Expanda 1.1.1

Integrated Corpus-Building Environment

Expanda

The universial integrated corpus-building environment.

Introduction

Expanda is an integrated corpus-building environment. Expanda provides integrated pipelines for building corpus dataset. Building corpus dataset requires several complicated pipelines such as parsing, shuffling and tokenization. If the corpora are gathered from different applications, it would be a problem to parse various formats. Expanda helps to build corpus simply at once by setting build configuration.

Main Features

• Easy to build, simple for adding new extensions
• Manages build environment systemically
• Fast build through performance optimization (even written in Python)
• Supports multi-processing
• Extremely less memory usage
• Don't need to write new codes for each corpus. Just write one line for adding new corpus.

Dependencies

• nltk
• ijson
• tqdm>=4.46.0
• mwparserfromhell>=0.5.4
• tokenizers>=0.7.0
• kss==1.3.1

Installation

With pip

Expanda can be installed using pip as follows:

$ pip install expanda

From source

You can install from source by cloning the repository and running:

$ git clone https://github.com/affjljoo3581/Expanda.git
$ cd Expanda
$ python setup.py install

Build your first dataset

Let's build Wikipedia dataset by using Expanda. First of all, install Expanda.

$ pip install expanda

Next, create workspace to build dataset by running:

$ mkdir workspace
$ cd workspace

Then, download wikipedia dump file from here. In this example, we are going to test with part of enwiki. Download the file through web browser, move to workspace/src and rename to wiki.xml.bz2. Instead, run below code:

$ mkdir src
$ wget -O src/wiki.xml.bz2 https://dumps.wikimedia.org/enwiki/20200520/enwiki-20200520-pages-articles1.xml-p1p30303.bz2

After downloading the dump file, we need to setup the configuration file. Create expanda.cfg file and write the below:

[expanda.ext.wikipedia]
num-cores           = 6

[tokenization]
unk-token           = <unk>
control-tokens      = <s>
                      </s>
                      <pad>

[build]
input-files         =
    --expanda.ext.wikipedia     src/wiki.xml.bz2

Current directory structure of workspace should be as follows:

workspace
├── src
│   └── wiki.xml.bz2
└── expanda.cfg

Now we are ready to build! Run Expanda by using:

$ expanda build

Then we can get the below output:

[*] execute extension [expanda.ext.wikipedia] for [src/wiki.xml.bz2]
[*] merge extracted texts.
[*] start shuffling merged corpus...
[*] optimum stride: 17, buckets: 34
[*] create temporary bucket files.
[*] successfully shuffle offsets. total offsets: 102936
[*] shuffle input file: 100%|████████████████████| 102936/102936 [00:02<00:00, 34652.03it/s]
[*] start copying buckets to the output file.
[*] finish copying buckets. remove the buckets...
[*] complete preparing corpus. start training tokenizer...
[00:00:59] Reading files                            ████████████████████                 100
[00:00:04] Tokenize words                           ████████████████████ 405802   /   405802
[00:00:00] Count pairs                              ████████████████████ 405802   /   405802
[00:00:01] Compute merges                           ████████████████████ 6332     /     6332

[*] create tokenized corpus.
[*] tokenize corpus: 100%|█████████████████████| 1749902/1749902 [00:28<00:00, 61958.55it/s]
[*] split the corpus into train and test dataset.
[*] remove temporary directory.
[*] finish building corpus.

If you build dataset successfully, you can get the following directory tree:

workspace
├── build
│   ├── corpus.raw.txt
│   ├── corpus.train.txt
│   ├── corpus.test.txt
│   └── vocab.txt
├── src
│   └── wiki.xml.bz2
└── expanda.cfg