|
|
--- |
|
|
license: mit |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- uncertainty-detection |
|
|
- social-media |
|
|
- text-classification |
|
|
widget: |
|
|
- text: "It seems like Bitcoin prices are heading into bearish territory." |
|
|
example_title: "Hedge Detection (Positive - Label 1)" |
|
|
- text: "Bitcoin prices have fallen by 42% in the last 30 days." |
|
|
example_title: "Hedge Detection (Negative - Label 0)" |
|
|
--- |
|
|
|
|
|
### Overview |
|
|
Fine tuned VinAI's BERTweet base model on the Wiki Weasel 2.0 Corpus from the [Szeged Uncertainty Corpus](https://rgai.inf.u-szeged.hu/node/160) for hedge (linguistic uncertainty) detection in social media texts. Model was trained and optimised using Ray Tune's implementation of Deep Mind's Population Based Training with the arithmetic mean of Accuracy & F1 as its evaluation metric. |
|
|
|
|
|
### Labels |
|
|
* LABEL_1 = Positive (Hedge is detected within text) |
|
|
* LABEL_0 = Negative (No Hedges detected within text) |
|
|
|
|
|
### <a name="models2"></a> Model Performance |
|
|
Model | Accuracy | F1-Score | Accuracy & F1-Score |
|
|
---|---|---|--- |
|
|
`BERTweet-Hedge` | 0.9680 | 0.8765 | 0.9222 |
|
|
|
