Unsupervised Counselor Dialogue Clustering for Positive Emotion - PowerPoint PPT Presentation

Unsupervised Counselor Dialogue Clustering for Positive Emotion Elicitation in Neural Dialogue System Nu Nurul l Lu Lubis is, Sakriani Sakti, Koichiro Yoshino, Satoshi Nakamura Information Science Division, Nara Institute of Science and Technology, JAPAN 13 July 2018 Nurul Lubis 1

Affective dialogue systems • Increase of dialogue system works and applications in various tasks involving affect High potential of dialogue • Companion for the elderly systems to address the [Miehle et al., 2017] • Distress clues assessment emotional needs of users [DeVault et al., 2014] • Affect-sensitive tutoring [Forbes-Riley and Litman, 2012] 13 July 2018 Nurul Lubis 2

Emotion elicitation Traditional emotion works Intent for emotion elicitation Expression Recognition Expression Emotion elici licitati tion: elici liciti ting g ch change of of emotion in in dia ialo logue • Using machine translation with target emotion (Hasegawa et al., 2013) • Using system’s affective personalities ( Skowron et al., 2013)  Have not yet considered the em emotional ben enefit it for the user 13 July 2018 Nurul Lubis 3

Positive emotion elicitation We aim to draw on an overlooked potential of emotion elicitation to im improve use ser emoti tional l states • A chat-based dialogue system with an implicit goal of posi sitiv ive emotio tion eli licit itatio ion Circumplex model of affect [Russell, 1980] 13 July 2018 Nurul Lubis 4

Different responses elicit different emotions  I failed the test. I failed the test. I failed the test. Oh, ag again in? You ou will ill do o better next xt tim time! Yeah… Thank you. arousal arousal valence valence Negative Positive Em Emoti tional impact 13 July 2018 Nurul Lubis 5

Neural chat-based dialogue system • RNN encoder-decoder [Vinyals et al., 2015] • Hierarchical recurrent encoder-decoder (HRED) [Serban et al., 2016] • Generating dialogue response with emotional expression [Zhou et al., 2018] Application towards emotion elicitation [Serban et al., 2016] is still very lacking. 13 July 2018 Nurul Lubis 6

Emotion-sensitive response generation: Emo-HRED [Lubis et al., 2018] in Proc. AAAI 2018 • Encodes emotional context and considers it in generating a response • Training data contains responses that elicit positive emotion • Significant improvement on perceived emotional impact Limit Lim itations 1. Has not yet learned strategies from an expert 2. Short and generic responses with positive-affect words 13 July 2018 Nurul Lubis 7

Challenge and proposal 2. 2. Go Goal: : in incr crease var ariety in in th the 1. 1. Go Goal: : Le Learnin ing elici licitation str trategy generated resp sponse to im improve from an fr an expert engagement • Challenge: Absence of data that shows • Challenge: Data sparsity • positive emotion elicitation in everyday • We hypothesize that higher level situations information, e.g. dialogue action, will • expert strategy in affective dialogue reduce data sparsity • Proposed: Construct a dialogue corpus • categorizing responses involving an expert in a positive • emphasizing this information in the emotion elicitation scenario training and generation process. 13 July 2018 Nurul Lubis 8

Proposed architecture: Multi context HRED (MC-HRED) A neural dialogue system which generate response based on multiple dialogue contexts • Dialogue history • User emotional state • Response action label MC-HRED architecture. 13 July 2018 Nurul Lubis 9

Corpus construction Positive Emotion Elicitation by an Expert 13 July 2018 Nurul Lubis 10

Data recording design • Goal: learn expert strategy for eliciting positive emotion • Collect: • Interaction between an exp xpert and a partic icipant • Condition the interaction with negative emotion • Expert guides the conversation to allow participant’s emotion recovery and reinstate positive emotion Emotion Briefing annotation Recorded Session Emotion induction Emotion processing Opening talk using video and recovery (neutral) (negative) (positive) 24-Jul-17 Nurul Lubis 11

Data collection and annotation • 60 sessions: 23 hours and 41 minutes of material • 1 counselor, 30 participants • 2 sessions per participant • 1 induced to anger • 1 induced to sadness • Self-report emotion annotation using Gtrace [Cowie et al., 2000] • Transcription 13 July 2018 Nurul Lubis 12

Unsupervised Clustering of Counselor Dialogue 13 July 2018 Nurul Lubis 13

Counselor dialogue clustering  Human annotation • Expensive, labor intensive Goal: To find high-level information • Low reliability • Information equivalent to dialogue acts  Standard dialogue acts classifier • Specific to the dialogue scenario • May not cover specific emotion-related • Retaining affective intents intent in the data  Unsupervised clustering 13 July 2018 Nurul Lubis 14

