A full-length pipeline to generate chatbots from content sources
There are thousands upon thousands of calls made each day from people who are looking for basic help. Someone can call looking for help changing their password or looking up information in their account. For many, it is easier and much quicker to call a number and ask a representative rather than look through an entire website for the right help documentation. Handling these calls requires hundreds of people and millions of dollars. And importantly, the information callers are seeking is noncomplex and readily available.
Organizations have already flagged this as a potential area for improvement. As such, some are shifting to using chatbots to interact --- through voice or through text --- with those requesting help. These chatbots are great for understanding and responding to users, but they need to be trained first. And people are training them manually. This takes a long time and effort for every single bot you want to build.
We propose a pipeline to automatically generate a chatbot for given content repository. Not only will this pipeline save time and effort in answering people's questions, but it will also save time and effort in producing the chatbots necessary to answer those questions.
Currently, we're focusing on Frequently Asked Questions sections of websites because they are relatively structured and are directly related to call center type questions. Theoretically, one could analyze what questions their call centers answer and compile an FAQ from that. From there, our pipeline could read in that FAQ and spit back out a chatbot to answer those questions.
Chatbots utilize artificial intelligence to understand and respond to users. However, to do properly understand users, chatbots must be pretrained. Specifically, chatbots must be fed information about the different reasons a user might be talking with them (intents), the different ways a user might express the same intent (sample utterances), and the different types of values a user can mention e.g. location, phone numbers, time (slots).
Our pipeline primarily uses Amazon Web Services (AWS) and Microsoft Azure. From a high-level, we use Microsoft Azure to parse FAQ webpages into question-answer pairs and then we use an AWS service called Lex to create, build, and publish our bot. From there, you can wrap the published bot into any platform e.g. Twilio, Facebook.
Let's take a detailed walkthrough of the process, from start to finish. If you are looking for more detailed developer documentation, look in the subfolders.
/create , /describe and /delete, for creating, describing, and deleting bots. There is currently no authentication or API key required to use these endpoints. When API Gateway receives a GET request, it will invoke an AWS Lambda function. For a /create GET request, the function ProcessQueue will be invoked with the query parameters passed in as part of the event object. For a /describe GET request, the function DescribeBotStatus will be invoked. And for a /delete GET request, the function DeleteBot will be invoked.main.py script on a running EC2 instance. The main.py runs through the entire pipeline. It will interact with a Microsoft Azure service to parse FAQ pages. An API key is required to use Microsoft Azure and that key is stored securely in an S3 bucket. Through a series of API calls, the pipeline scripts will feed Azure the FAQ url, download all the question-answer pairs that it parsed, and then delete the knowledge base from Azure. From here, we have the question and answers from the FAQ. We wrap these up and store them in a new AWS DynamoDB table, iteratively batch writing the intents into the table. Afterward, the scripts will interact with AWS Lex. Lex is the AWS chatbot service. To create a bot, you need intents to give it. And to create an intent, you need to provide sample utterances. The scripts create an intent for each question with the verbatim question text as a sample utterance. A bot is then created with those intents attached. All intents and the bot are built and an alias is applied to it. Right now, the Lex bot understands the different intents a user might express, but it does not know how to respond yet.Demo app download URL to try (screenshots)