Skip to content

Latest commit

 

History

History
107 lines (81 loc) · 5.9 KB

README.md

File metadata and controls

107 lines (81 loc) · 5.9 KB
Raw

Large XML Corpus Extractor

This repository is based on A Dissimilarity-based Concordancer and provides a tool which can be used to index and query a large XML-based text corpus using Elasticsearch. The results are then writable to a text file. This particular project was developed for the Icelandic Gigaword corpus (IGC; Steingrimsson et al. 2018) and contains sample data from this corpus.

Table of contents

Description

There are two modules included in this package: source and data.

source

  • initialize_index.py : constructs an index from the zip file(s) in data and is specified for maximum sentence length
  • query.py : queries the indexed corpus using a regex match and writes all results to a text output (very slow, but recommended for very big queries)
  • query_minimal.py : similar functionality to query.py but returns results faster and can be specified for different output sizes (can also output all results, but not recommended for very big queries)

data

  • Gigaword_small.zip : a single XML file in a nested directory
  • Gigaword_large.zip : multiple XML files in the same nested directory structure
  • corpus.json : metadata that you supply about your corpus

Setting up Elasticsearch

  1. Install Elasticsearch
  2. Get Elasticsearch up and running.
    • You can navigate to its installation directory cd YOUR_PATH/elasticsearch-7.13.2 and then run its binary bin/elasticsearch-7.13.2
    • Or, if installed using Homebrew on macOS, then you can start it as a service brew services start elastic/tap/elasticsearch-full
  3. Check curl http://localhost:9200/ to make sure that Elasticsearch has started

Running the Extractor

Step 1: Index the Corpus

In this step, we use XML header tags to ensure that every word in the corpus is a searchable string of the format token0lemma0tag. The separator is always 0. Sentence structures are preserved, and so when queried, every match will be returned embedded in its original sentence.

Navigate to the source directory cd YOUR_PATH/igc-corpus-reader/source and run the indexer, which takes 3 arguments (the corpus zip file --zip_file, the corpus metadata file --corpus_json, and the maximum sentence length --max_sent_len):

Only one zip file

python3 initialize_index.py --zip_file="Gigaword_small.zip" --corpus_json="corpus.json" --max_sent_len=1000

Multiple zip files

If you have multiple zip files in data that you would like to index, then you can create a list of all the zip files you'd like to index (e.g., Gigaword_list.txt) and use a simple for-loop:

for f in `cat Gigaword_list.txt`; do python3 initialize_index.py --zip_file=$f --corpus_json='corpus.json --max_sent_len=1000'; done

Reindexing the corpus (important!)

The indexed corpus will be saved to igc-corpus-reader/bin and so this indexation step only needs to be run once. If you want or need to re-index the corpus, remember to delete all the files in bin and run curl -XDELETE http://localhost:9200/bin or else the new index will just append to the old one and you could end up with duplicates.

Step 2: Query the corpus

Fast query (query_minimal.py)

This takes 2 arguments (the regex query itself and the maximum number of results to write to a csv):

python3 query_minimal.py "hola0nvfe" 10000

This will look for any string matches in the corpus with the sequence hola0nvfe, and 10000 indicates that 10,000 results will be printed to an output csv. If you want to print all results to the output csv, you can enter -1.

The results in the output csv will be printed as one sentence (each containing at least one match) per row.

Long query (query.py)

Navigate to the source directory cd YOUR_PATH/igc-corpus-reader/source and run the query, which takes 1 argument (the regex query):

python3 query.py "hola0nvfe"

This will look for any string matches in the corpus with the sequence hola0nvfe...

1:	<regex.Match object; span=(9, 23), match='Hola0hola0nvfe'>:	,,0,,0ta Hola0hola0nvfe Espania0Espania0nven-s !0!0!
2:	<regex.Match object; span=(42, 57), match='holna0hola0nvfe'>:	Á0á0af einum0einn0tfkeþ átján0átján0tfvfe holna0hola0nvfe golfhring0golfhring0nven gengur0ganga0sfg3en maður0maður0nken níu0níu0tfkfþ til0til0aa tíu0tíu0tfkfo kílómetra0kílómetri0nkfe .0.0.
3:	<regex.Match object; span=(88, 117), match='vinnsluholna0vinnsluhola0nvfe'>:	Það0það0fphen fól0fela0sfg3eþ í0í0af sér0sig0fpheþ borun0borun0nveo tveggja0tveir0tfvfe vinnsluholna0vinnsluhola0nvfe og0og0c varmaskiptastöðvar0varmaskiptastöð0nvee .0.0.

...and will then print a 4-column output csv (and not a sentence, as in query_minimal.py):

source token lemma tag
data_MIM_morgunbladid.zip.40579752 borhola borhola nvfe
data_MIM_morgunbladid.zip.40635343 borholna borhola nvfe
data_MIM_morgunbladid.zip.40686869 holna hola nvfe