Counselor dialogue clustering K-means K-means clustering labels Pre-trained Vectorized Counselor Counseling counselor word2vec Counselor responses Counselor data dialogue model DPGMM responses DPGMM labels responses clustering K-Means DPGMM • Need to predefine how many • No prior definition of model clusters complexity • We choose K empirically 13 July 2018 Nurul Lubis 15

Analysis K-means: 8 clusters K-means: 8 sub-clusters DPGMM: 13 clusters “So you feel frustrated.” “Have you thought about this An assortment of “I guess we all have to be careful.” kind of issue before?” “Mm mm .” shorter sentences “Yes, hm mm” “So who do you think is “Maybe, yes yes .” “Mm mm .” An assortment of responsible for it?” “Right.” “Yes, hm mm” longer sentences 13 July 2018 Nurul Lubis 16

Experiment 13 July 2018 Nurul Lubis 17

Experimental set-up Pre-train inin ing • SubTle corpus [Ameixa et al., 2014] ~5.5M dialogue pairs from movie subtitle • HRED to retain information across Pre-training SubTle dialogue turns Fin Fine-tunin ing Counselor Unsupervised Counseling dialogue Fine-tuning • Counseling corpus action label data clustering • Baseline: Emo-HRED • emotion context • Proposed: MC-HRED Testing • emotion and action contexts • Clust-HRED • Action context 13 July 2018 Nurul Lubis 18

Pre-training and fine-tuning Pre-training initializes the weights of HRED components Selective fine-tuning: only optimize parameters affected by new contexts MC-HRED is jointly trained on combined losses • NLL of target response • Emotion prediction error • Action prediction error MC-HRED architecture. 13 July 2018 Nurul Lubis 19

Objective evaluation: perplexity Mo Mode del Emo Emo Act Actio ion Perp rplexit xity Emo-HRED yes no 42.60 • Combining cluster label and emotion K-means 39.57 Clust-HRED no contexts DPGMM 30.57 • K-means cluster label shows improvements K-means 29. 29.57 57 MC-HRED yes • DPGMM cluster label slightly worsen DPGMM 32.04 13 July 2018 Nurul Lubis 20

Objective evaluation: perplexity • Perplexity on short and long queries Perp rplexit xity Model Mode Emo Emo Act Actio ion • Performance on short queries are all ll shor short long long consistently better than long ones Emo- yes no 42.60 35.74 61.17 HRED • MC-HRED with K-means obtains K-means 39.57 32.30 57.37 Clust- substantial improvement on long no HRED DPGMM 30.57 24.79 42.25 triples K-means 29.57 29 57 23.23 23 23 38 38.73 73 • The multiple contexts help, especially MC-HRED yes DPGMM 32.04 25.00 42.43 for long inputs 13 July 2018 Nurul Lubis 21

Subjective evaluation 4.2 • 100 queries, each judged by 20 crowd Emo-HRED MC-HRED 4.1 workers 4 • Naturalness • Emotional impact 3.9 • Engagement 3.8 • Improved engagement while 3.7 maintaining the emotional impact and 3.6 naturalness 3.5 • MC-HRED produce responses with 4.17 4.17 3.79 3.8 3.92 3.99 3.4 2.53 more words on average naturalness emo_impact engagement 13 July 2018 Nurul Lubis 22

Conclusion We presented • A corpus showing expert strategy in positive Future Work emotion elicitation • Multimodal information: • Unsupervised clustering of expert dialogue speech, visual • A multi-context neural response generation • Evaluation through user • Improves performance on longer queries interaction • Improves dialogue engagement • Produces longer responses 13 July 2018 Nurul Lubis 23

Thank you! 13 July 2018 Nurul Lubis 24

Examples U 1 are you a student here? U 2 uh yes, actually I just got, er that's my lab over there in U 1 oh how do you feel about that one. social computing yes (laughter). U 2 yes i heard the story. U 3 (Target) oh really. so you've been watching us going by. U 3 (Target) you heard it before. Emo-HRED Oh okay. Emo-HRED right. MC-HRED (laughter) it’s nice to meet you. MC-HRED it’s a big thing. 13 July 2018 Nurul Lubis 25

Traditional Works on Emotion Expression and Recognition Expression Recognition Em Emoti tion express ssion or or si simula lation Emoti Em tion rec ecognition • Conveying emotion to user • Recognizing user’s emotional state • Increasing closeness and satisfaction • Increasing task success [Forbes-Riley and Litman, 2012] [Higashinaka et al., 2008] 13 July 2018 Nurul Lubis 26