|
--- |
|
library_name: transformers |
|
language: |
|
- en |
|
--- |
|
### 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 ``<s>`` token, where the valence/arousal predictions for an ``<s>`` 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: |
|
|
|
 |
|
|
|
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. |
|
|
|
[More Information Needed] |
|
|
|
|
|
## 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:** |
|
|
|
[More Information Needed] |
|
|
|
## Model Card Contact |
|
|
|
For further inquiries, please contact lukas1[dot]christ[at]uni-a[dot].de |
