In the first part, the dataset must be pre-processed in order to give it to the model. This is done using the 🤗 Transformers and BERT tokenizers. Once this is done, finetuning is applied from bert-base-cased and using the 🤗 AutoModelForTokenClassification.

Finally, the model is trained obtaining the neccesary metrics for evaluating its performance (Precision, Recall, F1 and Accuracy)

Summary of executed tests can be found in: https://docs.google.com/spreadsheets/d/1lI7skNIvRurwq3LA5ps7JFK5TxToEx4s7Kaah3ezyQc/edit?usp=sharing

Model can be found in: https://huggingface.co/paulrojasg/NER-finetuned-BETO

Github repository: https://github.com/paulrojasg/nlp_4th_workshop

Training

Training Details

Epochs: 10
Learning Rate: 2e-05
Weight Decay: 0.01
Batch Size (Train): 16
Batch Size (Eval): 8

Training Metrics

Epoch	Training Loss	Validation Loss	Precision	Recall	F1 Score	Accuracy
1	0.0065	0.2077	0.8436	0.8564	0.8499	0.9712
2	0.0062	0.2345	0.8318	0.8513	0.8415	0.9683
3	0.0069	0.2156	0.8464	0.8470	0.8467	0.9674
4	0.0064	0.2189	0.8356	0.8490	0.8423	0.9686
5	0.0055	0.2383	0.8373	0.8488	0.8430	0.9687
6	0.0050	0.2315	0.8334	0.8543	0.8438	0.9694
7	0.0037	0.2343	0.8428	0.8573	0.8500	0.9703
8	0.0031	0.2493	0.8400	0.8555	0.8477	0.9694
9	0.0024	0.2421	0.8478	0.8617	0.8547	0.9704
10	0.0023	0.2497	0.8432	0.8598	0.8514	0.9703

Authors

Made by:

Paul Rodrigo Rojas Guerrero
Jose Luis Hincapie Bucheli
Sebastián Idrobo Avirama

With help from:

Raúl Ernesto Gutiérrez