RSR

A testing environment for "Refined Semantic Relatedness" (RSR), a distributional semantics model aiming to quantify how close two texts are conceptually. RSR uses an incrementally improvable index as a basis for assessing semantic relatedness. This work is a part of our Bachelor's degree project.

By Per Fahlander and Mattias Bergsström.

Usage

Calibrate the index:

./SSR/indexer <file_or_dir_to_learn_from> <association_span>

• <file_or_dir_to_learn_from> : A document file (or a directory containg such) with plain text that can be used for associating words in order to know which go together.
• <association_span> : How far two words can be apart and still get associated together.

List associations in index:

./SSR/indexer --list

Remove index to start over:

rm -r SSR/index

Purge (stop) words in index:

./SSR/indexer --purge [--preview]

• [--preview] : Shows what words would be removed without action

Benchmark SSR / evaluate SSR against a set of tests (human opinions): Shows the results from specified tests and calculates a Pearson correlations between the algorithmic and human assessments.

./benchmarker <tests_csv_file> [--sort]

• <tests_csv_file> : the set of tests for the evaluation (e.g. "LeePinCombeWelsh/complete.sh")
• [--sort] : The tests SSR performed the worst on will be placed at the end.

Resources

LeePincombeWelsh/
  complete.csv          # The data for the 12,227 tests
  short.csv             # The data for the short test of 6 text pairs.
  ... 50 text files     # All text files are located here as well...

Samples/                # The samples folder includes different text files which can be used to test the Indexer.

SSR/
  indexer.sh
  ...