What higher education leaders need to know about conversational AI
Artificial intelligence. Machine learning. Chatbots. As a leader in higher education, it’s likely you’ve heard one or more of these buzzwords in recent months. But what do they really mean – and why do they matter to today’s universities and colleges? Perhaps most importantly, what role can they play in helping you to improve student outcomes?
Here at AdmitHub, artificial intelligence (AI) is our area of expertise. Powerful AI is at the core of our conservational platform for higher education staff and students. To get started, let’s define a few of the most common terms in the AI landscape:
- Artificial Intelligence (AI): A computer program that automatically learns, improves, and performs tasks in a way that resembles what we consider human intelligence.
- Machine Learning (ML): The process that enables systems to automatically improve from experience and data patterns without being explicitly programmed.
- Natural Language Processing (NLP): The ability of a computer to decipher human conversation (or “natural language”). By using contextual clues, NLP can help machines make sense of what humans are trying to say.
AdmitHub’s conversational messaging platform is built with a specially trained AI that helps separate tasks into: (1) repetitive FAQs, and (2) complex, unique, or sensitive topics that deserve individual attention or require human interaction. These AI-powered chatbots, combined with our innovative conversational platform and our strategic communication services, help our partner institutions to significantly increase student engagement, elevate administrative productivity, and improve student outcomes.
When a student asks one of our chatbots a question via text, web, or Facebook messenger, a complex process ensues to analyze the question and provide an accurate, immediate response – all within a matter of seconds. Here’s how our AI handles any inquiry it receives.
Step 1: The chatbot uses AI to determine the meaning of the question
Before the chatbot can provide a student with an accurate answer, the AI must first figure out what their question is actually asking.
To start this process, our technology simplifies the question by cleaning and standardizing each word. Contractions like “I’m” and “would’ve” are expanded to “I am” and “would have”, respectively. Abbreviations are also expanded, so “LOL” becomes “laugh out loud” and “TBD” becomes “to be determined”. This allows the AI to match each word to existing words to start to understand context and meaning. This also helps enhance the Natural Language Processing to understand poor grammar, missing punctuation, and slang. Sentences that look different character by character but have the same meaning now look more similar to each other.
Once the question has been simplified, the bot’s AI then starts analyzing the words in the question to extract key information and map out precisely what the question is asking. This may seem trivial to a person, but to a machine, it is difficult to decipher the difference between similarly-worded questions.
For example, “What is the address of the library?” and “What are the directions to the library?” are worded similarly: They both start with “what” and both end “the library.” In between, they have the same grammatical elements: a verb (to be), a definite article, a noun, and a preposition. Yet their expected answers are different. In contrast, “What are the directions to the library?” and “Can you tell me how to get to the library?” are worded quite differently, but the expected answer is the same.
To help determine the actual meaning of a question, our AI model turns each combination of words into a high-dimensional mathematical representation of the question called a vector. This conversion can be done in many ways – some simple, some complex – and how well the chatbot works depends a lot on how the conversion is done. The key is that questions with similar meaning, but possibly very different surface forms, are converted to vectors that are geometrically near one another. Our AI team uses deep neural networks, among other things, to do this.
Step 2: The AI compares the new question to previous student questions to find the best response
Once the AI has a vector representation of the question, the AI compares it to all other existing vectors in the vector space, which maps out all questions in our knowledge base. The AI calculates the distance between representations, which measures the similarity of the questions. The closer two questions are, the more similar they usually are.
If this distance is within a certain threshold (determined and fine-tuned by AdmitHub’s expert team of engineers), the chatbot can return the answer associated with the matched question. Basically, if the question – or a question similar enough to it – has been correctly answered before, the bot can provide the same answer.
Groups of questions that have the same meaning and match to the same answer are called topics or understandings. Usually, these will be close to each other in the vector space (having shorter distances between them), although the AI can be trained to recognize that some very close questions are not part of the same understanding – and that some more distant questions do belong in the same understanding.
