Applying NLP in FinServ, Practical Considerations - with Luca Scagliarini of Expert.ai (formerly Expert System)
This is daniel fidel. And you're listening to the financial services podcast. There are a lot of potential use cases for a financial services. Let me see if you can tell what these three habit common. The first is a conversational interface. The second is the summer ization of call center. Data calls from our customer support operations and the third is documents search and discovery. But they all have in. Common is reliance on natural language processing applying. Nlp can happen in all different dark corners of banking and financial services. And this interview we speak with lucas scaglia. Rini luca is with expert system expert system for those of you who are enterprise customers here with emerge. You're aware of expert system from our ai. In banking opportunity landscape anybody who's ads access to that data as part of their work with us here to merge will be aware of the various products of experts system one of the over one hundred companies covered in that research and a big part of what they do. Lucas speaks to us about. Serve the differences between machine learning and other. Ai approaches and some of the practical considerations for applying natural language processing inefficiential services context. When is it the right move. When is it the wrong move. And what are some of the things that companies often get wrong. They could do better to have more success with their. Nlp applications if you're interested in bringing nlp to life in your business be sure to download our unlocking the business value of nlp pdf guide. It's short pdf brief. You can download it e. m. e. r. j. dot com slash. Nlp one that's nlp like natural language processing and then just the number one yemi rj dot com slash p. One again that's unlocking the business value of nlp if you're interested in more use cases going deeper with actual application checkout that guide otherwise. Enjoy this interview with luca. Glad he could join us on this episode with expert system. you're on the a infinity services. Podcast so luca. I want to kick things off by talking to you about serve the as you framed it the war between symbolic and machine learning techniques for nlp. I know there's kind of a wrestling match year as to which is better which is worse which is going to win. You've got a little bit of a different take on this. Can you talk a little bit about how you see that kind of wrestling match on. What i think is that i think trying to fight for with is the perfect solution for any possible applications kind of waste of time. I think it's because we have been dealing with this now for years. It should be clear that one technique like machine learning is pretty good in doing a certain kinds of things under certain scenarios for sure the best way to apply to business use cases and then it's also obvious that You know that in certain circumstances especially when you don't have a lot of training You know sad or you don't want to spend the hundreds of thousands of dollars just to train a system that then you need to retrain to improve the performance in approach is more symbolic my symbolic i mean. An approach is more similar to the way. The human brain works in terms of You know understanding concepts relationship between concepts. I think these kind of techniques can work much better in certain circumstances so i think that what we would see in the future and referring to that in the future in general application to nlp to business but obviously being the financing of being one of the first kind of adapter. It's probably way we'd see these. I at a kind of a makes What we started referring and an a person is referring hybrid approach so the possibility to actually understand the case. Understand what are the what is the specific situation and then based on those contextual information really peak models are mainly based on machine learning mail based on on symbolic. I really think this is kind of Appearing and i think it's going to be probably even more common in in the future. Maybe we could even talk about instances where you know as you had said a big overall. Who's going to win. The whole game is kind of a silly way to think about it. There's going to be different approaches for different kinds of applications and based on performance and based on maybe energy use in cycles or whatever we're trying to optimize for there's going to be other ways you in different ways to skin the cat for you. Are there instances where you know maybe representative instances in financial services where we could talk about the symbolic really being what we wanna lean on and maybe not making as much sense to machine learning and maybe situations where machine learning tends to be what we wanna lean on for. Nfl as as we can kind of. Put those in buckets and talk about when each might have their place without oversimplifying things. But i think that there are pretty obvious situations in which you know machine learning requires even if obviously the training requirements are not as demanding as they wear in the past but still machine learning techniques need to learn from examples right so he situations where use cases present naturally for an enterprise a lot of existing samples so a lot of ways where you to start from. They are a perfect candidate for For machine learning. I'm referring for example. The chat area is an area where machine learning fines very media. It's obviously superior because you'll have the samples the variety then the variance of the