es-categorizer

Extracting categories from an arbitrary block of text using Elasticsearch and Wikipedia.

TODO:

• Get better wiki article results.
  • Adapt according to text size?

How?

Given some text, term vectors and the OpenNLP Ingest Processor plugin can be used to extract keywords that exist only within that text, which is not suitable for category extraction. What is needed is a dataset that includes indexable content paired with tags, like an encyclopedia:

1. Index Wikipedia. See this guide and Setup below.
2. Use a More Like This Query to find documents (wiki articles) that match a given text.
3. Gather the wiki categories of the highest scoring documents.
4. Return a general category based on those categories.

Prerequisite Dependencies

• python 3
• elasticsearch 5.5.2
• ICU Analysis Plugin
• Wikimedia's Extra Queries and Filters API Extention Plugin
  • Also referred to as Elasticsearch Trigram Accelerated Regular Expression Filter

If you're using OS X and homebrew:

brew install python3
brew install elasticsearch
brew services start elasticsearch  # or just "elasticsearch" for foreground execution
elasticsearch-plugin install analysis-icu
elasticsearch-plugin install org.wikimedia.search:extra:5.5.2

Setup

1. Download the latest dump of the search index.

curl -O "https://dumps.wikimedia.org/other/cirrussearch/current/enwiki-20170925-cirrussearch-content.json.gz"
# If the URL is invalid, go to https://dumps.wikimedia.org/other/cirrussearch/current/ to find an alternative

2. Install this program's dependencies: pip install -r requirements.txt.

   • You should probably do this in a virtual environment.

3. Run ./main.py.

   • You will be prompted for the path of the file you downloaded in step 1.
   • If you've got limited storage, consider setting delete to True for load_chunks() beforehand.
   • This command is equivalent to:

./main.py delete_index
./main.py create_index
./main.py chunk /path/to/dump
./main.py load

4. Test wiki category extraction with ./main.py extract "Some text".

Sample output for ./main.py extract -s 3 "The Cubs are destroying the Mets right now.":

1) 21.56842 "1969 Chicago Cubs season"
   - Use mdy dates from November 2013
   - Articles with hCards
   - Chicago Cubs seasons
   - 1969 Major League Baseball season
2) 21.366503 "2015 Chicago Cubs season"
   - Use mdy dates from August 2015
   - Articles with hCards
   - 2015 Major League Baseball season
   - 2015 in sports in Illinois
   - Chicago Cubs seasons
3) 21.28783 "2015 National League Championship Series"
   - Pages using deprecated image syntax
   - 2015 Major League Baseball season
   - National League Championship Series
   - Chicago Cubs postseason
   - 2015 in sports in Illinois
   - 21st century in Chicago
   - 2015 in sports in New York City
   - New York Mets postseason
   - October 2015 sports events

NOTE: Results may differ between search index versions.

5. Test general category extraction with categorize.

./main.py categorize "The Cubs are destroying the Mets right now." should yield sports.