Extractor

Extractor is a tool for fast monolingual data collection from Common Crawl data. It processes Common Crawl WET files, removes invalid UTF-8, identifies languages and outputs file-separated monolingual data. Text is classified by language using Google's CLD2 language detector (https://github.com/CLD2Owners/cld2).

Following is an example of french output file:

...
f6fa1abb58549287111ba8d776733e9 uri:http://blogalolotte.canalblog.com/tag/pomme%20de%20terre language:fr bytes:59
pomme de terre : Tous les messages sur pomme de terre - Le
df6fa1abb58549287111ba8d776733e9 uri:http://blogalolotte.canalblog.com/tag/pomme%20de%20terre language:fr bytes:81
...

Each header line starts with the same hash code followed by metadata. Then, the extracted language-speific text (here french) starts on a new line and is ended by the next header line.

Apart from this, Extractor can also produce statistics on language distribution per domain.

Requirements

• GCC, C++11 compiler
• Boost 1.58.0 or later (Boost 1.65.1 or later for LZMA compression)
• Cmake 3.5 or later
• curl, libcurl4-openssl-dev

Installation

mkdir -p build && cd build
cmake .. -DWITH_LZMA=off -DBUILD_TEST=on -DCMAKE_BUILD_TYPE=Release
make -j 4

-DWITH_LZMA=on enables LZMA compression but requires Boost version 1.65.1 or higher.

Usage

Extractor can be used to process monolingual data with the following command:

cat <input_file> | ./mono --icompression <compression> --ocompression <compression> --workers <number_of_workers> --output <output_folder>

Extractor reads from standard input and expects a line-separated list of path names. Such list is then processed in parallel with the number of workers specified as an argument.

icompression and ocompression arguments set the input and ouput compresion formats respectively. The following compression formats are supported: gzip|bzip2|zlib|lzma|none.

One folder per worker is created in the output folder and each such folder contains file-separated monolingual data. Processed files are logged in done.log file located in each folder.

Optional arguments:

• --curl: uses cURL to download remote files
• --print_stats: prints language statistics

If --curl option is used, the input is expected to be a list of URL addresses instead of a list of path names. Extractor also supports multiple URL sources; If one URL source fails, it attempts to download the content from another one. To define additional sources, just add more URLs on the same line separated by a single space. Any cURL errors are logged in errors.log file in each worker folder.

If --print_stats option is enabled, it creates additional stats folder in each worker folder, and outputs language statistics on the processed files. Each file contains statistics on language distribution per domain with the following format:

<domain>	<language code>	<byte length>

Each file contains no duplicates and is alphabetically sorted.

Known Issues

CLD2 detects almost 300 different languages when processing Common Crawl. Since each worker opens a new file for every detected language, the total number of opened files might exceed the user's hard or soft resource limits on open file descriptors. You can type ulimit -a to get a list of all current resource limits.

For the Common Crawl purposes, executing the following command should increase the limit of open file descriptors sufficiently:

ulimit -n 32768