech recognition and Cepstral text-to-speech
Three Types of Tutoring Corpora Human Tutoring –14 students / 128 dialogues (physics problems) –5948 student turns, 5505 tutor turns Computer Tutoring –77 students / 385 dialogues –both synthesized and pre-recorded tutor voices Wizard /Computer Tutoring – 81 students / 405 dialogues – human performs speech recognition, semantic analysis – computer performs dialogue management
Experimental Procedure College students without physics –Read a small background document –Took a multiple-choice Pretest –Worked 5-10 problems (dialogues) with tutor –Took an isomorphic Posttest Goal was to optimize Learning Gain – e.g., Posttest – Pretest
Outline The State of the Art: A Brief Survey Adaptation to Student Uncertainty in Tutorial Dialogue: A Case Study – ITSPOKE: System and Corpora – Uncertainty Detection – System Adaptation – Experimental Evaluation Summing Up
Standard Empirical Detection Methodology Manual annotation of user states that will trigger system adaptation –Naturally-occurring spoken dialogue data Prediction via machine learning –Use speech and language processing to automatically extract features from user turns –Use extracted features and annotations to learn a model for predicting user state(s) in new data –Significant reduction of baseline error
Example Features What a user says –words (speech recognition), stems (morphology) –part-of-speech, syntactic constituents (parsing) –correctness (semantic analysis) –dialogue moves (pragmatics and discourse) How a user says it –acoustic-prosodic analysis
Extracting Pitch Features
Extracting Energy Features
Temporal Features Duration = end time - begin time Tempo (speaking rate) = #syllables/duration
Detecting Neg/Pos/Neu in ITSPOKE - Baseline Accuracy via Majority Class Prediction
Detecting Neg/Pos/Neu in ITSPOKE -Use of prosodic (sp), recognized (asr) and/or actual (lex) lexical features outperforms baseline
Detecting Neg/Pos/Neu in ITSPOKE -As with other applications, highest predictive accuracies are obtained by combining multiple feature types
Outline The State of the Art: A Brief Survey Adaptation to Student Uncertainty in Tutorial Dialogue: A Case Study – ITSPOKE: System and Corpora – Uncertainty Detection – System Adaptation – Experimental Evaluation Summing Up
System Adaptation: How to Respond? Our initial focus: responding to student uncertainty –Most frequent user state in our data –Focus of other studies –.62 Kappa Approaches to adaptive system design –Theory-based –Data-driven
Theory-Based Adaptation: Uncertainty as Learning Opportunity Uncertainty represents one type of learning impasse, and is also associated with cognitive disequilibrium – An impasse motivates a student to take an active role in constructing a better understanding of the principle. [VanLehn et al. 2003] –A state of failed expectations causing deliberation aimed at restoring equilibrium. [Craig et al. 2004] Hypothesis: The system should adapt to uncertainty in the same way it responds to other impasses (e.g, incorrectness)
Data-Driven Adaptation: How Do Human Tutors Respond? An empirical method for designing dialogue systems adaptive to student state –extraction of “dialogue bigrams” from annotated human tutoring corpora –χ 2 analysis to identify dependent bigrams –generalizable to any domain with corpora labeled for user state and system response
Example Human Tutoring Excerpt S: So the- when you throw it up the acceleration will stay the same? [Uncertain] T: Acceleration uh will always be the same because there is- that is being caused by force of gravity which is not changing. [Restatement, Expansion] S: mm-k. [Neutral] T: Acceleration is– it is in- what is the direction uh of this acceleration- acceleration due to gravity? [Short Answer Question] S: It’s- the direction- it’s downward. [Certain] T: Yes, it’s vertically down. [Positive Feedback, Restatement]
Findings Statistically significant dependencies exist between students’ state of certainty and the responses of an expert human tutor –After uncertain, tutor Bottoms Out and avoids expansions –After certain, tutor Restates –After mixed, tutor Hints –After any non-neutral, tutor increases Feedback Dependencies suggest adaptive strategies for implementation in computer tutoring systems
Outline The State of the Art: A Brief Survey Adaptation to Student Uncertainty in Tutorial Dialogue: A Case Study – ITSPOKE: System and Corpora – Uncertainty Detection – System Adaptation – Experimental Evaluation Summing Up
