<img src="//bat.bing.com/action/0?ti=5739614&amp;Ver=2" height="0" width="0" style="display:none; visibility: hidden;">

When developing an AI personal benefits assistant, you need software engineers, machine learning specialists and... a linguist? 

Lexi_Blog

You bet! While luckily, language is second nature to many of us, machines don't know or understand language unless we teach them. And that's no small feat. I sat down with our Computational Linguist, Lexi to understand her role in developing Sofia and got some very interesting insights into the complexities of language as a whole. **The below copy has been edited for length and clarity. 

Tell us a little bit about your role and background. 

I combine coding with linguistic knowledge to help Sofia understand how language works. I have a background in psychology and linguistics. I really wanted to look for ways to apply linguistics to a project. So I got my masters degree in computational linguists. I strengthened my coding knowledge which basically brought together my two passions, computing and linguistics. I wanted to work on projects that thought about and put into practice meaningful language processing techniques to help machines interpret language. 

What kind of projects are you working on with Sofia?

One of my core projects is maintaining Sofia’s training data. There is a saying in linguistics that language is infinite. Basically meaning that there are infinite ways to ask a question or say a sentence. I work on building and understanding certain "utterances" or questions that we want Sofia to understand. Once we build her database of utterances she is able to understand the intent behind the question. The more utterances she knows, the better she learns the intents and can give the right answers. Here are some examples of a similar structural utterance with different intents. 

“I want to add my son as my dependent." - The intent is that someone wants to add a dependent onto an insurance plan. If we feed Sofia many similar utterances like "I want to add my daughter as my dependent" or "I want to add my husband to my health plan" she can learn the intent behind the process they want to start. 

“I don’t want to add my son as a dependent.” - This is where linguistics can really help Sofia learn the difference between these two utterances and the intents. Although this sentence is very similar to the one above, it's actually asking the opposite thing or intent. It's difficult to teach a machine the difference between these two sentences since they are very similar in terms of content. However, if you teach Sofia the negation in the sentence "don't" that can help her understand how she can respond and the intent behind this statement. 

Our world of benefits and our lingo is complex and complicated. And it’s very specific. Helping someone use their natural language to communicate with Sofia and get the right answers when they need them is an important aspect of my role. 

What has been your favorite project so far? 

I really enjoy going deeper into the intent of an utterance. What is the topic this person is asking about? New topics and ways to ask questions come up all the time. By understanding what someone is actually saying and teaching Sofia that intent with syntactic parsing, helps her understand language better. At the end of the day, a sentence is a list of words. But they are structured in a specific way and that matters to the understanding of that sentence. For example: 

  • Time flies like an arrow.
  • Fruit flies like a banana.

Although these two sentences look very similar, in the first sentence, "flies" is the verb, or the action, but in the second sentence "like" is the verb. Dependency parsing helps us get more out of what’s in those sentences. Additionally, we take user data from previous months and take a look at what people have asked and how they've asked it. By using real data, Sofia is always getting smarter and learning new ways to say things and answer questions. 

How do you see HR and AI evolving?

When you’re building artificial intelligence you are giving people a non-biased product to help people navigate a complex situation, like benefits. It's important to keep non-biased techniques in mind when we develop Sofia. We like to call it, technology with heart. AI can help HR continue to reach and empower employees with the right information, at the right time with diverse experiences and backgrounds in mind.

You can listen to the full interview here: 

 

Check out our other interview with our Machine Learning Research Scientist

To learn more about how we develop technology with heart read our resource below. 

New call-to-action

View all Posts by Bridget Mortland