EricPostMaster
commented
2 years ago Added trf to requirements.txt and loaded it into fortune_cookie.py. Performance has improved, but I wonder how it might differ with the en_core_web_lg (https://spacy.io/models/en#en_core_web_lg) model. According to the spaCy website, the trf model outperforms the lg model in various accuracy metrics, including part-of-speech tagging, sentence segmentation, and named entities. Maybe I'll try it if I find myself with decent wifi to download the 560MB model.
The current app only uses the
en_core_web_smlibrary, but thetrflibrary is what was originally used. The transformer model is too large for loading with the website (460 MB), so maybe we can just do all of the rule-based transformations ahead of time and load the completed movie synopses into the application.