In the web 2.0, a large part of online data is generated by users, such as movie reviews, amazon product reviews, hotel reviews. According to a recent survey, 93% of users feel User-Generated Content (UGC) are helpful when making a purchasing decision. Also, the sellers regard the sentiment over UGC as a good indicator for the product or service feedback. Thus, more and more companies built their own sentiment analysis model to monitor their brand reputation. However, the models were built with different qualities. Some are good, and some are bad. and there is a lot of duplicated work. The problem is how to develop some good models and share with others.
What it does
To predict sentiment score of user generated contents. End users can also re-train the models based on their own dataset.
How we built it
Please view our video here: https://youtu.be/ITZW_xZI1Rg or our attached deployment documents
Challenges we ran into
How to retrain the model? How to encapsulate the data preprocessing with the model, thus the end users only need to provide the raw text data? How to build more accurate model? How to generalize the model to various applications?
Accomplishments that we're proud of
Three state-of-the-art deep learning models, which acquire good model performance. Ensemble method to future improve the model performance. Added re-train function to generalize different user scenarios.
What we learned
How to package and share your models to others? How to improve model performance? How to use dockers?
What's next for Sentiment_Analysis_For_Online_User_Generated_Content
Data visualization for our prediction. Deploy the microservice into Amazon cloud and provide the API to the public.