Towards Empathetic Open-domain Conversation Models: A New Benchmark and Dataset

Paper · Source
Psychology EmpathyEmotions

This work proposes a new benchmark for empathetic dialogue generation and EMPATHETICDIALOGUES, a novel dataset of 25k conversations grounded in emotional situations. Our experiments indicate that dialogue models that use our dataset are perceived to be more empathetic by human evaluators, compared to models merely trained on large-scale Internet conversation data.

Each dialogue is grounded in a specific situation where a speaker was feeling a given emotion, with a listener responding (Figure 2). The new resource consists of crowdsourced one-on-one conversations, and covers a large set of emotions in a balanced way. This dataset is larger and contains a more extensive set of emotions than many similar emotion prediction datasets from other text domains

Surprised got,shocked,really that's,good,nice
Excited going,wait,i'm that's,fun,like
Angry mad,someone,got oh,would,that's
Proud got,happy,really that's,great,good
Sad really,away, get sorry,oh,hear
Annoyed get,work,really that's,oh,get
Grateful really,thankful,i'm that's,good,nice
Lonely alone,friends,i'm i' m,sorry,that's
Afraid scared,i'm,night oh,scary,that's
Terrified scared,night,i'm oh,that's,would
Guilty bad,feel,felt oh,that's,feel
Impressed really,good,got that's,good,like
Disgusted gross,really,saw oh,that's,would
Hopeful i'm,get,really hope,good,that's
Confident going,i'm,really good,that's,great
Furious mad,car,someone oh,that's,get
Anxious i'm,nervous, going oh,good,hope
Anticipating wait,i'm, going sounds,good,hope
Joyful happy,got,i'm that's,good,great
Nostalgic old,back, really good,like,time
Disappointed get,really,work oh,that's,sorry
Prepared ready,i'm,going good,that's,like
Jealous friend,got,get get,that's,oh
Content i'm,life,happy good,that's,great
Devastated got,really, sad sorry,oh,hear
Embarrassed day,work,got oh, that's,i'm
Caring care,really,taking that's,good,nice
Sentimental old,really,time that's,oh,like
Trusting friend,trust,know good,that's,like
Ashamed feel,bad,felt oh, that's,i'm
Apprehensive i'm,nervous, really oh,good,well
Faithful i'm,would,years good,that's,like

As a result, we developed the Empathetic Question Taxonomy (EQT) with two distinguished branches: question acts describe semantic-driven features of questions (e.g., ask for confirmation, positive rhetoric), whereas question intents characterize their emotion-regulation functions targeted at the interlocutor’s emotional state (e.g., sympathize, amplify excitement). As it will be revealed further (§7), an empathetic listener can use different question acts to deliver the same intent, justifying the proposed branching.

Question acts Request information (38.7%): Ask for new factual information

Ask about consequence (21.0%): Ask about the result of the described action or situation

Ask about antecedent (17.1%): Ask about the reason or cause of the described state or event

Suggest a solution (8.7%): Provide a specific solution to a problem in a form of a question

Ask for confirmation (5.8%): Ask a question to confirm or verify the listener’s understanding of something that has been described by the speaker

Suggest a reason (5.2%): Suggest a specific reason or cause of the event or state described by the speaker in a form of a question

Irony (1.3%): Ask a question that suggests the opposite of what the speaker may expect, usually to be humorous or pass judgement