[Dataset] Empathetic Dialogues

[anngedeus]

A collection of open-domain conversations annotated with emotion labels, designed to study empathetic response generation. The corpus contains 25k conversations (about 81k utterances) between two crowdworkers, where one speaker describes a personal situation grounded in a specific emotion and the other responds.

Towards Empathetic Open-domain Conversation Models: a New Benchmark and Dataset
Hannah Rashkin, Eric Michael Smith, Margaret Li, Y-Lan Boureau. ACL 2019.
Paper Link

Dataset details

• Number of Speakers: 796
• Number of Utterances: 88500
• Number of Conversations: 23063

Speaker-level information

Speakers in this dataset are the two dialogue participants in each EmpatheticDialogues conversation.

• In the original dataset they are indexed as "0" and "1" (per conversation).
• These indices are used as the speaker IDs in the ConvoKit corpus.

Utterance-level information

Each conversational turn is represented as an utterance. For each utterance, we provide:

• id: unique identifier for the utterance
• speaker: the speaker who produced the utterance
• conversation_id: the identifier of the conversation this utterance belongs to
• reply_to: id of the preceding utterance in the conversation (or None if it is the first utterance)
• timestamp: always None or Null (no timestamps are provided in EmpatheticDialogues)
• text: textual content of the utterance

Metadata for each utterance include:

• selfeval: self-evaluation string with numerical ratings from the dataset
• tags: extra tags if present (usually empty)
• utterance_idx: the turn index of the utterance within its conversation
• parsed: parsed version of the utterance text, represented as a SpaCy Doc
• split: dataset partition the utterance came from (train, valid, or test)

Conversation-level information

Each conversation corresponds to one dialogue in EmpatheticDialogues.

• id: identical to the original field
• meta:
  • label: the emotion label for the conversation (from context in the csv)
  • situation: the one-sentence scenario provided by the Speaker (from prompt in the csv).

Contact Information

Ann-Kareen Gedeus, ag2637@cornell.edu

convokit_data.zip