advertisement

personage

50 %
50 %
advertisement
Information about personage
Entertainment

Published on August 11, 2007

Author: GenX

Source: authorstream.com

advertisement

Generating Language with Personality for Dialogue:  Generating Language with Personality for Dialogue Prof. Marilyn Walker (Mairesse andamp; Walker, To Appear in ACL2007) University of Sheffield 18 April 2007 Why generate language with personality?:  Why generate language with personality? Cognitive Science: Can we computationally model findings from personality and social psychology w.r.t. naturally occurring conversation and text genres? Manipulate a computational artifact to reproduce and explore cross-domain generalizability of these findings? Human Computer Interaction / Dialogue Systems: Similarity attraction: holds in HCI as well as human-human and affects all aspects of user satisfaction (Nass andamp;Lee 2001) Linguistic style (politeness) improves learning outcomes in intelligent tutoring (Poraysta-Pomsta andamp;Mellish04; Wangetal, 04) Extraverts prefer ECA with social chitchat (Bickmoreandamp;Cassell02) System has a personality whether you design it in or not; Possible Applications:  Possible Applications Interactive Narrative and Computer Gaming Characters (Lester etal 1999, Mateas 2002) Intelligent Tutoring Systems (Poraysta-Pomsta and Mellish; Wang etal, 2004) Dialogue Systems (Nass and Lee 2001) Virtual Training Environments (Johnson etal 2004) Scripted Dialogue: Edutainment, Sales and Marketing (Walker etal, 1997; Andre etal 2002) Outline:  Outline Why model personality-based language variation? A bit of personality psychology Psychology findings on the linguistic markers of personality PERSONAGE: a personality generator Evaluation Experiment Statistical modelling for continuous variation Conclusion Personality Psychology:  Personality Psychology Personality The complex of all the attributes − behavioural, temperamental, emotional and mental − that characterize a unique individual. Do personality traits exist? Created to maintain an illusion of consistency in the world?  Find the most essential independent traits Lexical Hypothesis Important traits are encoded in the language (Allport andamp; Odbert, 1936) Participants describe people using a pool of adjectives Factor analysis The Big Five Personality Traits:  The Big Five Personality Traits Factor analysis  5 dimensions (Norman, 1963) Extraversion Sociability, assertiveness vs. quietness Emotional stability Calmness vs. neuroticism, anxiety Agreeableness Kindness vs. unfriendliness Conscientiousness Need for achievement, organization vs. impulsiveness Openness to experience Imagination, insight vs. conventionality Biological Causes:  Biological Causes Eysenck (1985) Introverts are more aroused than extravert, i.e. more stimuli reach the brain More sensitive to volume and lemon juice! Neuroticism is related to the activation threshold of the sympathetic nervous system More easily into ‘fight or flight’ state Increase of heart rate, sweating, muscular tension, etc. Genetically inherited (partially) Monozygotic vs. dizygotic twins (Henderson, 1982) Correlation for extraversion: 0.51 vs. 0.18 Correlation for neuroticism: 0.46 vs. 0.20 Some Linguistic Reflexes of Personality:  Some Linguistic Reflexes of Personality Extraversion (Furnham, 1990, Scherer1979) Talk more, faster, louder and more repetitively Fewer pauses and hesitations Lower type/token ratio Less formal, more references to context (Heylighen andamp; Dewaele, 2002) More positive emotion words (Pennebaker andamp; King, 1999) E.g. happy, pretty, good Neuroticism (Pennebaker andamp; King, 1999) 1st person singular pronouns Negative emotion words Conscientiousness (Pennebaker andamp; King, 1999) Fewer negations and negative emotion words Low but significant correlations Personality in Language: Ear Snippets:  Personality in Language: Ear Snippets Example (Mehl et al., 2006) Personality in Language: Stream of Consciousness Essays:  Personality in Language: Stream of Consciousness Essays Example (Pennebaker et al., 2003) So how can we use these findings? :  So how can we use these findings? Research Challenges:  Research Challenges Can linguistic variables observed in naturally occurring genres like SOC and EAR be mapped to parameters in a spoken language generator? Will utterances generated using these parameters be recognisable as manifesting personality types by humans In a very specific domain as in typical dialogue systems? Within a single utterance? Given the weak correlations previously observed What types of speech acts in particular domains manifest personality? Can language manifesting all Big Five personality traits be generated? PERSONAGE’s Framework:  PERSONAGE’s Framework Hypothesis that evaluative utterances in fixed domains will manifest personality: recommendations,comparisons Develop initial generator based on SPARKY statistical trainable generator in restaurant domain (Walker, Rambow and Rogati 2002, Stent etal 2004, Mairesse and Walker 2005) Extraversion: important trait, many findings Hypothesize a mapping from psychology findings to parameters in NLG standard architecture Extend SPARKY by implementing those parameters Explore two generation methods: Direct generation: use parameter values suggested by literature to generate extremes Overgenerate and rank: generate randomly and train statistical ranking model Evaluate Variation in the NLG Pipeline Architecture:  Variation in the NLG Pipeline Architecture What to say Content Planner How to Say It Sentence Planner Surface Realizer Prosody Assigner What is Heard Speech Synthesizer Dialogue Manager Spoken Language Generation Speech Synthesis Variation in the NLG Pipeline Architecture:  Variation in the NLG Pipeline Architecture What to say