questions tend to be always in eighty twenty. So with the you know the the twenty percent of questions to actually answer eighty percent of their parents. It fits pretty well. There are other situations in which these kind of reliability of A big set of data to train the system is not there and so you know having a system that is based the knowledge and existing knowledge is knowledge graph and that has an understanding of language these independent from the training so it's sexually based on On general sani of the language need to deal with finds a much more natural application. I'm using some examples. That are maybe not so. Common depends on vary from country to country but for example the area of managing automatically garnishments for example that requires a lot of going after reading in deep in depth documents extracting and for. You don't have samples that can really be representative all the cases and now the situation in field that is linked to the financial institution is for example claims management insurance or even areas like Contract contract is a contract comparison comparing a contract with a new version of the contract. You make have to make sure that all the exclusions are over represented correctly. Can the inclusion are represented correctly. All of these are situations that usually are. He does that from a major train. Sat. and so those situations where naturally symbolic a person to be. Better and let me add. One thing is that is not that The situation loves to change change over time. And that's the example of boats that have to deal suddenly after kobe with a lot of queries from customers about things that will never mentioned before is a good example. Also maybe pick the machine during this structure change and then suddenly you untapped to train. The system is so again i am. I'm here to say that it's really a question of not being is not like a religious war you know. It's not the is just. That is just a question of being pragmatic impractical. Yup and that will depend on the circumstance circumstance. Being the use case circumstance. Also being the time like you said maybe if you're leveraging machine learning system but the the world changes radically maybe that's no longer the best suited use case so hopefully some of these distinctions. You've called out are useful for people listening in now in wondering sort of. How do they want to approach things in hopefully for the folks listening in if you hear somebody really making this into you. Know as luca put a religious war as opposed to a practical consideration. You'll know that that's not exactly a strong footing to be standing on. If you're really trying to solve business problems speaking of business problems one facet that you and i chatted about before the interview here other wanted to get into was around the kind of acquisition and cleaning of data when people think about nlp. They think about often what it can do. Oh you know. I want to have a chat bot customers. Oh i want to search for documents or find certain clauses in in my legal contracts or whatever the case may be but they often aren't thinking quite as much about what it takes to acquire and clean data so that. Nlp can actually do. Its job a wanted to know maybe for you if you could kind of t up for the audience really. What if some of the big concerns there as business leaders who might not be writing the code. What do we need to understand. And think about when we're thinking about using nlp in this respect for position cleaning. Yeah i'm i'm making some some simple examples because obviously this can become pretty technical brady fast. For example i can see the Like the need to. Let's say process incoming documents in this in different formats okay so it is obvious that you can have a significant noise in input if this format is not ideal you know like for example or cr are working extremely well but it's enough that the document needs to be yard. It's the facts for example that's that's-that's make a typical example. You lose a lot of quality important so even if it seems in our structure document it seems something should be handled pretty easily than the turning that into a digital version. You include so much. The noise than the downstream part of the process is pretty critical and other examples are documents that are for example say structured right so let's say documents include tables and all of these are areas that for a business person the app but it's a document the document and so for me the fact that that's three tables or the fact that taza just paragraph it's exactly the same because he includes unstructured information and it's correct but the difference in doing a great job to turn the table into data can be used makes all the difference in the final solutions i i i spoke sometimes with customers that give me like you know handwritten documents even if now there are dr are software that turn a handwritten documents into digital aversion steel data a lot of noise that this kind of middle of the road the middleware software actually creating the price so i think that downplaying the need of having a a very rich way and very once again. It's not the one a one-size-fits-all like having the right set of tools that enable you to do the the cleaning it's perceived as being extremely important in that in that issues were data numbers. But there's not the same perception or at least the same understanding in In nlp so again. i make. I made very simple examples. I'm sure that people that are more technical. They conceive them. Maybe not not very intelligent