Building on the Semantic Text Similarity datasets of SemEval, this repository seeks to evaluate computationally-efficient approaches to identify short text segments that have nearly the same semantic meaning in large-scale datasets.
The main entry point to the code in this repository is test_textsim.py.
That file contains fuller comments describing the approaches being evaluated.
There are several similar files that build upon test_textsim.py.
test_unisent_multilingual.pyevaluates the Multilingual Universal Sentence Encoder (which requirestensorflow>=2.0.0see comments in file)test_flair.pyevaluates various transformer embeddings using the flairNLP librarytest_hindi.pyuses XLM-R embeddings (from flairNLP) to test performance on a new Hindi dataset. That dataset is not in this repository currently. Please contact the maintainer (see below) if needed.
density_plots.R plots out results, and mwe* files are minimal working examples for some approaches.
Further information is available from Scott Hale. Meedan team members can contact Scott via Slack and others can reach out to Scott via comments/issues on this repository or via direct message on Twitter