len-sla / nlp_flair_texhero_distilbert Goto Github PK

Predict which tweets are about real dissters and which ones are not in a couple lines of code. Data for calculation taken from https://www.kaggle.com/c/nlp-getting-started

If you need to get quickly some initial results in typical NLP task than using packages Flair, Texthero and DistilBERT would give quite good results.

1. Flair Embeddings
2. FastText Embeddings
3. TransformerWordEmbeddings
4. How to use flair with keras

Project is: in progress,

Project inspired by Kaggle nootebook

Rev_B_real_or_not.ipynb Results were worse compared with initial simple automatic approach. That proves how good/opimised Flair Framework is to get best results. Tweaking does not give better results. Maybe more extensive text cleaning and deciphering abbreviation and other shorcuts could do better result. Things like hero.visualization.wordcloud, kmeans, custom_pipeline were checked.

Created by [email protected]