Motivation

Motivation

• No Child is safe in the Digital World, 90% of children have access to the internet

Motivation

• Social media has been connected to various detrimental consequences on people's psychology and mental health. (Even our close friends do not know it is cyberbullying.)

"Towards Facilitating Empathic Conversations in Online Mental Health Support: A Reinforcement Learning Approach" - Sharma et al., 2021

"A Computational Approach to Understanding Empathy Expressed in Text-Based Mental Health Support" - Sharma et al., 2021

"DYME: A Dynamic Metric for Dialog Modeling Learned from Human Conversations" - von Unold et al., 2021

"Text Coherence Analysis Based on Deep Neural Network" - Cui et al. - 2017

Literature Review

Key FINDINGS

• GPT-2 used for empathetic rewriting task
• Sentence Coherence is lower
• The Edit rate is higher compared to other Transformers
• Uses EPIOME as Rewards for Training

"Towards Facilitating Empathic Conversations in Online Mental Health Support: A Reinforcement Learning Approach" - Sharma et al., 2021

Key FINDINGS

• Aimed to develop a theoretically-grounded framework for finding empathy in text-based conversations
• Bi-encoder Model with Attention (RoBERTA)
• Emotional Reactions
• Interpretations
  • Understand in-depth feelings
• Explorations
  • Understand exploring feelings and experiences not given in the context

"A Computational Approach to Understanding Empathy Expressed in Text-Based Mental Health Support" - Sharma et al., 2021

〞

I feel really sad for you

Everything will be fine

vs

Emotional Reactions​​

• ​Warmth, Compassion & Concern

〞

This must be terrifying

Interpretations

• understanding in-depth the feelings
• cognitive understanding increases awareness of hidden feelings and experiences

I understand how you feel

vs

〞

What happened?

Explorations

• improve understanding of the seeker by exploring feelings and experiences not put in the post
• understand the outside context of the post

Are you alone right now?

vs

Key FINDINGS

• Uses EPIOME in metrics

• Utterance length

• Diversity

• Utterance Repetition

"DYME: A Dynamic Metric for Dialog Modeling Learned from Human Conversations" - von Unold et al., 2021

Findings

so far

User: "My diet becomes f**ked when I get depressed... I can’t control myself I eat so much when I’m de- pressed."

BayMax: "The only way to stop is a hard and fast recovery, but you gotta keep pushing through it."

GPT-2 with Fine Tuning

Responses from BayMax (GPT-2) with fine tuning

Part - A (Recap)

BayMAX Core

• GPT-2 Model structure
• A1: Selecting a position in the response post
• A2: Generating an Empathetic response
  • Rewritten the sentence with the help of Position from A1 and A2 (candidate sentence)

(Better understanding of architecture on next page)

Baymax Core

Friend & PRotector

• BayMax as Friend
  • Enhancing positive emotion in chatbot-based conversations.
• BayMax as Protector
  • converting negative into positive emotions in online interactions.

Demo with Baymax

Evaluation with other LMs

Seeker: When my current girlfriend said yes to dating me.

Baymax: It sounds like you're feeling really happy and excited. It's great to hear that your girlfriend said yes and that you're in a good place in your relationship.

PARTNER: That is amazing! Are you happy and excited?

BART: oh nice! I hope you are happy.

T5: Oh wow, that's really nice of her.

Performance of Baymax With other Lms

I don't feel so great today, I need a friend to talk to.

Seeker:

I am here for you. Let's talk about what's been on your mind and see how we can make you feel better.

Baymax:

Why don't you talk to them?

T5:

What's wrong with you?

BART:

I'm here for you. What do you want to talk about? Like School or Work?

PARTNER:

Evaluation with Epiome

• T5, DialoGPT, and BART Evaluation with EPIOME are in-progress
• Since it requires manual intervention to find the rationales.
• Example in next page.

EPIOME Rationales of EX, IP of PARTNER

EX, IP Rationales