brenomatos / portuguese-pos-tagging

POS-Tagger for portuguese using a LSTM

Part-Of-Speech Tagging (Portuguese)

Code regarding the second practical assignment from Natural Language Processing class @ UFMG. The assignment was to perform a POS-Tag task on a portuguese corpus.

Files Needed

In this assignment, I used the Mac-Morpho dataset. Mac-Morpho is a corpus of Brazilian Portuguese texts annotated with part-of-speech tags. It's files and manual (in portuguese) can be found here.

Running the code

First, clone this repository.

Then, execute the command:

python3 main.py

Note that you can change the # of epochs, batch size and the sliding window size in the main method:

def main(window_size,epochs,batch_size)

The results will be stored in the 'results' folder.

Results

For each model created, there will be 2 results files: one is named like 'x-y.csv', being 'x' the sliding window size, and 'y' being the # of epochs. The other file is named 'total_accuracy.csv' and is structured like

window_size,epochs,accuracy
3,1,0.9615463528245525
4,1,0.9615463534395741
5,1,0.9615463534328539

The first file contains the accuracy of the model regarding every grammatical class considered. The second file contains the accuracy of the model as a whole, considering total hits/misses, for all models.

Graphs

This code will also create graphs for each model created, stored in the 'results/graphs/' folder. The first kind of graph displays each class's accuracy, for each model created, like the example below:

The second kind of graph created represents the data stored in the "total_accuracy.csv" file, like the example below:

Built With

• Keras
• Scikit-learn

Contributing

Pull requests are welcome!

Acknowledgments

Some methods from 'graphs.py' were inspired by my friend and brazillian developer Rafael. Check his implementation for a portuguese POS-Tagger here

License

MIT