README.md

metadata

dataset_info:
  features:
    - name: id
      dtype: string
    - name: annotation
      struct:
        - name: emotion
          dtype: string
    - name: meta
      struct:
        - name: situation
          struct:
            - name: eng
              dtype: string
            - name: rus
              dtype: string
    - name: utterances
      list:
        - name: annotation
          struct:
            - name: emotion
              dtype: string
        - name: role
          dtype: string
        - name: text
          struct:
            - name: eng
              dtype: string
            - name: rus
              dtype: string
  splits:
    - name: train
      num_bytes: 31583670
      num_examples: 24856
  download_size: 15553828
  dataset_size: 31583670
task_categories:
  - text-classification
language:
  - ru
  - en
tags:
  - empathy
  - psychology
  - psytechlab
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

EmpatheticIntents-ru

This dataset was originally published here. We translated its final part that placed in ./datasets/empatheticdialogues_annotated to Russian by our pipeline that relies on a power of modern LLMs. You can find translated texts by attributes that contain a suffix "_ru". Next we provide original description.

For the translation the Qwen-2.5-72b (primary) и GPT-4o (secondary for the hard cases) was used.

Introduction

In empathetic human social conversations, the speaker often carries a certain emotion, however, the listener being empathetic does not necessarily carry a specific emotion. Instead, by means of a question or an expression of acknowledgement or agreement, a listener can show his empathy towards the other person. By manually analyzing a subset of listener utterances in the EmpatheicDialogues datasets (Rashkin et al., 2019) containing 25K empathetic human coversations, we discovered specific means or intents that a listener uses to express his empathy towards the speaker. The following are the most frequent intents that were discovered:

1. Questioning (to know further details orclarify) e.g. What are you looking forward to?

2. Acknowledging (Admitting as beingfact) e.g. That sounds like double good news. It was probably fun having your hard work rewarded

3. Consoling e.g. I hope he gets the help he needs.

4. Agreeing (Thinking/Saying the same) e.g. That’s a great feeling, I agree!

5. Encouraging e.g. Hopefully you will catch those great deals!

6. Sympathizing (Express feeling pity orsorrow for the person in trouble) e.g. So sorry to hear that.

7. Suggesting e.g. Maybe you two should go to the pet store to try and find a new dog for him!

8. Wishing e.g. Hey... congratulations to you on both fronts!

We have extended the number of examples per each intent by searching through the rest of the dataset using words and phrases that are most indicative of the intent. For example, words and phrases such as 100%, exactly, absolutely, definitely, i agree, me neither, me too and i completely understand are indicative of the category Agreeing.

In total dataset has 33 categories: 24 emotional classes of the original EmpatheticDialogues that are related to the speaker utterences, 8 discovered intentens that are related to the listener utterences and neutral category.

Bibliography

Welivita, A. and Pearl Pu. “A Taxonomy of Empathetic Response Intents in Human Social Conversations.” International Conference on Computational Linguistics (2020).

Hannah Rashkin, Eric Michael Smith, Margaret Li and Y-Lan Boureau. 2019. Towards Empathetic Open-domain Conversation Models: A New Benchmark and Dataset. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5370–5381, Florence, Italy.