pttagger 0.1.1

Simple PyTorch-based tagger

pttagger is a simple PyTorch-based tagger which has the following features:

• stacked bi-directional RNN (GRU or LSTM)

• variable-sized mini-batches

• multiple inputs

Quick Start

Installation

Run this command in your terminal:

$ pip install pttagger

Pre-processing

Suppose that you have the following examples of named entity recognition:

• Joe/B-PER Smith/I-PER goes/O to/O Japan/B-LOC ./O

• Jane/B-PER Smith/I-PER belongs/O to/O Kyoto/B-ORG University/I-ORG ./O

• …

First, give the examples to construct a Dataset object like this:

from pttagger.dataset import Dataset

examples = [
    {'Xs': [['Joe', 'Doe', 'goes', 'to', 'Japan', '.']],
     'Y': ['B-PER', 'I-PER', 'O', 'O', 'B-LOC', 'O']},
    {'Xs': [['Jane', 'Doe', 'belongs', 'to', 'Kyoto', 'University', '.']],
     'Y': ['B-PER', 'I-PER', 'O', 'O', 'B-ORG', 'I-ORG', 'O']},
    ...
]
dataset = Dataset(examples)

You can also use multiple inputs as a value of Xs. In the following case, Xs has not only word information but also POS information:

examples = [
    {'Xs': [['Joe', 'Doe', 'goes', 'to', 'Japan', '.'], ['NNP', 'NNP', 'VBZ', 'TO', 'NNP', '.']],
     'Y': ['B-PER', 'I-PER', 'O', 'O', 'B-LOC', 'O']},
    {'Xs': [['Jane', 'Doe', 'belongs', 'to', 'Kyoto', 'University', '.'], ['NNP', 'NNP', 'VBZ', 'TO', 'NNP', 'NNP', '.']],
     'Y': ['B-PER', 'I-PER', 'O', 'O', 'B-ORG', 'I-ORG', 'O']},
    ...
]
dataset = Dataset(examples)

Now, dataset has the following two indices:

• x_to_index: e.g., [{'<PAD>': 0, '<UNK>': 1, 'Joe': 2, 'Doe': 3, ...}]

• y_to_index: e.g., {'<PAD>': 0, '<UNK>': 1, 'B-PER': 2, 'I-PER': 3, ...}

If you use multiple inputs, x_to_index has indices for each input.

Training

Construct a Model object and train it as follows:

from pttagger.model import Model

EMBEDDING_DIMS = [100]  # if you use multiple inputs, set a dimension for each input
HIDDEN_DIMS = [10]  # the same as above
x_set_sizes = [len(x_to_index) for x_to_index in dataset.x_to_index]
y_set_size = len(dataset.y_to_index)
model = Model(EMBEDDING_DIMS, HIDDEN_DIMS, x_set_sizes, y_set_size)
model.train(dataset)

You can also use the following parameters:

• use_lstm - If True, uses LSTM. Default: False (uses GRU)

• num_layers - Number of recurrent layers. Default: 1

Test

Predict tags for test examples like this:

test_examples = [
    {'Xs': [['Richard', 'Roe', 'comes', 'to', 'America', '.']],
     'Y': ['B-PER', 'I-PER', 'O', 'O', 'B-LOC', 'O']}
]
test_dataset = Dataset(test_examples)
results = model.test(dataset)