The goal of this project is to be able to predict what rating customers on social media platforms would give to products. This enables businesses to better understand what customers think of their products as social media platforms such a Twitter and Youtube do not have rating systems.
This web application can search through Twitter and extract tweets which relate to a given keyword and classify the tweets into 5 categories. These categories represent ratings (out of 5) where 1 is bad and 5 is excellent. Ideally, the keywords should be products but, the webpage can also take in just about anything so long as people are talking about it on Twitter.
This web application utilizes a neural network and BERT (Bidirectional Encoder Representations for Transformers) to make the classifications of the tweets. The machine learning models are based on the Is Bigger Better? Text Classification using state-of-the-art BERT with limited Compute research paper by: Ayaz Nakhuda, David Ferris and Jastejpal Soora. This paper can be visted using this link: https://github.com/AyazNakhudaGitHub/BERT_Customer_Reviews_Classification/blob/main/Report_Group_24.pdf
Python, Django, Flask, HTML5 and CSS3 were mainly used.
To run this project locally one will need to:
-
Download all the files and maintain the directory structure.
-
Ensure that all neccessary modules are installed.
-
Download the machine learning models from this Google drive link: https://drive.google.com/drive/folders/1S1WCRVz2fr8oZd7EysAFK36Lhwab0Eb6?usp=sharing
-
Place the machine learning models into this directory for the API: BERT_API/test/ml_models
- Get the credentials for access to the Twitter API and input them into the file sentiment_BERT_Web_Project/sentiment_BERT_Web_Project/views.py
- Run the API as seen in the image below:
- Enter this command to get the wepage running: python manage.py runserver
Future plans to host this web application and the API on the Google Cloud Platform is currently in the works.
While a GIF is included, a video is provided to give a live demo: