Emotion-Transformer

Emotion Detection with Transformer models 😃😡😱😊

To test our model capacity to predict emotions we use the GoEmotions Corpus. This corpus consists of 58k reddit comments annotated with 28 different emotions.

Model Architecture

Our model is built on top of a pretrained Transformer model such as RoBERTa. To get a sentence representation we apply a pooling technique (average, max or CLS) and pass that representation to a classification head that produces an independent score for each label.

Install

virtualenv -p python3.6 emot-env
source emot-env/bin/activate

https://github.com/HLT-MAIA/Emotion-Transformer
cd Emotion-Transformer
pip install -r requirements.txt

Command Line Interface:

Train:

To set up your training you have to define your model configs. Take a look at the example.yaml in the configs folder, where all hyperparameters are briefly described.

After defining your hyperparameter run the following command:

python cli.py train -f configs/example.yaml

Monitor training with Tensorboard:

Launch tensorboard with:

tensorboard --logdir="experiments/"

Interact:

Fun command where we can interact with with a trained model.

python cli.py interact --experiment experiments/{experiment_id}/

Testing:

After training we can test the model against the testset by running.

python cli.py test --experiment experiments/{experiment_id}/

This will compute the precision, recall and F1 for each label and the macro-average results.

Results:

Ekman Emotion

Model	Macro-Precision	Macro-Recall	Macro-F1
biLSTM Reported	-	-	0.53
BERT-base Reported	0.59	0.69	0.64
Mini-BERT	0.43	0.69	0.51
RoBERTa-base	0.58	0.69	0.62

Note: The results reported were achieved with default parameters. With some search over hyper-parameters better results can be achieved.