Capstone project for school, uses NLP via a CNN-LSTM model to analyze strings of text for sentiment, assigning a float value of 0.0 - 1.0 (Neg-Pos). Was trained using sentiment140 with binary labeling and implemented to learn to predict on a continuous scale.
The sentiment model for predictions needs downloaded from
https://f004.backblazeb2.com/file/school-bucket/sentiment_model.pt as it is too large to host on github.
A CLI interface is provided to make predictions on single string of text, file of multiple sentiment strings and perform evaluation of the model against test sets. Several graphing functions are also included.
Model performs with an accuracy of 87.55 against the test set from the sentiment140 dataset, and with 90% accuracy when evaluated against a set of 60 GPT4 generated sentiments.
Doesn't perform well with small strings of text, as they tend to just get assigned in the neutral(0.5~) range, but performs relatively well on long strings of text.