Dividing the ontology alignment task with neural embeddings

This repo contains two scripts that prepare clusters of ontology classes from the lexical index, which is computed with LogMap. The clusters themselves are computed from the neural embeddings computed with StarSpace. Therefore it is required that you have starspace installed somewhere on your system.

Python packages requirements

Install requirements from requirements.txt with

pip install -U pip
pip -r install requirements.txt

Convert LogMap index into StarSpace-compliant format

First, you need to convert LogMap stem -> concept_index mappings of the form

stem_1;stem_2|concept_index1,concept_index2

into the form which is required by the starspace toolkit

word1 word2 __label__concept_index1 __label__concept_index2

So we assume that we have mappings (stem -> word) and (concept_index -> __label__concept_index).

To convert your LOGMAP_INDEX call the script like so

python3 logmap_to_starspace.py LOGMAP_INDEX

# see `python3 logmap_to_starspace --help` for help

Computing clusters

compute_clusters.py takes in the converted lexical index files found in LOGMAP_DIR, and per each lexical index file it computes embeddings with starspace: word1 x1 x2 ... xn. Then, it computes aggregated vectors, and finally outputs clusters per set of words. This script will populate:

• CONVERTED_DIR with starspace model and embeddings
• RESULTS_DIR with the clusters
• LOG_DIR running time and other logging info

see python3 compute_clusters.py --help for more.

Set the variable STARSPACE inside compute_clusters.py to point to the path to your local binary of starspace.