While the vector space is a wildly complex 400-dimensional world, here is a 2-dimensional version that might help illustrate the concept. Questions within the same topic are close to one another because they have the same or similar meanings.
The more data the AI has to learn from, the denser the vector space becomes. This means that as the chatbot learns, it becomes increasingly more likely that a new question will be similar to an existing question, so it can be matched successfully and answered accurately. This also teaches the AI about English language syntax and how questions can be asked in several different ways with the same meaning, looking for the same answer.
At AdmitHub, we have a very dense vector space with around 100,000 different vector representations (questions). As a result, the vast majority of student questions coming into our bots have previously been answered or are highly similar to questions previously answered.
Here’s another way to think of it: Imagine you’re outside on a partly cloudy night, staring up at the sky. Each star is a question from a student, and each constellation is a topic or understanding (a group of questions with similar meanings and the same answer). If clouds are obscuring most of the sky, you may be able to see individual stars, but it would be difficult to determine if they were part of or near to a specific constellation. This is a vector space with very few known question-answer combinations.
Now imagine it’s a perfectly clear night and you can see all the stars. It’s much easier to see which stars are part of each constellation. This is a vector space with a lot of question-answer combinations. The AI training process “clears” the sky, adding more and more stars to the AI’s view. If you were asked to identify which constellation a new star (or a new question from a student) was part of (or closest to), it would be much easier on a clear night with tons of stars than on a cloudy night when you can’t see many.
These stars in the sky are the questions in our existing knowledge base, which becomes more dense over time and allows the AI to more easily identify similarities as it learns.
Step 3: The chatbot provides an answer to the student’s question
To provide an answer to a student’s question, the chatbot must be able to assess how well it understood and analyzed the question. We measure accuracy based on how similar a new question is to other existing questions. This goes back to the threshold we’ve determined around distances in the vector space (the shorter the distance between representations, the more similar the questions are and the more confident the bot is that its answer will be accurate).
If an AdmitHub chatbot is confident that a matched response will correctly answer the question it’s been given, the response is immediately sent to the student.
If the chatbot is uncertain how to respond, it will send something along the lines of “I’m not sure I have the best answer for you, type #followup to forward your question to a human” which lets the student know they’ve been heard and they can reach someone if needed.
This part of the review process includes escalation to university staff. Additionally, if the question is flagged as a sensitive topic or meets the criteria for configured escalations, it is escalated to the predetermined point of contact at the university. That person is able to send a response directly back to the student in the same conversation thread.
In summary, our AI allows the chatbot to take on the task of providing confident answers to the questions it understands, while humans can intervene in unique, complex, and novel situations.
What does this mean for you and your students?
Because AdmitHub’s AI-powered conversational messaging can effectively understand, compare, and answer student questions, our platform allows you to transform communication with students.
Our AI is always on, so students can receive answers to their specific questions at any time of day or night, without your staff needing to work 24/7. Students can also count on getting correct answers to a wide range of questions, even if they use slang or emojis to express their queries. And, if complex or sensitive topics come up, your staff can rest assured that those questions will be escalated for a human response.
At the same time, your staff gains reach and bandwidth, enabling them to support a greater number of students and helping them to spend more time building 1:1 relationships with those students. Your team can easily monitor the conversations students are having with the chatbot, giving them invaluable insights into campus trends that they can use to further elevate and enhance their daily work. Plus, the chatbot becomes more effective over time – especially since your team can further train and refine its knowledge base – allowing it to grow with you as your campus’s needs change.
The result? Happier, more engaged students who are more likely to enroll and persist, and happier, more effective staff who can reach and support more students than ever before.
What’s next for AI at AdmitHub?
We are always working diligently to improve our technology with the highest standard. Here’s what we’re focused on next:
- Leveraging new state-of-the-art NLP research, including some very effective models developed at Google
- Improving our knowledge management tools to help schools specialize their chatbot for their institution
- Adding to our network of higher education partners, so we can share the insights we’ve gained from our AI – and continually learn and refine our AI based on their experiences
Download our free guide to learn more about the benefits of introducing a chatbot to your campus.