---
language: 
- en
tags:
- text-classification
- emotion
- pytorch
license: mit
datasets:
- emotion
metrics:
- accuracy
- precision
- recall
- f1
---

# DistilBERT-Base-Uncased-Emotion

## Model Description

`distilbert-base-uncased-emotion` is a specialized model finetuned on a combination of unify-emotion-datasets (https://github.com/sarnthil/unify-emotion-datasets), containing around 250K texts labeled across seven emotion categories: neutral, happy, sad, anger, disgust, surprise, and fear. This model was later adapted to a smaller set of 10K hand-tagged messages from StockTwits. The model is designed to excel at emotion detection in financial social media content such as that found on StockTwits. 

Model parameters were as follows: sequence length of 64, learning rate of 2e-5, batch size of 128, trained for 8 epochs. For steps on how to use the model for inference, please refer to the accompanying Inference.ipynb notebook.

## Training Data

The training data was obtained from the Unify Emotion Datasets available at https://github.com/sarnthil/unify-emotion-datasets.

## Evaluation Metrics

The model was evaluated using the following metrics:
- Accuracy
- Precision
- Recall
- F1-score

## Research

The underlying research for emotion extraction from social media can be found in the paper "EmTract: Extracting Emotions from Social Media". 
The paper is available at https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3975884.

### Research using EmTract
[Social Media Emotions and IPO Returns](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4384573)
[Investor Emotions and Earnings Announcements](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3626025])

## License

This project is licensed under the terms of the MIT license.