README.md

---
library_name: transformers
language:
- en
inference: false
---
## Emotion Prediction for Written Stories

<!-- Provide a quick summary of what the model is/does. -->

This model is intended to predict emotions (valence, arousal) in written stories. For all details see [the paper (TODO)](#) and [the accompanying github repo (TODO)](#).



### Model Description

<!-- Provide a longer summary of what this model is. -->

As described in [the paper (TODO)](#), this model is finetuned from [DeBERTaV3-large](https://huggingface.co/microsoft/deberta-v3-large) and predicts sentence-wise valence/arousal values between 0 and 1. 

This particular checkpoint was trained with a window size of 4. 

All available checkpoints and their performance measured by Concordance Correlation Coefficient (CCC):

| Model                                                                  | Valence dev/test   | Arousal dev/test   |
|------------------------------------------------------------------------|--------------------|--------------------|
|[stories-emotion-c0](https://huggingface.co/chrlukas/stories-emotion-c0)| .7091/.7187        | .5815/.6189        |
|[stories-emotion-c1](https://huggingface.co/chrlukas/stories-emotion-c1)| .7715/.7875        | .6458/.6935        |
|[stories-emotion-c2](https://huggingface.co/chrlukas/stories-emotion-c2)| .7922/.8074        | .6667/.6954        |
|[stories-emotion-c4](https://huggingface.co/chrlukas/stories-emotion-c4)| .8078/.8146        | .6763/.7115        |
|[stories-emotion-c8](https://huggingface.co/chrlukas/stories-emotion-c8)| **.8223**/**.8237**| **.6829**/**.7120**|

We provide the best out of 5 seeds for each context size. Hence, the numbers in this table differ from the result table in the paper, where the mean performance across 5 seeds is reported.

Technically, this model predicts token-wise valence/arousal values. Sentences are concatenated via the ``[SEP]`` token, where the valence/arousal predictions for an ``[SEP]`` token
are meant to be the predictions for the sentence preceding it. All other tokens' predictions should be ignored. For reference, see the figure in the paper:

![image](tales_vertical.png)

The [accompanying repo](TODO) provides a convenient script to use the model for prediction.

### Model Sources

<!-- Provide the basic links for the model. -->

- **Repository:** [Github](TODO)
- **Paper:** [ArXiv](TODO)

## Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
This model is intended to predict emotions (valence, arousal) in written stories. It was mainly trained on stories for children. 
Please note that the model is not production-ready and provided here for demonstration purposes only. 
For details on the datasets used, please refer to the [paper (TODO)](#).

In the [github repository](TODO), a convenient script to predict V/A in existing texts is provided. Example call: 

``
python3 predict.py --input_csv input_file.csv --output_csv output_file.csv --checkpoint_dir chrlukas/stories-emotion-c4 --window_size 4 --batch_size 4
``

## Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->
Please see the *Limitations* section in [the paper](#). Please note that the model is not production-ready and provided here for demonstration purposes only. 


## Citation [optional]

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**


## Model Card Contact

For further inquiries, please contact lukas1[dot]christ[at]uni-a[dot].de