Parts-of-Speech Tagger

The purpose of this project was to learn how to implement RNNs and compare different types of RNNs on the task of Parts-of-Speech tagging using a part of the CoNLL-2012 dataset with 42 possible tags. This repository contains:

1. a custom implementation of the GRU cell.
2. a custom implementation of the RNN architecture that may be configured to be used as an LSTM, GRU or Vanilla RNN.
3. a Parts-of-Speech tagger that can be configured to use any of the above custom RNN implementations.

Requirements

• python 3.5
• pytorch
• torchtext

Organisation

The code in the repository are organised as follows:

• gru.py: custom GRU
• rnn.py: custom RNN
• model.py: POS Tagger Model
• train.py: training/validation/testing code
• main.py: driver code

The raw dataset is in RNN_Data_files/.

Usage

Preprocessing datasets

Use preprocess.sh to generate tsv datasets containing sentences and POS tags in the intended data_dir (RNN_Data_files/ here).

$ ./preprocess.sh RNN_Data_files/train/sentences.tsv RNN_Data_files/train/tags.tsv RNN_Data_files/train_data.tsv
$ ./preprocess.sh RNN_Data_files/val/sentences.tsv RNN_Data_files/val/tags.tsv RNN_Data_files/val_data.tsv

Training/Testing

usage: main.py [-h] [--use_gpu] [--data_dir PATH] [--save_dir PATH]
                    [--rnn_class RNN_CLASS] [--reload PATH] [--test]
                    [--batch_size BATCH_SIZE] [--epochs EPOCHS] [--lr LR]
                    [--step_size N] [--gamma GAMMA] [--seed SEED]

PyTorch Parts-of-Speech Tagger

optional arguments:
  -h, --help            show this help message and exit
  --use_gpu
  --data_dir PATH       directory containing train_data.tsv and val_data.tsv (default=RNN_Data_files/)
  --save_dir PATH
  --rnn_class RNN_CLASS
                        class of underlying RNN to use
  --reload PATH         path to checkpoint to load (default: none)
  --test                test model on test set (use with --reload)
  --batch_size BATCH_SIZE
                        batchsize for optimizer updates
  --epochs EPOCHS       number of total epochs to run
  --lr LR               initial learning rate
  --step_size N
  --gamma GAMMA
  --seed SEED           random seed (default: 123)

Results

Results.pdf compares the results for LSTM, GRU and Vanilla RNN based POS Taggers on various metrics. The best accuracy of 96.12% was obtained using LSTM-based POS Tagger. The pretrained model can be downloaded from here.