Experimental Design: 4 Conditions Manipulate tutor responses to student uncertainty and investigate impact on learning and efficiency Experimental-Basic: treat all uncertain turns as incorrect (theory) Experimental-Empirical: for uncertain or incorrect turns, provide original content but vary dialogue act (human tutor analysis) Control-Norm: ignore uncertainty (as in original system) Control-Random: ignore uncertainty, but treat a percentage of random correct answers as incorrect (to control for additional tutoring)
TUTOR: Now let’s talk about the net force exerted on the truck. By the same reasoning that we used for the car, what’s the overall net force on the truck equal to? STUDENT: The force of the car hitting it? [uncertain+correct] TUTOR (Control-Norm): Good [Feedback] … [moves on] TUTOR (Experimental-Basic): Fine. [Feedback] We can derive the net force on the truck by summing the individual forces on it, just like we did for the car. First, what horizontal force is exerted on the truck during the collision? [Remediation Subdialogue] –Same tutor response if student had been incorrect Treatments in Different Conditions
Platform: Adaptive WOZ-TUT System Modified version of ITSPOKE –Dialogue manager adapts to uncertainty »system responses based on combined uncertainty and correctness –Full automation replaced by some Wizard of Oz (WOZ) components »human wizard recognizes student speech »human also annotates uncertainty and correctness »provides upper-bound speech and NLP performance
WOZ-TUT Screenshot
Experimental Procedure subjects in each condition –Native English speakers with no college physics –Procedure: 1) read background material, 2) took pretest, 3) worked training problem with WOZ-TUT, 4) took user Brief Survey, 5) took posttest
Experimental Results Two-way ANOVA indicated students learned (F(1,77) = , p = 0.000, MSe = 0.009) Amount depended on condition (F(3,77) = 3.275, p = 0.025, MSe = 0.009) One-way ANOVA with post-hoc Tukey tests determined which conditions learned more
Experimental Results Two-way ANOVA indicated students learned (F(1,77) = , p = 0.000, MSe = 0.009) Amount depended on condition (F(3,77) = 3.275, p = 0.025, MSe = 0.009) One-way ANOVA with post-hoc Tukey tests determined which conditions learned more
In Addition… Learning Efficiency also improved –Two Efficiency Measures »(Normalized Learning Gains) / (Total Student Turns) »(Normalized Learning Gains) / (Total Time in Minutes) –Experimental-Basic > Control-Norm (p <.05) Current Directions –New evaluation of Experimental-Basic »fully-automated ITSPOKE –New methods for designing Experimental-Empirical »educational data mining using reinforcement learning –Other student states
Outline The State of the Art: A Brief Survey Adaptation to Student Uncertainty in Tutorial Dialogue: A Case Study – ITSPOKE: System and Corpora – Uncertainty Detection – System Adaptation – Experimental Evaluation Summing Up
Summing Up: I Spoken Dialogue Systems for Adaptive Training –Natural language dialogue is a key aspect of human one- on-one training –Using presently available technology, successful conversational computer training environments are now being built –Evidence that more adaptive versions of such systems will further enhance performance
Summing Up: II Adaptive Training in turn provides many other opportunities and challenges for researchers in Speech and Natural Language Processing
Acknowledgements ITSPOKE group past and present –Hua Ai, Min Chi, Joanna Drummond, Kate Forbes-Riley, Alison Huettner, Michael Lipschultz, Beatriz Maeireizo-Tokeshi, Greg Nicholas, Amruta Purandare, Mihai Rotaru, Scott Silliman, Joel Tetreault, Art Ward –Columbia Collaborators: Julia Hirschberg, Jackson Liscombe, Jennifer Venditti –Jan Wiebe, Rebecca Hwa, Wendy Chapman, Paul Hoffmann, Behrang Mohit, Carol Nichols, Swapna Somasundaran, Theresa Wilson, Chenhai Xi Why2-Atlas and Human Tutoring groups –Kurt Vanlehn, Pamela Jordan, Uma Pappuswamy, Carolyn Rose –Micki Chi, Scotty Craig, Bob Hausmann, Margueritte Roy
Thank You! Questions? Further Information –
The End
Overview: Towards Adaptive Spoken Dialogue Systems User State Detection u Promising across user states and applications, e.g.: u Craig et al., 2006 u Litman & Forbes-Riley, 2006 u Lee & Narayanan, 2005 u Vidrascu & Devillers, 2005 u Batliner et al., 2003 Adaptation u Sparse, can be difficult to show adaptation improves performance u Some used basic adaptations and showed likeability increases u For other performance metrics, basic adaptations not clear a priori System Adaptation Health Assessment Stress Empathy [Liu & Picard 2005] Gaming Frustration Apology [Klein et al. 2002] Application Tutoring ??????????????
