A chatbot is often described as one of the most advanced and promising expressions of interaction between humans and machines. For this reason I wanted to create one in order to become affiliated with Natural Language Processing and Deep-Learning through neural networks.
In light of this contest's theme of adventure, I wanted to tackle a very important issue. If an adventurer were to set out in our current situation, that being the coronavirus pandemic, they would almost definitely be infected.
Due to the current pandemic, we are truly living in an unprecedented time. As the virus' spread continues, it is important for the adventurers of the world to stay educated and updated on the latest information regarding COVID-19, in order to continue on with said adventures. So, I designed a chatbot named Kovi who a user can talk to, and get latest information regarding COVID-19.
(This bot is designed to function similarly to ones used on websites for companies such as Amazon or Microsoft, in which users can interact with the bot to ask questions they would normally ask to a customer service member, although through the power of AI and deep learning, the bot can answer these questions for the customer on it's own)
What it does
Kovi answers questions the user has regarding COVID-19.
More specifically, the training data I fed into my feed-forward neural network to train Kovi falls under 5 categories:
- Deaths from COVID-19
- Symptoms of COVID-19
- Current Cases of COVID-19
- New Technology/Start-up Companies working to fight coronavirus
I also added three more categories of data for Kovi to learn, those being greetings, thanks and goodbyes, so the user can have a conversation with Kovi which is more human-like.
How I built it
First, I had to create my training data. Commonly referred to as 'intentions', the data I used to train Kovi consisted of different phrases that a user could potentially ask. I split up all of my intentions into 7 categories, which I listed above, and these were called 'tags'. Under my sub-branch of tags, I would provide Kovi several phrases the user could ask about that tag, and also gave it responses to choose from to answer questions related to that tag. Once the intentions were made, I put this data in a json file for easy access in the rest of the project. (I used google cloud's COVID-19 data as my intents)
Second, I had to use 3 artificial-intelligence, natural language processing, techniques to process the data, before it was fed into my training model. These were 'bag-of-words', 'tokenization' and 'stemming'. First, bag-of-words is a process which took a phrase, which were all listed under the tags, and created an array of all the words in that phrase, making sure there are no repeats of any words. This array was assigned to an x-variable. A second y-variable delineated which tag this bag-of-words belonged to. After these bags-of-words were created, tokenization was applied through each bag-of-words and split them up even further into individual words, special characters (like @,#,$,etc.) and punctuation. Finally, stemming created a crude heuristic, i.e. it chopped off the ending suffixes of the words (organize and organizes both becomes organ), and replaced the array again with these new elements. These three steps were necessary, because the training model is much more effective when the data is pre-processed in this way, it's most fundamental form.
Next, I made the actual training model. This model was a feed-forward neural network with 2 hidden layers. The first step was to create what are called hyper-parameters, which is a standard procedure for all neural networks. These are variables that can be adjusted by the user to change how accurate you want your data to be. Next, the network began with 3 layers which were linear, and these were the layers which inputted the data which was pre-processed earlier. Next, these were passed on into what are called activation functions. Activation functions output a small value for small inputs, and a larger value if its inputs exceed a threshold. If the inputs are large enough, the activation function "fires", otherwise it does nothing. In other words, an activation function is like a gate that checks that an incoming value is greater than a critical number.
The training was completed, and the final model was saved into a 'data.pth' file using pytorch's save method.
Finally, the user interface for communicating with Kovi was made. To run Kovi, I created a pathway through terminal using pytorch, and a user simply needs to run the chat.py file through command line.
Challenges I ran into
The most obvious challenge was simply time constraints. I spent most of my time trying to make sure my training model was efficient, and had to search up several different articles and tutorials on the correct methodology and API's to use. Numpy and pytorch were the best ones.
Accomplishments that I'm proud of
This was my first deep-learning project so I am very proud of completing at least the basic prototype. Although I was aware of NLP techniques such as stemming and tokenization, this is my first time actually implementing them in action. I have created basic neural nets in the past, but also never a feed-forward one which provides an entire model as its output.
What I learned
I learned a lot about deep learning, neural nets, and how AI is trained for communication in general. This was a big step up for me in Machine Learning.
What's next for Kovi: Deep Learning Chatbot (TigerHacks 2020 Submission)
I will definitely improve on this in the future by updating the model, providing a lot more types of questions/data related to COVID-19 for Kovi to be trained on, and potentially create an external UI, although I do like the intuitive ease of use that comes with the integration to the terminal.