- Use of custom models for different categories (tech, food, books,...) to be automatically (using context classfication) or manually selected(by the client). (different datasets applied)
- Run multiple models per dataset and derive weighted average results.
- Developing a layered classification use fast/slow classification (divide the dataset using confidence index to strong and weak groups; the weak group will be analysed further using Roberta model).
- Aspect based analysis (attach sentiment to specific aspects rather than sentence/opinion) and word cloud (for word frequencies) to show insights of the reviews. (Amazon comprehend model)
- Use of lemmatization, opinion unit extractor, subjectivity index and multiclass classification(love, sad, angry,...) for better accuracy and data enrichment.
- Test of a sent-ngrams lexion sentiment analysis (SO-CAL).
- Use of client dataset to fine-tune the model. (Ideation phase)
- Twitter airline
- IMDB
- Yelp (preprocessing phase)
- Amazon
- 140sentiment twitter
- Textblob
- Flair (RNN)
- Vader
- Roberta-large
- Bert_multilingual