Spam Slayer is a webapp used to detect deceptions from Amazon products' reviews. Users can copy a link to the product and paste directly paste it to our page. Using our trained deep learning model, Spam Slayer will distinguish between true/fake reviews and show them an adjusted rating without fake reviews.
The main program for this project is ./span_slayer_runner.py, which also has a copy in the front-end folder ./Spam_Slayer/slayer/
We used dataset found on Kaggle, which is adapted from raw data provided by Amazon. The data contains 21,000 reviews, 10,500 labeled as truthful and 10,500 labeled as deceptive.
We have our own word-embedding generators, using vocab from the abovementioned dataset. Using two convolution layers, one for sentence convolution and another for document convolution, a maxpooling layer and a softmax layer for classification, the final validation accuracy is ~66%. Because of the limit of time, we don't have the chance to improve our architecture, but we got a significant 26% enhancement compared to human's success in distinguishing deceptive reviews (~40%).
We built a web application with python and Django framework. Our app crawls data from a given url (currently it works for an url from Amazon, but similar websites like Yelp can easily be done), and turn a list of reviews to our trained model to classify whether each review as real or computer-generated. They are then displayed on the UI.
To use the app on localhost:
pip install django
python manage.py runserver
Our model is based on research by Yingqi Ding: https://github.com/dyq0811/Opinion-Spam-Detection-BiRCNN