Grammatical Profiling for Semantic Change Detection

This is the repository for the CoNLL-2021 paper "Grammatical Profiling for Semantic Change Detection" by Mario Giulianelli, Andrey Kutuzov, and Lidia Pivovarova.

The full_profiling_pipeline.sh script can be used right out of the box.

Run the script as

./full_profiling_pipeline.sh TARGETS CORPUS0 CORPUS1

• TARGETS is a list of target words, one per line
• CORPUS0 and CORPUS1 are your earlier and later corpora, tagged and parsed into the CoNLL-U format. Any tagger and parser of your choice can be used (UDPipe and Stanza come to mind immediately).

The script will process your CONLL-U files and generate morphosyntax-based predictions of graded and binary semantic change for each target word. Find the aggregated change scores as predicted by our algorithm in the output directory. The script already uses the best hyperparameters from the paper.

Do not forget to first install dependencies:

pip3 install -r requirements.txt

In case you are interested, the output/tsv directory contains separate predictions from morphological and syntax features.

More details

• stanza_process.py produces parsed CONLL files from raw texts using Stanza;
• collect_ling_stats.py reads a CONLL file and dumps frequencies for morphological and syntax properties of the target words into JSON files;
• compare_ling.py produces TSV files with binary and graded change predictions from two JSON files;
• merge.py reads two JSON files (e.g. with morphological and syntax properties) and produces combined predictions.

Use eval.py to evaluate the resulting TSVs with regards to gold scores.

The raw profiles (feature counts) from the paper are in the /features directory.