metaphor_identify_glue.py identify metaphors from glue/empathy corpus.
Should use labeling_no_token_type model, since predicting metaphors has no token_type clues.
Then run run_glue.py to have glue results file saved as glue/val_results_with_vua/vua_and_result_{task}.
At the end, using produce_metaphoricity_for_glue() from funcs.py to compute metaphor score and glue results, and make all human-readable.