/ngram-lm-wiki

Scripts to train a n-gram language models on Wikipedia articles

Primary LanguagePythonMIT LicenseMIT

ngram-lm-wiki

Train n-gram language models (LM) on Wikipedia articles, mainly intended for Automatic Speech Recognition (ASR) applications.

Dependencies

The following dependencies should be available on your system:

  • KenLM: KenLM is used to create the n-gram LM. Please see KenLM docs for more details.
  • Scripts are tested with Python 3.7 and required packages are listed in the requirements.txt.

Process Wikipedia Dump

The script process_wiki.py currently support Turkish language in terms of sentence tokenization but can be easily configured for other languages.

  1. It extracts Wikipedia (e.g. trwiki-latest-pages-articles.xml.bz2) article dump into individual json files.
  2. Then, loads each extracted wiki file and tokenizes & cleans sentences for LM training.
  3. Finally, it saves tokenized and cleaned sentences in the output directory.

Example usage:

python process_wiki.py \
    --wiki_dump trwiki-latest-pages-articles.xml.bz2 \
    --output data \
    --language_id tr \
    --processes 8

Train n-gram LM

KenLM binary lmplz is used to create a n-gram LM in arpa format and then the arpa LM is converted to binary format using build_binary.

The following usage creates a 4-gram LM with top 200000 most frequent words and pruning.

For ASR applications, --vocab option can be used with ASR output vocabulary of characters to filter unigrams.

python generate_lm.py \
    --input data/sentences.txt \
    --output lm \
    --kenlm_bins kenlm/build/bin \
    --vocab vocab.txt \
    --top_k 200000 \
    --order 4 \
    --prune "0 0 1" \
    --binary_type "trie" \
    -a 255 \
    -q 8 \