Generation for Positive Emotion Elicitation Nuru Nurul Fithri ria - - PowerPoint PPT Presentation

Affect-sensitive Dialogue Response Generation for Positive Emotion Elicitation

Nuru Nurul Fithri ria Lubi ubis Augm ugmen ented ed Human Com

• mmunication
• n (AHC) Lab

Na Nara Inst nstitute e of Sci cien ence e and nd Technol

• log
• gy (NAI

NAIST ST)

Affective dialogue systems

High potential of dialogue systems to address the emotional needs of users

15 March 2019 Nurul Lubis 2

• Increase of dialogue system works

and applications in various tasks involving affect

• Companion for the elderly

[Miehle et al., 2017]

• Distress clues assessment

[DeVault et al., 2014]

• Affect-sensitive tutoring

[Forbes-Riley and Litman, 2012]

Emotion elicitation

Emo motio ion elic licit itatio ion: elic licit itin ing change e of

• f emotio

ion in dialo 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 emotio ional l benefit it for the user

15 March 2019 Nurul Lubis 3

Expression

Recognition

Traditional emotion works

Expression Intent for emotion elicitation

Research goal: Positive emotion elicitation

We aim to draw on an overlooked potential of emotion elicitation to improve us user emotional st states

• A chat-based dialogue system

with an implicit goal of po posi siti tive emotion elicitation

15 March 2019 Nurul Lubis 4

Circumplex model of affect [Russell, 1980]

Different responses elicit different emotions

15 March 2019 Nurul Lubis 5



I failed the test.

I failed the test. Oh, Oh, aga gain? Yeah… I failed the test. You will do do be better er nex next time! e! Thank you.

arousal valence

Emotiona nal impa pact

Negative

arousal valence

Positive

User Sys System User

Positive emotion elicitation does NOT mean always responding with positive emotion

13 December 2018 Nurul Lubis 6

How was your day?

• Terrible. Work did not go well.

That’s too bad. How was your day?

• Terrible. Work did not go well.

I’m glad to hear that!

There are situations where “happy responses” can lead to negative impact Expressing negative emotion can lead to positive impact

• System should learn the

proper strategy

Neural chat-based dialogue system

• End-to-end modeling of chat dialogue
• 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]

15 March 2019 Nurul Lubis 7

Not yet an application towards emotion elicitation

[Serban et al., 2016]

Proposed approach

15 March 2019 Nurul Lubis 8

Emotion-sensitive response generation: Emo-HRED

Encodes emotional context and considers it in generating a response Train on responses that elicit positive emotion

15 March 2019 Nurul Lubis 9

Training Emo-HRED

Opti Optimization

• n
• Train on combined losses
• Negative Log Likelihood (NLL) of target

response

• Emotion prediction error
• The emotion encoder targets the

emotion label of the dialogue turn

• The final cost is used to optimize the

entire network

• Adam optimizer

Pre Pre-training g and nd selec ective fine-tuning

• Emotion-annotated data is limited
• Start by pr

pre-training HRED with large- scale conversational data

• Learning semantic and syntactic

knowledge

• Select

ctivel ely fine-tune Emo-HRED with the emotion-annotated data

• Only train parts that are affected by

emotion context

• Avoid over-fitting or destabilizing

15 March 2019 Nurul Lubis 10

Datasets

Existing da data SubTle Database

• For pre-training
• Large-scale conversational corpus

from movie subtitles (5.5M triples) SEMAINE Database

• Small amount of conversation

between user and listening agent in WoZ fashion (2K triples) Posi sitive-emotion

• n eliciting da

data SEMAINE-positive

• For fine-tuning
• Augmenting an existing corpus
• Contains positive-emotion eliciting

responses

15 March 2019 Nurul Lubis 11

SEMAINE Database default dataset system

response

human judgement positive- emotion eliciting dataset dialogue system

Evaluation

15 March 2019 Nurul Lubis 12

Objective evaluation: Perplexity

Model Parameter er upda update Perp rplexity

• n SEMAINE-

posit itiv ive test set

Baseline HRED standard 121.44 selective 100.94 Proposed Emo-HRED selective 42.26 26

15 March 2019 Nurul Lubis 13

Emotion information can be leveraged in response generation to reduce perplexity Pre Pre-training: g: SubTle Fine-tuning: SEMAINE-positive Testing: SEMAINE-positive

Subjective evaluation

• Evaluation via crowdsourcing
• 100 test queries, 20 judgments each
• Likert scale 1 to 5 (higher is better)
• Naturalness
• Positive emotional impact

15 March 2019 Nurul Lubis 14

3.26 3.22 3.27 3.39 3 3.1 3.2 3.3 3.4 3.5 naturalness positive emotional impact Likert score HRED positive Emo-HRED positive

The proposed model is perceived as more natural and significantly elicits a more positive emotion (𝑞 < 0.05)

Conclusion

15 March 2019 Nurul Lubis 15

Conclusion

• We proposed: a dialogue response generator to elicit positive emotion
• Considers emotional context of dialogue
• Trained on constructed corpus that contains responses with positive emotional impact
• Subjective and objective evaluation show improvement over system that ignores

emotion information

• More natural
• Elicit a more positive emotion impact
• Future work
• Collect and utilize more emotion rich dialogue data
• Richer dialogue context
• Other modalities
• Longer context

15 March 2019 Nurul Lubis 16

Thank you

15 March 2019 Nurul Lubis 17

Automatically retrieve responses with positive impact by utilizing example-based dialogue system approach

• Semantic similarity:

text cosine similarity between query and example query

• Emotion correlation: valence

& arousal Pearson’s score between query and example query

• Emotional impact: valence

change in the example triple

25-Mar-19 18

Example Database Semantic similarity scoring Emotion correlatio n scoring Emotional impact scoring Query Response Response best best emotion text emotional change

10-best 3-best

Nurul Lubis

Traditional EBDM Proposed EBDM Evaluation shows that the proposed EBDM is perceived as more natur ural and elicit a more positive impact

(Lubis et al., 2017) in Proc. IWSDS 2017