Detecting and Responding to Student States Opportunity –Adaptive spoken dialogue system technology can improve student learning and other measures of performance [Aist et al. 2002; Pon-Barry et al. 2006] Challenges – What to detect? – How to respond? – Evaluation?
Example Student States in ITSPOKE ITSPOKE: What else do you need to know to find the box‘s acceleration? Student: the direction [UNCERTAIN] ITSPOKE : If you see a body accelerate, what caused that acceleration? Student: force [CERTAIN] ITSPOKE : Good job. Say there is only one force acting on the box. How is this force, the box's mass, and its acceleration related? Student: velocity [UNCERTAIN] ITSPOKE : Could you please repeat that? Student: velocity [ANNOYED]
Prediction via Machine Learning Multiple feature types per student turn, e.g. –Acoustic-prosodic –Lexical –Identifiers –System and student performance Sample research questions –Relative utility of feature types –Impact of speech recognition –Speaker and task dependence –Impact of learning algorithm, amount of training data
Detecting Neg/Pos/Neu in ITSPOKE - However, relative feature utility differs in tutoring (e.g., for speech features: temporal > energy > pitch)
Example from ITSPOKE ITSPOKE: What is the net force acting on the truck equal to? Student: I don’t know [Uncertain, Frustrated] ITSPOKE: Since net force is defined as the vector sum of all the forces exerted on an object, let's first find the forces acting on the truck. Try to name them. Student: The impact force, the normal force, and gravity [Certain, Neutral] ITSPOKE: Yep.
Approaches to Evaluation “Correlational” Studies, e.g. –Student uncertainty positively correlates with learning [Craig et al. 2004] –Adding uncertainty and frustration metrics to regression models increases model fit [Forbes-Riley et al. 2008] “Causal” Studies, e.g. –Adding human-provided emotional scaffolding to a reading tutor increases student persistence [Aist et al. 2002] –Experimentally manipulate tutor responses to student uncertainty and investigate impact on learning [Pon-Barry et al. 2006]
Treatments in Different Conditions TUTOR: Now let’s talk about the net force exerted on the truck. By the same reasoning that we used for the car, what’s the overall net force on the truck equal to? STUDENT: The force of the car hitting it? [uncertain+correct] TUTOR (Control-Norm): Good [Feedback] … [moves on] TUTOR (Experimental-Empirical): That’s exactly right, but you seem unsure, so let’s sum up. [Feedback] The net force on the truck is equal to the impact force on it… [New Bottom Out] –New tutor responses for incorrect +/- uncertainty answers as well
In Closing Synergy between Intelligent Tutoring and Spoken Dialogue Systems can provide –Better scientific understanding of how dialogue facilitates learning –Long-term benefit for scaling spoken dialogue systems to new and complex domains
What is Tutoring? “A one-on-one dialogue between a teacher and a student for the purpose of helping the student learn something.” [Evens and Michael 2006] Human Tutoring Excerpt [Thanks to Natalie Person and Lindsay Sears, Rhodes College]
Intelligent Tutoring Systems Students who receive one-on-one instruction perform as well as the top two percent of students who receive traditional classroom instruction [Bloom 1984] Unfortunately, providing every student with a personal human tutor is infeasible – Develop computer tutors instead