Content Planner How to Say It Sentence Planner Surface Realizer Prosody Assigner What is Heard Speech Synthesizer Dialogue Manager Spoken Language Generation Speech Synthesis Parametrized Pragmatic Variation Related Work:  Related Work Stylistic generation (Hovy88, DiMarcoandamp;Hirst90, Inkpenandamp;Hirst04, Paivaandamp;Evans04) Bandamp;L Politeness in generation (Walker,Cahnandamp;Whittaker97, Bickmoreandamp;Cassell02, Poraysta-Pomstaandamp;Mellish04; Johnsonetal04, Wangetal04, Rehmandamp;Andre07) Believable agents (Loyallandamp;Bates97, Lesteretal99, Mateas02) Personality in Agents/ECAs (Reevesandamp;Nass96, Andreetal00, Nassandamp;Lee01, Isbisterandamp;Nass00, Piwek03, Isardetal06) Individual Linguistic Adaptation (Coulston,Darvesandamp;Oviatt02, Wardandamp;Nakagawa02, Mairesseandamp;Walker05, Belz06, Lin06, Reiteretal05) Statistical Generation Models with Corpus-Based LM ranking (Langkildeandamp;Knight98,Isardetal06, Belz06, Lin06) Limitations of previous work:  Limitations of previous work While previous work suggests a clear utility for individual differences/personality/politeness/pragmatic variation in generation: There has been no systematic exploration of personality parameters suggested by psycholinguistic findings; Unclear which findings have impact in particular application domains; Very little evaluation to date showing it is possible to produce recognizable personality variation. PERSONAGE’s Architecture:  PERSONAGE’s Architecture Restaurant attributes E.g. food = 4/5 Content Planner Syntactic Structure Selection Pragmatic Structure Transformation Lexical Choice Realization Generation dictionary 29 Parameters E.g. Verbosity = 0.9 WordNet (Fellbaum, 1998) RealPro (Lavoie and Rambow, 1997) Output utterance Aggregation SPaRKy (Stent et al., 2004) Slide19:  Gricean Intention/ Communicative Goal: Recommend Le Marais:  Gricean Intention/ Communicative Goal: Recommend Le Marais Phase 1: Content Planning and Aggregation:  Phase 1: Content Planning and Aggregation Sentence Plan Tree Content Plan Tree Clause combining operation Content Planner Content item selection Relation insertion Content item ordering Aggregation Selection of clause combining operations Generation Decisions: Content Planning:  Generation Decisions: Content Planning Verbosity Content items selection, e.g. food quality, price, service, etc. Choice of content based on polarity Zagat scalar ratings, e.g. food = 2/5, service = 5/5 Insertion of restatements/repetitions Small generation dictionary WordNet based paraphrasing, e.g. 'the food is awful, terrible' Concession of content items with different polarity E.g. 'Even if Wok Mania has awful food, it’s cheap' Position of positive content Generation Decisions: Aggregation operations:  Generation Decisions: Aggregation operations Many ways to combine information (Scott and DeSouza90, Wang and Hovy 92, Stent et al. 2004) Justification 'since', 'because', 'so', etc. Contrast 'but', 'however', 'on the other hand', etc. Inference Relative clause, e.g. 'X, which has good food, is …' Merge, e.g. 'X has good food and decent service' Conjunction, period, etc. Concession 'Even if X has awful food, …' 'Although X has awful food, …' Restatement Comma, e.g. 'X has awful food, it has bad food' Merge, e.g. 'X has awful food, bad food' Phase 2: Syntactic Template Selection:  Phase 2: Syntactic Template Selection Generation dictionary Content item  Syntactic Structures Sentence Plan Tree Wok Mania class: proper noun number: sg have class: verb awful class:adjective food class: noun number: sg article: none Obj Subj ATTR assert-reco- food-quality(Wok Mania, 0.1) Phase 2:Syntactic Template Selection:  Phase 2: Syntactic Template Selection Need to choose from different syntactic templates Each template is associated to a point in a feature space E.g. syntactic complexity, self-reference, positive connotation Select the closest claim template to the input target values Self references Syntactic complexity (e.g. depth) 0 1 2 0.0 0.5 1.0 X is the best I am sure you would like X target Phase 3: Pragmatic Transformations:  Phase 3: Pragmatic Transformations Sequential modifications of the syntactic tree Subject implicitness 'X has awful food'  'The food is awful' Negation insertion WordNet antonyms of adjectives 'X has good food'  'X doesn’t have bad food' Hedge insertion Downtoners: 'kind of', 'quite', 'It seems that', 'around', etc. Fillers: 'err…', 'I mean', 'like' Acknowledgments: 'right', 'I see', 'well', 'yeah' Emphasizers: 'really', 'basically', 'you know', exclamation, etc. Tag question insertion 'X has good food, doesn’t it?' Example of Pragmatic Transformation:  Example of Pragmatic Transformation Negation insertion 'X has awful food'  'X doesn’t have good food' Wok Mania class: proper noun number: sg have class: verb awful class:adjective food class: noun number: sg article: none Obj Subj ATTR WordNet Database Look for antonym 'good' - Negate verb - Replace adjective by antonym Wok Mania class: proper noun number: sg have class: verb negated: true good class:adjective food class: noun number: sg article: none Obj Subj ATTR Phase 4: Lexical Choice:  Phase 4: Lexical Choice Need to choose from WordNet synonyms Each word is associated to a point in a feature space E.g. word frequency of use, word size, etc. Machine-readable dictionary Select the closest synonym to the input target values Word size Normalized frequency 1 2 3 4 5 6 7 8 9 10 11 0.0 0.5 1.0 cheap inexpensive target PERSONAGE Outputs: `Recommend Le Marais’:  PERSONAGE Outputs: `Recommend Le Marais’ Evaluation Experiment:  Evaluation Experiment Two methods for generation: Direct Control: Use parameters suggested by psycholinguistic findings; Collect judges’ ratings on extraversion dimension; Evaluate if judges perceive extraversion as