During this webinar, industry experts discussed NLP in the Enterprise and more specifically, what the requirements are for an organization that wants to start doing NLP. We’ve included a short transcription of the webinar, beginning at 3:45 of the webinar.

 

J.T. Wolohan, Booz Allen Hamilton: So Jennifer, let's start with you and you can kick us off with an answer to what are the requirements for an organization that wants to start doing natural language processing and at Boeing, where have you guys started focusing on natural language processing?

Jennifer Klemisch, Boeing: There are a couple ways and it always takes a little bit to figure out the best process but starting with the team, bring in your instrumentation engineers, bring in the data scientist. Also bring in the subject matter experts. The ones that help curate the data and that are going to end up working with the data in the end and working with your models and honing your models. So starting with that core group.

Also having a customer representative and especially the sponsor. It is so easy to have the technical community get together and not have the sponsorship engaged from the beginning. So starting with your team and really identifying what do you want to do for natural language processing or any model for that matter? What questions are you looking to answer? Can you actually write out the question and get it back there? From there, consider if there are any applications that will do what you want to do. If you're looking to do analysis of charts or graphs, go get software that will do that because it will do it better than you do in house. 

If you still need to do an in house model, a lot of what we do is mixing engineering with natural language processing and there's not a lot out there with that physics based modeling to begin with. So from there, we actually start with the data science high-paid needs and identify gaps in the structure. So we have the data, is it being stored? Is it accessible? A lot of times with our data, at least, they're in different systems that were developed 24 years ago, so accessing the data will be challenging. So we start with finding out the basic levels or the foundation of the hierarchy of needs and identify those gaps.

We start with MBPs. We do a number of hackathons internally so we can get to a mobilized product and then we improve the architecture as the questions get more challenging so make that a never-ending process. Then as we go with each problem, we really try to identify what the object is. So it's easy to get lost and try to do too much at the beginning so taking one step at a time with each problem.

J.T. Wolohan, Booz Allen Hamilton: Johann, you've been in AI for a long time, I imagine that you've seen a lot of organizations try to start the process of getting into NLP. What are some of the success stories you've seen there? And the organizations that do start up NLP well, what have they done? 

Johann Beukes, Levatas: I would say one of the first things that we always try to do is do a data audit. We want to see what data is available; but not just available, we want to see what’s consumable.. Is it consumable? What's the quality?

Here, I'll give you two examples. I like to kind of give you one success and one really bad one, just so you can kind of contrast. We did a project for a company that had a recommendation engine that they didn’t want to change out. They wanted to add some natural language interface like a chatbot to make recommendations and, in a soft way, somebody to make the purchase with upsell, cross-sell of products. The problem we ran into is that NLP was not the problem. It was the recommendation engine that ended up really making the product not very useful. 

Switching to a good use case, another company in the same kind of situation where that the recommendation engine and, back to what Jennifer was saying about the just a data, this company, even though the project overall was successful, we started off with no data. All the data was basically stored in menus. So if you think about that you have to extract it, digitize it, tag it. But ultimately, we had a back-end system that made really good recommendations. We tied it with the third party data source for food specific ingredients and we made it pretty rich where most people's experience of chatbots aren't really great. This was a very narrow focus, very specific type of application. So to boil it down, I think if an organization wants to succeed the first thing they need to expand their data to understand how they are either going to utilize it or if they are not at that point, how they are going to capture it to utilize later. 

J.T. Wolohan, Booz Allen Hamilton: Yeah, it's great. Wes, do you have similar experiences?

Wes Barlow, USAA: I do, I guess one of the one of the big things I probably press on are some of the comments that Jennifer made around understanding the business objectives and the requirements around it, because that's a data scientist. I think in general we're very curious and we get excited about new challenges and new things and you can get really stuck chasing something down that really isn't going to meet the real goal. Especially using some of these more exciting and new methods that are coming out. It’s understanding what the business really needs in the beginning and whether or not NLP is the right avenue to go to meet that objective because it is new and neat and we do enjoy doing it. But we've definitely had instances where as you really refine what their requirements are, you find out there's a much better solution that is not what the current whatever buzzword has been pitched to the executives. 

J.T. Wolohan, Booz Allen Hamilton: Yeah. Yeah, I think there’s a really good trend brewing across what the three of you have said starting with the business problem and then working back into a data science problem. So you don't necessarily want to start with NLP as the answer unless the question suggests that. If you've identified that NLP is the answer, there are requirements that you need to have, like textual data to work with.

 

Learn more and watch the full video on YouTube: https://youtu.be/CNH_yphj0P8.