We created an Artificial-intelligence powered chatbot to serve the customers of our online clothing store: Hudson's Drip. We used an agile SDLC methodology, with our development cycle consisting of two scrums.

Our software consists of two input JSON files used to train the neural network and make personalized answers based on customer information. The training.py imports data from intents.py and trains the neural network. training.py outputs words and classes pk1 files, which are used by the predict.py module to estimate the probability that a user input addresses a particular topic in the intents.JSON. A random response from the most likely topic is outputted by the chatbot. Our bot is also able to empathize with the user and take into account the mood of the user when generating responses.

Several libraries were used: tkntr, random, json, pickle, numpy, nltk (Porterstemmer, word net lemmatizer, Sentiment intensity analyzer), and tensorflow.

JIRA Roadmap.
Reference material.

• Guiherme Durvan António Zandamela
• Lakshay Karnwal
• Abdulaziz Almutlaq
• Ravil Bigvava
• Jordan onwuvuche

The GUI was implemented using the Python tkinter module for graphical interfaces. Using this module we created a text box where the user enters their query/text. This helps the app become intuitivey easy-to-use for the user.

The Automated Unit Testing Framework used for this was pytest due to its easy to use module functions. We created multiple test cases for all the important functions which checks if all the functions are working as desired. You can run the Unit testing file using the pytest command. After running the test file, pytest displays a summary of failed and passed functions. If the functions do not pass that means the chatbot will have errors, if all the functions pass that means the program is working as desired.

In the screenshot example above, one of the functions failed because the value of probability was not correct. This is a useful feature as it reduces effort to debug the code.

We used a natural language toolkit "Sentiment intensity analyzer" to obtain positivity and negativity scores from the words in the user input. This information was used to make conditional statements in which the response of the bot is either complemented or replaced (depending on the intensity) by a response designed to address the sentiment of the user.

Word stemming was used in addition to unstemmed input. Our algorithm stems the user input if the bot is not able to find a probable intent (probability > 0.50). This helps to prevent inaccurate responses and can address suffixes.

Flickr API: The Flickr API helps the user to show what the words used by the chatbot mean and it shows the associated images with that particular word used in the conversation held. It displays relevant pictures about the words on request made by the customer.

Wikipedi API: The Wikipedia API helps the user be redirected to a Wikipedia page if they want to learn more about something. The user is being given a link to read and learn more. This helps the customers get more information on request and help them make wise decisions.