An encoder-decoder architecture chatbot with an attention mechanism, designed to provide people wit a virtual companion.
The chatbot was trained on google collab using the depression data available on reddit.
The project was divided into 5 stages.
- Data Collection : The data for training the chatbot was collected from Reddit using the name tags depression and psychiatry. The data was extracted using PRAW (A python reddit API wrapper). A dataset of around 3750 sample points was built.
- Data Cleaning : All the spelling mistakes, word contraptions, symbols were corrected and removed. The dataset was cleaned for appropriate sample points.
- Building Model Architecture The model was using encoder decoder architecture for Recurrent Neural Network, along with Bahdanau attention mechanism for better model performance via subclassing.
- Model Training : Carried out on Google Colaboratory files. Model was trained only for 50 epochs, to provide a decent result, due to the computational expense.
- Model Evaluation :
Performance was kept track of recording the loss values at each epoch.
Further attention analysis can also be carried out.