Model Card for Kerassy/bert_base_tp_123k

Classifies sentiment of user review text as either positive, neautral or negative.

• Developed by: Jay Broughton (Kerassy)
• Language(s) (NLP): En
• License: MIT
• Finetuned from model: distilbert/distilbert-base-uncased

How to Get Started with the Model

from transformers import pipeline

reviews = ["Really long wait time for drinks and food and then for food order wrong. Bowling however that was great no issues at all.",
           "Took my son and friends for a birthday party 6 children under 12. A week later get a £200 fine for parking in a free car park!!Apparently you have to type your reg number into something to validate your stay. They don’t care as a third party operate the car park so it’s not there problem if you get a fine!! Everywhere I’ve ever been with one of these systems the person at reception will tell you when booking to enter your details into the machine.  But not Hollywood bowl Oxford!!I see they have replied to me saying to get in touch I already did they said it’s not their problem due to them not owning the car park!",
           "Beyond brilliant! Rachel's energy is something else and she entertains them children thoroughly throughout the entire time of booking, plenty of dancing and music and she matches her routine to the ages of the children.I've used for a couple of years now and have no hesitation in booking her for every kids event we have, She makes it a proper party and the adults love it too!",
           "Terrible Got charged 5 times Over half the times smashed and still waiting for someone to contact me..."]

pipe = pipeline(task="sentiment-analysis", model="Kerassy/bert_base_tp_123k", device="cuda") # or "cpu"

preds = pipe(reviews)
print(preds)

Categories

• 'LABEL_O' - Negative
• 'LABEL_1' - Neutral
• 'LABEL_2' - Positive

Training Data

The model was fine-tuned using the following dataset: Kerassy/trustpilot-reviews-123k

Accuracy

The performance of the model was evaluated using an anonymous Trustpilot reviews dataset sourced from Kaggle, in conjunction with Scikit Learn metrics libraries. The results are as follows:

• Accuracy: 0.9543
• Precision: 0.9670
• Recall 0.9543
• F1-Score: 0.9597