Monitoring Student State (motivation) Tutor26: Let me ask my question in a more general way. What happens to the velocity of a body when there is no force acting on it? Student27:dammit (ASR: it is) Tutor28 :Could you please repeat that? Student29 :same (ASR: i same) Tutor30 :Right. And what happens to the HORIZONTAL velocity of a body when there is no HORIZONTAL force acting on it? Student31:zero (ASR: the zero) Tutor32 :Well... OK, let's consider how forces in the vertical direction affect the motion of a body in a horizontal direction in a different scenario Student33:oh gravity you already said this (ASR: beats gravity you're exert less) Tutor34:Fine. Are there any other forces acting on the apple as it falls? Student35:no why are you doing this again (ASR: no y and to it yes) Tutor36:Correct. The only force acting on the apple as it falls to the earth is the earth's gravitational force. In which direction does gravitational force act? Student37:downward you computer (ASR: downward you computer)
What to Annotate? Information-Access and Customer Care Systems –Negative: Angry, Annoyed, Frustrated, Tired –Positive/Neutral: Amused, Cheerful, Delighted, Happy, Serious [Ang et al. 2002; Shafran et al. 2003; Lee and Narayanan 2005; Liscombe et al. 2005]
What to Annotate? Information-Access and Customer Care Systems –Negative: Angry, Annoyed, Frustrated, Tired –Positive/Neutral: Amused, Cheerful, Delighted, Happy, Serious [Ang et al. 2002; Shafran et al. 2003; Lee and Narayanan 2005; Liscombe et al. 2005] Tutorial Dialogue Systems –Negative: Angry, Annoyed, Frustrated, Bored, Confused, Uncertain, Contempt, Disgusted, Sad –Positive/Neutral: Certain, Curious, Enthusiastic, Eureka [Litman and Forbes-Riley 2006, D’Mello et al. 2006]
Theory-Based Adaptation In tutoring, not all negatively-valenced states are bad! –While frustration/anger/annoyance is often frustrating… –Frustration can also be an opportunity to learn Example from AutoTutor – neutral flow confusion frustration neutral [Thanks to Sidney D‘Mello and Arthur Graesser, University of Memphis]
Bigram Dependency Analysis EXPECTED Tutor IncludePos Tutor OmitsPos neutral certain uncertain mixed OBSERVED Tutor IncludesPos Tutor OmitsPos neutral certain uncertain mixed71161 χ2 = (critical χ2 value at p =.001 is 16.27) - “Student Certainness – Tutor Positive Feedback” Bigrams
Bigram Dependency Analysis (cont.) EXPECTED Includes Pos Omits Pos neutral OBSERVED Includes Pos Omits Pos neutral Less Tutor Positive Feedback after Student Neutral turns
Bigram Dependency Analysis (cont.) EXPECTED Includes Pos Omits Pos neutral certain uncertain mixed OBSERVED Includes Pos Omits Pos neutral certain uncertain mixed Less Tutor Positive Feedback after Student Neutral turns - More Tutor Positive Feedback after “Emotional” turns
Adaptation to Student Uncertainty: A First Evaluation Most systems respond only to (in)correctness Recall that literature suggests uncertain as well as incorrect student answers signal learning impasses Experimentally manipulate tutor responses to student uncertainty and investigate impact on learning
Experimental Design: 4 Conditions Experimental-Basic: treat all uncertain turns as incorrect Experimental-Empirical: for uncertain or incorrect turns –provide original content, but vary dialogue act (human tutor analysis) –provide additional feedback on uncertainty (beyond propositional content ) Control-Norm: ignore uncertainty (as in original system) Control-Random: ignore uncertainty, but treat a percentage of random correct answers as incorrect (to control for additional tutoring)