Last month, I documented my experience “dogfooding” our AI chatbot, Wendy, to help us find a great Natural Language Processing (NLP) Data Scientist. I outlined a general understanding of how Wendy works but this time around, I’d like to get under the hood and deep dive into the things that make Wendy great. In this post, I’ll detail the tasks she performs, the processes that enable her functionalities, and what makes her an amazing AI chatbot.
What Does an AI Chatbot Do?
Our AI chatbot, Wendy, performs four main tasks:
1. Analyzes Job Descriptions
When a new job enters our system, Wendy reads the description and identifies essential information—including title, location, main responsibilities, and required education, skills, and experience.
2. Designs Interviews
Once Wendy has identified the important information, she plans the dialogue flow of the interview. One of the biggest architectural challenges while designing this capability was to enable Wendy to execute a completely different dialogue based on a job description and each company’s needs. Here’s how she structures interviews:
a. Introduction and a brief description of the process and role
b. Company overview
c. Update candidate’s profile
d. Qualification questions
e. Logistics questions
f. Closing
Some of the items in this list can vary depending on the need of the company.
3. Conducts Interviews
Wendy works hard to say the right things and ask the right questions during interviews. She’s also responsible for the following:
- Noting relevant candidate information including past roles, skills, and companies.
- Responding to questions brought up during interviews.
4. Evaluates Candidates
After the interview process, Wendy evaluates each candidate based on their responses to the qualification questions and provides a numerical score that translates to one of two values: Qualified or Unqualified.
How Does Wendy Carry Out Her Tasks?
Wendy is powered by four microservices:
- NLP Service: Parses job descriptions and extracts relevant information
- Chat Generation Service: Generates the dialogue flow of interviews
- Chat API: Communicates with Kore.AI, our chatbot platform, to execute interviews
- Evaluation Service: Evaluates interview responses, calculates scores, and generates outcomes
Wendy executes these services sequentially for each incoming job and creates an evaluation outcome for each candidate. That process looks a bit like this:
Underlying these microservices is our graph database. Each microservice retrieves input values and saves output values from the graph. Requirements from job descriptions, chat dialogue flow, interview transcripts, enriched candidate profiles, and evaluation results are all stored within. As a candidate moves through the process, this graph database grows with each response.
But Where Is The AI?
So, what makes Wendy such an incredible and unique AI chatbot? Her ability to effectively leverage a wide array of powerful machine learning models and proprietary algorithms across all four tasks.
1. Job Descriptions
NLP service leverages most Machine Learning (ML)/NLP models and deploys the following:
- Skill Extraction Model: Identifies required skills specified in a job description
- Skill Recommendation Model: Recommends skills when a job description doesn’t contain enough information
- Industry Classification Model: Categorizes the role to identify relevant skills and titles
2. Interview Design
The Chat Generation Service generates a dialogue flow for each interview instance depending on the company, job description, and applicant by applying NLP Service output and key variable data (e.g., applicant/company name, job location, etc.) to a predefined dialogue template that’s automatically selected by a proprietary algorithm based on job title.
3. Conducting Interviews
Wendy is an expert at making candidate interactions more conversational and less like simple directed dialogue. She does this by leveraging machine learning models in the following ways:
a. Entity Extraction: Wendy leverages entity extraction models to efficiently aggregate key elements from candidates’ responses (e.g., titles, skills, experience, company names, URLs, etc.).
b. Intent Detection: Wendy uses an effective intent detection model to identify and respond to candidate questions that arise during a chat before moving forward.
4. Evaluating Candidates
Wendy capably evaluates candidates using a proprietary algorithm that’s based on the Bayesian Network. She calculates candidates’ scores using a joint probability function that assigns a conditional probability to each qualification, which she adjusts based on the qualification type (e.g., must-have skills, nice-to-have skills, etc.).
Conclusion
Wendy is a highly capable AI chatbot. She carefully executes each task with precision and leverages some of the most powerful and advanced proprietary algorithms and machine learning models in the market. By applying NLP outputs to predefined templates, she can conduct different interviews for different job descriptions and companies. She can even score candidates and generate outcomes for recruiters and talent professionals. These advanced capabilities help make Wendy a powerful force in the world of AI recruitment.