COVID-19 Chatbot with Rasa 2.0: open source conversational AI
Table of Contents
introduction
As natural language processing (NLP) technology and chatbot systems over the past few years have evolved quickly, also the usefulness of chatbots has increased. The motivation of chatbots is productivity; they have an instant access to information they refer to and are efficient in assisting users. (Brandtzaeg, 2017, Why people use chatbots. COVID-19 chatbot is an excellent use case example for the technology.
The content of a chatbot consists of the personality, conversation flows and the information it can deliver to the user. Personality is created by interactions and responses and by acting differently in different situations. These responses should be designed so that it maximises the engagement between the bot and the user (Katz, 2019, The Ultimate Guide to chatbot personality, Chatbots Magazine). The COVID-19 chatbot described here aims to use these principles, however due to the efforts required, in a rather minimalistic way leaving plenty of room for future improvements. e.g. in the area of how to handle chitchat.
COVID-19 data
The COVID-19 data format chosen here is defined by (https://api.rootnet.in/covid19-in/stats/history), which provides COVID-19 data freely for developers.
Conversation Flow
The conversation is initiated by the end-user. A greeting or a goodbye should reset any prior assumptions or knowledge collected by the bot during previous interactions. When time or COVID-19 detail are not contained in the query, the bot shall report the current and generic COVID-19 conditions. When the city is not provided in the query, the bot shall request for it. Any further specifics in the query should be answered in more detail if information is available.
3 Major queries that a user can perform apart from greeting,goodbye etc:
As this is a demo modal of how to integrate API in rasa. So I have designed this COVID-19 bot to answer limited queries as follows:
Query1
user can ask for current confirmed covid cases of any city in India
Examples:
-what is the number of cases currently in Delhi?
- how's the covid currently in Delhi?
- Tell me about covid currently in Maharashtra
- covid cases in Punjab currently
Query2
user can ask for total current confirmed cases of any two cities in India
Examples:
- Hey, what’s the total count of confirmed cases in Delhi, Maharashtra altogether?
- total confirmed cases in Delhi and Maharashtra together?
- total cases currently in Delhi and Maharashtra
- Get me the total current confirmed cases in Delhi and Maharashtra
- Tell me the total current confirmed cases in Delhi and Maharashtra
Query3
user can ask total cases between any dates(format: yyyy-mm-dd)
Examples:
- What’s the confirmed case count from 2020-10-01 to 2020-10-12?
- Hey, what's the current confirmed cases from 2020-10-01 to 2020-10-12?
- currently cases from 2020-10-01 to 2020-10-12?
- Hey, what’s the total count of confirmed cases from 2020-10-01 to 2020-10-12?
- total confirmed cases from 2020-10-01 to 2020-10-12 in India?
- total cases from from 2020-10-01 to 2020-10-12? in India
- Get me the total current confirmed cases from 2020-10-01 to 2020-10-12? in India
- Tell me the total current confirmed cases from 2020-10-01 to 2020-10-12? in India
Implementation
All components are defined to support the conversation flow . The end-user intents here are: who_are_you, covid_in_city, covid_in_two_city, covid_in_period, covid_without_city, greet, goodbye, affirm, deny, mood_great, mood_unhappy, bot_challenge, how_are_you, capabilities In Rasa, the slots can be used for passing information to and back between Rasa and external actions. Three slots are required: city,city2, init_date,final_date.
The responses where the personality is also largely created are: utter_greet, utter_goodbye, utter_ask city (triggers city_form), utter_iamabot, utter_capabilities, utter_im_well so on. This also includes the external action, action_covid, which fetches the COVID-19 data, parses it and generates the COVID-19 response sentence.
External actions are user defined functions written in python. Only one action, action_covid, is required. It is split in two separate functionalities here: actions.py which receives slots: city,city2,init_date and final_date from Rasa. It then queries the COVID-19 data for specific city from covid_api.py where a function covid_data(city,city2,init_data,final_data) is defined. The function returns the COVID-19 data(totalcases) after getting filtered out from (https://api.rootnet.in/covid19-in/stats/history) onecall json format to action_covid, which then forms a response sentense to be passed back to Rasa.
The user intents, stories and rules are used for training the NLP model. These intent examples cover tens of different ways of asking questions, and explaining to the model how to find the values for the three slots and what is the intent the user has. The stories contain the conversation flows and rules that will stop any conversation and force a different path.
Installation
Installation assumes existing installation of miniconda or anaconda. https://www.anaconda.com/
pip3 & Rasa
Below are the simple steps for creating a virtual environment, install pip3 and Rasa Open Source 2.0.
conda create -n RasaEnv python=3.7.6
conda activate RasaEnv
conda install -c anaconda pip3
pip3 install rasa==2.8.11
In case of issue, please refer to Rasa Open Source installation pages: https://rasa.com/docs/rasa/installation/
Creating and initialising a new project:
mkdir rasa
cd rasa
rasa init --no-prompt
This will create a new directlry, under which rasa creates all necessary directories and files.
Replace all files in the rasa directory with the files in the project.
Train the model and run the bot
Train the model with command
rasa train
There are additional actions that need to be started before starting the bot evaluation. These are in actions.py
and covid_api.py
files. To do so, run below commands on two different terminals:
rasa run actions
Start the discussion with rasabot:
rasa shell
References
- Rasa. (n.d.). Rasa: Open source conversational AI. URL: https://rasa.com