examples but the reality is that these are examples that everybody can understand and and the way where you have issues in terms of not having a clean. Acquisition of data can be represented by those examples. Cool and when. I think about The consideration bring up your on acquisition cleaning. I sometimes think about how much transformation has to happen to the business in order to bring ai to life a lot of the time if we're stepping into a universe where the data is really gobbledygook. And it's not not exactly easy to set up in integrate. A i in clean things well enough so we can have some outputs that are useful. Sometimes we just need a system that can drink in that ugly data and clean it reliably and then use that to train a model other times we need to talk to a client potentially around kind of changing how they date infrastructure in the first one is changing how they do intake in the first place. Is there a bit of a wrestling match. You know when we stepped into clients and arianna workflow between. Hey what can we just vacuum clean and tweak up in the back end in our of our own system versus what we really need you to systematically kinda fix for the sake of your data health long-term mr client so that we can actually make this happen. Is there often a bit of both luca. I mean it's a very very good point and So i think again. It's a question of if you want you. Can i think there's this perception that is you know magical ride that they saw gold possible issues because those warling artificial intelligence building the definition reality all these aspects. That are extremely important. So and i'm bringing it when even furthest step is not only about The way you actually structure sometimes need to have work to be done before you actually can implement official intelligence but also the way you have designed your processes you cannot do a autonation which includes usually a piece of natural anger understanding if you don't have processes that can be automated so the same way it applies data so i think we need to need to be honest. I mean the point. Is that the successor away. I is really around turning things being practical. I should be perceived as an investment in any kind of other software. It's not something that is you know kind of In its own completely different category and when you face reality and it we practical pragmatic chooses the solution that might be eighty percent today and then do the pre work to make sure that you can turn to under percents in the course of the next month. I think this is a important for people that are enterprises. The dr initiate investing in initiating journey inside the world over deficient. But he's also something that vendors need to be much more kind of You know practical in the communication the custom. Sometimes things cannot be done. And the risk of putting together in something that is unsuccessful than it impacts the overall perception of the enterprise of this new technology that can bring value today if you follow very practical and pragmatic approach. Speaking to that point. luca. I know that the ai vendors that we've spoken with over the years who have mature companies really have had to learn a lot about what it takes to work with clients and have a bit of a white glove approach to teaching folks may be what would it actually takes to bring ai into an organization. Sometimes it's easy. Sometimes it's plug and play most of the time it's not when it comes to serve encouraging clients to invest in their own data infrastructure the cleanliness and access whether it be real time access or just access in general to their data. Sometimes that feels like this. Extra hurdle is extra cost to leveraging an ai system but in fact there's another side of that where it's also an extra bit of maturity that we can kind of invest in so we can use that data more productively for thousand applications in the future. How you speak with clients about being able to and willing to invest in maturity maybe the kinds of maturity wouldn't need for traditional it system. But something that may be a i would need. What does it look like to kind of explain that in a way that executives understand it can make a rational decision about Find the usually a kind of productivity to consider that the arrow i off An investment in deficient intelligence sort of compounded effect. It means that the the kind of investment. You're doing maybe to do the first implementation that could be the two example to me right to create an infrastructure for data that can be acquired in a much more effective way or could be that to designing a process that is actually designed with mission in mind both those investments and it could be also creating a sort of you know kind of general knowledge across the enterprise. Zones is a knowledge. Graph three describes the language of the there. What needs to understand. Is that the return on that. Investment is not the only in the first project that they actually put in place. Their return investment is as a multiplication fact. When you actually turn these seen to cover during different areas in your organization and that's where you know having approach that he's around creating if you on actively intelligence as sort of infrastructure for the company where today you implement your shot tomorrow implement your contract comparison. The next day you implement something to automate europe wants to create the infrastructure than you actually can compound the effect of the return on investment. So when you actually face the discussion in terms it's something that i think