https://github.com/kensk8er/langdist
Multilingual Language Modeling Toolkit
https://github.com/kensk8er/langdist
character-embeddings data-science deep-learning language-model lstm machine-learning multilingual natural-language-generation natural-language-processing neural-network nlp python recurrent-neural-networks tensorflow
Last synced: 3 months ago
JSON representation
Multilingual Language Modeling Toolkit
- Host: GitHub
- URL: https://github.com/kensk8er/langdist
- Owner: kensk8er
- License: mit
- Created: 2017-03-12T13:49:48.000Z (over 9 years ago)
- Default Branch: master
- Last Pushed: 2017-05-25T12:15:22.000Z (about 9 years ago)
- Last Synced: 2025-12-15T12:16:17.283Z (6 months ago)
- Topics: character-embeddings, data-science, deep-learning, language-model, lstm, machine-learning, multilingual, natural-language-generation, natural-language-processing, neural-network, nlp, python, recurrent-neural-networks, tensorflow
- Language: Python
- Homepage:
- Size: 5.91 MB
- Stars: 11
- Watchers: 2
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
langdist - Character-level Multilingual Language Modeling Toolkit
=================================================================
`langdist` is a Python project for experimenting *Character-level Multilingual Language Modeling*, which is to see how learning a character-level language model in one language helps learning another character-level language model in a different language. The project is still **under development** and can offer limited functionality.
Features
--------
- Download and preprocess multilingual parallel corpora ([Multilingual Bible Parallel Corpus](http://christos-c.com/bible/))
- Train a *monolingual language model*
- This is a language model trained in one language
- Train a *bilingual language model*
- This is a language model that is trained on top of another language model (the parameters are initialized using another language model's parameters)
- Generate texts using a trained language model
Installation
------------
- This repository can run on Ubuntu 14.04 LTS & Mac OSX 10.x (not tested on other OSs)
- Tested only on Python 3.5
`langdist` depends on [NumPy and Scipy](https://www.scipy.org/install.html), Python packages for scientific computing. You might need to have them installed prior to installing `langdist`.
You can install `langdist` by:
```bash
pip install langdist
```
This installs `langdist` package to your Python, as well as `langdist` command and add it to your `PATH`.
`langdist` also depends on `tensorflow` package. In default, it tries to install the CPU-only version of `tensorflow`. If you want to use GPU, you need to install `tensorflow` with GPU support by yourself. (C.f. [Installing Tensorflow](https://www.tensorflow.org/install/))
Usage
-----
After installing, `langdist --help` will print help of how to use `langdist` command.
### 1. Download and preprocess a corpus
`langdist` implements a command to download and preprocess a corpus from [Multilingual Bible Parallel Corpus](http://christos-c.com/bible/). The following command will download an English corpus and save it to `./en_corpus.pkl`.
```bash
langdist download-bible en en_corpus.pkl
```
Note that `en` here is the language code of English. Specifying an invalid language code will raise an error message that shows the valid language codes.
### 2. Fit an encoder on the characters used in corpora
You need to fit an encoder to the character used in corpora before you train a language model on them. Note that the same encoder will be used when you train a new language model on top of another language model (*multilingual language model*). Therefore, you need to fit an encoder to all the corpora you will train multilingual language models on.
The following command will fit an encoder to English, French, and Japanese corpora and save it to `./en_fr_ja_encoder.pkl`:
```bash
langdist fit-encoder en_fr_ja_encoder.pkl en_corpus.pkl fr_corpus.pkl ja_corpus.pkl
```
Note that `xx_corpus.pkl` is a pickle file of a corpus, which can be generated by `langdist download-bible` command. You can also create a list of texts by yourself and save it to a pickle file. (Each element of the list would correspond to a segment such as sentence, paragraph, article, etc. depending on your purpose.)
### 3. Train a language model from the scratch (*monolingual language model*)
The following command will train a French language model and save it to `./fr_model` directory:
```bash
langdist train fr_corpus.pkl en_fr_ja_encoder.pkl fr_model --patience=819200 --logpath=fr.log
```
Note that using an encoder that was not fit to the corpus will throw an exception. `--patience` option specifies how many iterations you want to keep training and `--logpath` option specifies the path to the log file that records the progress of the training (no log file will be created if you don't specify the option).
During the training, various stats are dumped to `path_to_model_dir/tensorboard.log` directory. You can visualize them using `tensorboard` by `tensorboard --logdir=path_to_model_dir/tensorboard.log`. The model is saved every time after computing validation perplexity and is available to use before finishing the training.
Check the output of `langdist --help` to know what other options are available for training a language model.
### 4. Train a new language model on of another language model (*multilingual language model*)
The following command will train an English language model on top of the French language model we have trained and save it to `fr2en_model` directory:
```bash
langdist retrain fr_model en_corpus.pkl fr2en_model --patience=819200 --logpath=langdist.log
```
Note that you don't have to specify the path to an encoder because the model in `fr_model` includes it. If the encoder that was used when training `fr_model` was not fit to characters in `en_corpus.pkl`, it will throw an exception.
During the training, various stats are dumped to `path_to_model_dir/tensorboard.log` directory. You can visualize them using `tensorboard` by `tensorboard --logdir=path_to_model_dir/tensorboard.log`. The model is saved every time after computing validation perplexity and is available to use before finishing the training.
Check the output of `langdist --help` to know what other options are available for training a language model.
### 5. Generate texts using a trained language model
Once you have trained a language model, the following command will generate texts using the trained language model:
```bash
langdist generate fr2en_model --sample-num=50
```
`--sample-num` option decides the number of texts to generate. Note that each text is independently generated (sampled) by the language model.
Check the output of `langdist --help` to know what other options are available for training a language model.
### Use `langdist` from Python
`langdist` can be used as a normal python package by importing `langdist` package, which is installed to your Python environment by `pip install langdist`. Reading `langdist/cli.py` is a good way to figure out how to use the package.
Model Architecture
------------------
The language model is implemented using Character-level Multilayer LSTM. The architecture is roughly as follows:
1. Character Embedding Layer
2. 1st LSTM Layer
3. 2nd LSTM Layer
4. Fully Connected Layer
For the details, look at `_build_graph()` method in `langdist/langmodel.py`, which implements the computational graph of the architecture in `tensorflow`.
TODO: Add a link to the blog post *Bilingual Character-level Neural Language Modeling*