Skip to content

mayhewsw/pytorch-truecaser

Repository files navigation

Truecaser

Binder

This is a simple neural truecaser written with allennlp, and based loosely on (Susanto et al, 2016). They have an implementation here, but being written in Lua, it's a little hard to use.

We provide pre-trained models that can be used for truecasing English and German right out of the box. The English model is trained on the standard Wikipedia data split from (Coster and Kauchak, 2011), and achieves an F1 score of 93.01 on test. This is comparable to the best F1 of (Susanto et al 2016) of 93.19.

Requirements

Model

This model treats each sentence as a sequence of characters (spaces are included in the sequence). Each character takes a binary label of "U" if uppercase and "L" if lowercase. For example, the word tRuEcasIng would take the labels LULULLLULL

We encode the sequence using a bidirectional LSTM with 2 hidden layers, 50 dimensional character embeddings (input), 150 dimensional hidden size, and dropout of 0.25.

Scoring

A cautionary note is in order. The pytorch model optimizes for character level F1 score, but it is more common to measure on the word level. So, after training a model, get a comparable score using word_eval.py (which I copied from here)

For example, to score on the Wiki test data:

$ allennlp predict wiki-truecaser-model.tar.gz data/data.v1.split/normal.testing.txt --use-dataset-reader --output-file out_preds.txt --include-package mylib --predictor truecaser-predictor
$ python word_eval.py data/data.v1.split/normal.testing.txt out_preds.txt

Usage

If you just want to predict, you can run:

$ allennlp predict wiki-truecaser-model.tar.gz data/test.txt --output-file test-out.txt --include-package mylib --use-dataset-reader --predictor truecaser-predictor

Where data/test.txt is a file with one sentence per line.

See example.py for an example of how to use it programmatically.

Training

The dataset reader requires text that has one sentence per line. The model expects tokenized text. If your text is already tokenized (the Wiki data is), then you can use just_spaces as the word_splitter in the config. If you want to tokenize text first, you can use spacy.

You can get the Wikipedia data by running:

$ cd data
$ ./get_data.sh

Run:

$ allennlp train truecaser.json --include-package mylib -s /path/to/save/model/

If you have a GPU, set cuda_device to 0 in truecaser.json. This will make training much faster.