POS Tagger using Deep Learning

Running the program

Training:

python buildtagger.py <train_file_absolute_path> <model_file_absolute_path>

# Example:
python buildtagger.py sents.train model-file

Testing:

python runtagger.py <test_file_absolute_path> <model_file_absolute_path> <output_file_absolute_path>

# Example:
python runtagger.py sents.test model-file sents.out

Evaluation:

python eval.py <output_file_absolute_path> <reference_file_absolute_path>

# Example:
python eval.py sents.out sents.answer

Penn Treebank Tagset

The list of POS tags can be seen here.

Results

The following are the results when using a Bidirectional LSTM model (without character embeddings). Training time is restricted to 10 minutes. Testing time for 6.test is restricted to 2 minutes.

Parameters and Hyperparameters 1:

LSTM
- dropout = 0.5
- num_lstm_units = 50
- embedding_dim = 32
Batch size b = 32
Max words in a sentence max_length = 150
Num epochs = 7
Learning rate lr of SGD optimizer
- For epochs 0-3 = 0.1
- For epochs 4-5 = 0.05
- For epoch 6 = 0.005

Results of the POS tagger after training on sents.train.

Final loss printed = 0.1626

Test case	Accuracy
`sents.test`	0.94363
`2.test`	0.84554
`2a.test`	0.83023
`2b.test`	0.82338
`3.test`	0.80642
`4.test`	0.93334
`5.test`	0.93843
`6.test`	0.93548

Parameters and Hyperparameters 2:

LSTM
- dropout = 0.5
- num_lstm_units = 50
- embedding_dim = 32
Batch size b = 32
Max words in a sentence max_length = 150
Num epochs = 7
Learning rate lr of SGD optimizer
- For epochs 0-3 = 0.25
- For epochs 4-5 = 0.05
- For epoch 6 = 0.005