|
--- |
|
license: cc-by-sa-3.0 |
|
language: |
|
- ja |
|
tag: |
|
- emotion-analysis |
|
datasets: |
|
- wrime |
|
widget: |
|
- text: "車のタイヤがパンクしてた。。いたずらの可能性が高いんだって。。" |
|
--- |
|
|
|
# WRIME-fine-tuned BERT base Japanese |
|
|
|
This model is a [Japanese BERT<sub>BASE</sub>](https://huggingface.co/cl-tohoku/bert-base-japanese-v2) fine-tuned on the [WRIME](https://github.com/ids-cv/wrime) dataset. It was trained as part of the paper ["Emotion Analysis of Writers and Readers of Japanese Tweets on Vaccinations"](https://aclanthology.org/2022.wassa-1.10/). Fine-tuning code is available at this [repo](https://github.com/PatrickJohnRamos/BERT-Japan-vaccination). |
|
|
|
# Intended uses and limitations |
|
|
|
This model can be used to predict intensities scores for eight emotions for writers and readers. Please refer to the `Fine-tuning data` section for the list of emotions. |
|
|
|
Because of the regression fine-tuning task, it is possible for the model to infer scores outside of the range of the scores of the fine-tuning data (`score < 0` or `score > 4`). |
|
|
|
# Model Architecture, Tokenization, and Pretraining |
|
|
|
The Japanese BERT<sub>BASE</sub> fine-tuned was `cl-tohoku/bert-base-japanese-v2`. Please refer to their [model card](https://huggingface.co/cl-tohoku/bert-base-japanese-v2) for details regarding the model architecture, tokenization, pretraining data, and pretraining procedure. |
|
|
|
# Fine-tuning data |
|
|
|
The model is fine-tuned on [WRIME](https://github.com/ids-cv/wrime), a dataset of Japanese Tweets annotated with writer and reader emotion intensities. We use version 1 of the dataset. Each Tweet is accompanied by a set of writer emotion intensities (from the author of the Tweet) and three sets of reader emotions (from three annotators). The emotions follow Plutchhik's emotions, namely: |
|
|
|
* joy |
|
* sadness |
|
* anticipation |
|
* surprise |
|
* anger |
|
* fear |
|
* disgust |
|
* trust |
|
|
|
These emotion intensities follow a four-point scale: |
|
|
|
| emotion intensity | emotion presence| |
|
|---|---| |
|
| 0 | no | |
|
| 1 | weak | |
|
| 2 | medium | |
|
| 3 | strong | |
|
|
|
# Fine-tuning |
|
|
|
The BERT is fine-tuned to directly regress the emotion intensities of the writer and the averaged emotions of the readers from each Tweet, meaning there are 16 outputs (8 emotions per writer/reader). |
|
|
|
The fine-tuning was inspired by common BERT fine-tuning procedures. The BERT was fine-tuned on WRIME for 3 epochs using the AdamW optimizer with a learning rate of 2e-5, β<sub>1</sub>=0.9, β<sub>2</sub>=0.999, weight decay of 0.01, linear decay, a warmup ratio of 0.01, and a batch size of 32. Training was conducted with an NVIDIA Tesla K80 and finished in 3 hours. |
|
|
|
# Evaluation results |
|
|
|
Below are the MSEs of the BERT on the test split of WRIME. |
|
|
|
| Annotator | Joy | Sadness | Anticipation | Surprise | Anger | Fear | Disgust | Trust | Overall | |
|
|---|---|---|---|---|---|---|---|---|---| |
|
| Writer | 0.658 | 0.688 | 0.746 | 0.542 | 0.486 | 0.462 | 0.664 | 0.400 | 0.581 | |
|
| Reader | 0.192 | 0.178 | 0.211 | 0.139 | 0.032 | 0.147 | 0.123 | 0.029 | 0.131 | |
|
| Both | 0.425 | 0.433 | 0.479 | 0.341 | 0.259 | 0.304 | 0.394 | 0.214 | 0.356 | |
