CLARIN-Emo / README.md
KoconJan's picture
Update README.md
a1bdd5f verified
|
raw
history blame
6.13 kB
metadata
dataset_info:
  features:
    - name: text
      dtype: string
    - name: Joy
      dtype: bool
    - name: Trust
      dtype: bool
    - name: Anticipation
      dtype: bool
    - name: Surprise
      dtype: bool
    - name: Fear
      dtype: bool
    - name: Sadness
      dtype: bool
    - name: Disgust
      dtype: bool
    - name: Anger
      dtype: bool
    - name: Positive
      dtype: bool
    - name: Negative
      dtype: bool
    - name: Neutral
      dtype: bool
  splits:
    - name: train
      num_bytes: 869804
      num_examples: 7169
    - name: val
      num_bytes: 160871
      num_examples: 1401
    - name: test
      num_bytes: 176693
      num_examples: 1431
  download_size: 840179
  dataset_size: 1207368
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: val
        path: data/val-*
      - split: test
        path: data/test-*
task_categories:
  - text-classification
language:
  - pl
tags:
  - emotion
  - sentence-classification
  - emotion recognition
task_ids:
  - multi-label-classification
license: cc-by-4.0
size_categories:
  - n<1K

Dataset

The dataset is made up of consumer reviews written in Polish. Those reviews belong to four domains: hotels, medicine, products, and university. This collection also contains non-opinion informative texts belonging to the same domains (meaning they are mostly neutral). Each sentence, as well as all the reviews as a whole, are annotated with emotions from the Plutchnik's wheel of emotions (joy, trust, anticipation, surprise, fear, sadness, disgust, anger), as well as the perceived sentiment (positive, negative, neutral), with ambivalent sentiment being labeled using both positive and negative labels. The dataset was annotated by six people who did not see each other's decisions. These annotations were aggregated by selecting labels annotated by at least 2 out of 6 people, meaning controversial texts and sentences can be annotated with opposing emotions. While each sentence has its own annotation, they were created in the context of the whole review.

For more information about this dataset, see references 1 and 2.

Training set

Training data consists of 776 reviews containing 6393 sentences randomly selected from the whole dataset. The split was done on the level of whole reviews, meaning no reviews are split between sets.

Test sets

Two test sets contain 167 reviews, each containing 1234 and 1264 sentence annotations.

Dataset format

The datasets are stored in three directories (training and two test sets). All datasets have the same format.

Input rows contain ordered sentences of reviews. Each review ends with a sentence made out of only the symbol #. This sentence annotation corresponds to the annotation of the whole review and is not a sentence annotation. This sentence is not a part of the original review and should not be treated as such. It only marks the end of the current review and the row that contains the corresponding review annotation. The next row after such a sentence corresponds to the first sentence of a different review.

Example:

This fragment of the training input file:

Była to pierwsza wizyta ale moze i ostatnia.
Lakarz troche apatyczny, nie wypowiadajacy sie jasno.
Mam zrobic jakies badanie ale nie dardzo wiem jakie.
Nie napisal skierowania/zalecenia, chyba mowil o gastrologii.
Powinnam byla byc bardzej wymagajaca i dopytujaca.
Nie polecam tego lekarza.
###########################

corresponds to annotations:

False	False	True	False	False	True	False	False	False	True	False
False	False	False	False	False	True	True	False	False	True	False
False	False	False	True	False	True	False	False	False	True	False
False	False	False	True	False	True	False	False	False	True	False
False	False	False	True	False	True	False	True	False	True	False
False	False	False	False	False	True	False	False	False	True	False
False	False	False	True	False	True	False	False	False	True	False

meaning sentences are labeled as:

"Była to pierwsza wizyta ale moze i ostatnia." - anticipation, sadness, negative
"Lakarz troche apatyczny, nie wypowiadajacy sie jasno." - sadness, disgust, negative
"Mam zrobic jakies badanie ale nie dardzo wiem jakie." - surprise, sadness, negative
"Nie napisal skierowania/zalecenia, chyba mowil o gastrologii." - surprise, sadness, negative
"Powinnam byla byc bardzej wymagajaca i dopytujaca." - surprise, sadness, anger, negative
"Nie polecam tego lekarza." - sadness, negative

and the review as a whole, starting from "Była to pierwsza wizyta ale moze i ostatnia." and ending at "Nie polecam tego lekarza." is labeled as: surprise, sadness, negative.

References

1. Koptyra, Bartłomiej, et al. "CLARIN-Emo: Training Emotion Recognition Models Using Human Annotation and ChatGPT." International Conference on Computational Science. Cham: Springer Nature Switzerland, 2023.

2. Kocoń, Jan, et al. "ChatGPT: Jack of all trades, master of none." Information Fusion (2023): 101861.

Licensing information

CC BY 4.0

Citation information

@article{kocon2023chatgpt,
  title={ChatGPT: Jack of all trades, master of none},
  author={Koco{'n}, Jan and Cichecki, Igor and Kaszyca, Oliwier and Kochanek, Mateusz and Szyd{\l}o, Dominika and Baran, Joanna and Bielaniewicz, Julita and Gruza, Marcin and Janz, Arkadiusz and Kanclerz, Kamil and others},
  journal={Information Fusion},
  volume={99},
  pages={101861},
  year={2023},
  publisher={Elsevier}
}

@inproceedings{koptyra2023clarin,
  title={Clarin-Emo: Training emotion recognition models using human annotation and ChatGPT},
  author={Koptyra, Bart{\l}omiej and Ngo, Anh and Radli{'n}ski, {\L}ukasz and Koco{'n}, Jan},
  booktitle={International Conference on Computational Science},
  pages={365--379},
  year={2023},
  organization={Springer}
}

@inproceedings{koptyra2024poleval,
  title={PolEval 2024 Task 2: Emotion and Sentiment Recognition},
  author={Koptyra, Bart{\l}omiej and Koco{'n}, Jan},
  booktitle={Proceedings of the PolEval 2024 Workshop},
  year={2024},
  organization={Institute of Computer Science, Polish Academy of Sciences}
}