predicted. Overgenerate and Rank: Generate many alternatives with random parameter settings; Learning ranking model from judges’ ratings; Learn extraversion parameters from judges’ ratings. Evaluation Experiment (2):  Evaluation Experiment (2) 3 judges: anthropologist, historian, psychologist Training: Examine adjectives associated with Big Five traits (Goldberg 1990) TIPI (Ten Item Personality Inventory; Gosling03) =andgt; Extraversion score between 1 (introvert) and 7 (extravert) Naturalness: could have been said by human (1...7) 10 utterances for 20 content plans (200 total) 2 generated with introversion parameters 2 generated with extraversion parameters 6 generated with random parameters Direct Personality Generation:  Direct Personality Generation Generate extravert or introvert language with parameter settings from the psychology literature (80 out of 200) Evaluation: can judges recognise both types of language? Inter-rater correlations: 0.70, 0.74, and 0.76 Binary classification of output: 90% accuracy Significant at p andlt; .01 Predefined Parameter Sets:  Predefined Parameter Sets So where are we?:  So where are we? Can produce recognizable variation at extreme ends of extraversion scale using findings from psycholinguistics; Hypothesized mapping appears to be fairly successful; NOW: Can refine mapping and parameter values based on randomly generated utterances. Learn parameter settings:  Learn parameter settings What makes an utterance extraverted? Parameters can be refined based on correlation between generation decision and ratings Significant positive correlations over 120 random utterances (p andlt; .05) Learn Parameter Settings (2):  Learn Parameter Settings (2) Significant negative correlations (p andlt; .05) Slide37:  What if we want to produce language close to an arbitrary input personality value ? Overgenerate and Rank:  Overgenerate and Rank Generator 'WokMania is a great place, because the food is good, isn’t it?' Statistical Regression/Ranking Model Estimates scores from features, e.g. verbosity, hedges 'Yeah, even if WokMania is expensive, the food is nice, I am sure you would like WokMania!' 'Err… this restaurant’s not as bad the others.' … Input personality score e.g. 2.5 out of 7 Closest estimate, utterance 2: 'Err… this restaurant’s not as bad the others.' Feature vector 1 Feature vector 3 Feature vector 2 Feature vector n Trainable NLG:  Trainable NLG Where do we get the regression model from? Train the statistical model from any language source The ratings can reflect anything we want to control Utterance quality (SPOT, SPARKY) Perceived personality (PERSONAGE, also see Isard etal 06) Data Collection Text 1  Rating 1 Text 2  Rating 2 … Text n  Rating n Feature Extraction Text 1  Feature vector 1 Text 2  Feature vector 2 … Text n  Feature Vector n Function Approximation Find ‘simple’ F such as F(feature vector) = rating Can be learned statistically E.g. ranking models, decision tree Dataset Features + rating pairs Statistical model F(new features) = rating estimate Statistical Models of Personality Perception:  Statistical Models of Personality Perception Data: 120 random utterances rated by 3 judges Features based on generation decisions Results for predicting extraversion Best model: Support Vector Machines Correlation between rating and prediction: r = 0.5 Average prediction error on a scale from 1 to 7: absolute error = 0.9 Better than inter-rater correlations: 0.31, 0.41, and 0.44 Less natural than extreme utterances E.g. regression tree model Slide41:  Output Extraversion Score Overgenerate and Rank:Pros and Cons:  Overgenerate and Rank: Pros and Cons Pros Can use other features than generation decisions Emergent properties, e.g. word count, n-grams Psycholinguistic features using machine-readable dictionaries MRC Psycholinguistic database (Coltheart, 1981) General Inquirer (Stone et al., 1966) LIWC: Linguistic Inquiry and Word Count (Pennebaker and Francis, 2001) E.g. word familiarity, concreteness, ratio of positive words, etc. Cons Computation time Random parameter settings can produce inconsistencies E.g. positive claim supported by negative content items 'X is the best, it has bad food and awful service' Statistical model of naturalness? Research Challenges:  Research Challenges Can linguistic variables observed in naturally occurring genres like SOC and EAR be mapped to parameters in a spoken language generator? Will utterances generated using these parameters be recognisable as manifesting personality types by humans In a very specific domain as in typical dialogue systems? Within a single utterance? Given the weak correlations previously observed? What types of speech acts in particular domains manifest personality? Can language manifesting all Big Five personality traits be generated? Summary and Future Work:  Summary and Future Work Systematic framework for fully automatic generation of language manifesting personality exploiting findings from psycholinguistics Evaluation shows that PERSONAGE produces utterances with recognisable personality in a restricted domain Meaning?: Gricean/Discourse intention =andgt; ``Recommend Le Marais’’ Prosodic parameters? Introvert (Alt 5) 'Err... it seems to me that Le Marais isn’t as bad as the others.' Extravert (Alt 2) 'Basically, actually, I am sure you would like Le Marais. It features friendly service and acceptable atmosphere and it’s a French, kosher and steak house place. Even if its price is 44 dollars, it just has really good food, nice food.' Solving for parameter values and re-generate (Paivaandamp;Evans04); Extend to other Big Five traits (neuroticism, agreeableness, conscientiousness, openness to experience) Applications in interactive narrative, dialogue systems, computer gaming, intelligent tutoring systems

Add a comment

Related presentations

Related pages

Dramatis personae – Wikipedia

Dramatis personæ (lat. „Personen der Handlung“) bezeichnet die handelnden Figuren in einem Bühnenstück. Der ersten Szene eines Theatertextes werden ...
Read more

Personage - Wikipedia

Een personage is de uitbeelding van een persoon in een verhalend of dramatisch kunstwerk zoals een roman, toneelstuk of film. Personages ondergaan of ...
Read more

Personage - Ein hypersoziales Storytelling Experiment im Web

Ein hypersoziales Storytelling Experiment im Web. Personage verbindet die Kanäle des Social Webs um so eine fiktive Story zu erzählen. Verfolge die ...
Read more

dict.cc | personage | Wörterbuch Englisch-Deutsch

Übersetzung für personage im Englisch-Deutsch-Wörterbuch dict.cc.
Read more

personage - Deutsch-Übersetzung - bab.la Englisch-Deutsch ...

Übersetzung für 'personage' im kostenlosen Deutsch-Wörterbuch. Weitere Deutsch-Übersetzungen für: male personage, royal personage
Read more

Personage | Define Personage at Dictionary.com

Personage definition, a person of distinction or importance. See more.
Read more

Personage | Definition of Personage by Merriam-Webster

The premiere was attended by no less a personage than the president himself.
Read more

Persorange | persOrange GmbH – Personalberater ...

Personalvermittlung persOrange GmbH – langjährige Erfahrung und weiterentwickelte Technologien für Personalverantwortliche mit individuell ...
Read more

personage - Übersetzung Englisch-Deutsch - duden.de

personage, Übersetzung, Wörterbuch, Englisch-Deutsch, Duden ... Duden online bietet Ihnen umfassende Informationen zu Rechtschreibung, Grammatik und ...
Read more

dict.cc Wörterbuch :: personage :: Deutsch-Englisch ...

Englisch-Deutsch-Übersetzung für personage im Online-Wörterbuch dict.cc (Deutschwörterbuch).
Read more