Executive reading understand very well. And then that can remake the difference. Compared to do kind of you know kind of set them forget kind of strategy. Well i think this is a really important way to think about things for the listeners. Who are tuned in. We have an article called critical capabilities. You go to google you. Type in emerged critical capabilities. Get an understanding of some of these sort of prerequisites day i deployment data infrastructure being one of many. These are things that luca as you're kind of articulating. Now we're talking about sometimes. They're seen as hurdles right. A company will think to themselves that they're not familiar with ai. Will jeez this is a lot of work. I mean i just wanna make this chat bot happen but you know what you and i are talking about the fact that he s. Maybe these are additional time and money. But also this. This is how we unlock future capabilities. I'd love your thoughts on this as someone you guys are. Pretty mature vendor company had been around for decades very few. Ai companies have been around as long as you guys have in terms of seeing that maturity investment. Be something that clicks with executives. I wonder what that really implies. We could talk about conceptually. Hey you can upgrade your your data now and then in the future you can also use chat bots can also do this. There's a bit of a wrestling match here. Because it feels like number one we could just talk about like we could talk about all the different things we could do with the data and then in all those different circumstances. What's the foundation of that infrastructure. That what would need to look like to enable all the cool things we wanna do in the future. That's one thing a second thing would be. We have a single project and we just upgrade the infrastructure for that single project. It feels like the second one might be less abrasive to the buyer because oh it's less of an investment it's less thinking in strategy after bringing less subject matter experts but it's also unlocking less of that grand future potential weird. He kind of draw the line between really opening up the conversation about what we want to create mr buyer for everything we can unlock together. And you can unlock in the future versus. Hey what do we need to unlock to get you the darn documents search program that you hired us to put in place in the first place. How do you factor in both of those conversations because it really feels like they both have merits. I mean obviously. They both have married. I think one one aspect to think about. It's also how things are changing with time. Okay so obviously. There was much more resistance or at least was perceived much more scifi dot com Four five years ago right so The objective was to find these kind of fast win. That may be. It was creating the end of the day. What gartner caused the technical a gap right. So you find. You do the minimum required to do that solution. makes sense economically. You are out of that solution but then if you have to either one you need to redo things not from scratch but palmos. I think this was sort of tactical way. F- years ago. That was probably the only way to actually go into into the market. I think that there's enough awareness right now to understand that. Implementing is radius structure change for the organization. Where this discussion around. We are not here to boil the ocean but if you just focus on the limited things enable you to put in place one one solution. You're actually missing an actually. You're not creating the system that enabled you to leverage more. I think there's more awareness. And i think that i'm talking with our experienced for sure. Other enders have seen in our experience. Is that now. You can actually count on showing things that happened in other organizations around using approach that you know we have we. I'm sure i saw the vendor have several companies that have for whatever reason a combination of More of innovative approach the combination of the right people right moment have chosen to avoid the easy win or just focusing on these wing Stop this kind of infrastructure. And then they start getting the actual return so now few years went by and you can actually talk Nine ten twelve project in production. That are all around and be for example blase seeming images so so you can actually see that you have this proof if you wanted something that is changed in the last few years yet is. I'm trying to come up with a nutshell insight. Luca that we can end the interview with year. Maybe one way to think about this. Is that on. The one end of the spectrum we do the absolute bare minimum for any kind of integration. In order to just sort of plug in the solution is the downside here is that we don't build maturity in terms of teams skills culture importantly data infrastructure. But we kind of get it up. Maybe we can even get some results with it. The other end of the spectrum is with a goal in mind. We just think of one hundred things we could do with a and we think about the foundation we could build on that would really be the most fluid set of teams in skills and culture date infrastructure. That's go unlock all the future. Ai stuff that we wanna work on that second. One is pretty unrealistic. Nobody's going to invest purely in hypothetical in foundations. The the first one is a bit immature. I think most of our listeners certainly are aware that a with no focus on on the maturity is a pretty bad way to go it. Sounds like all folks that are procuring. These technologies luca thinking about them in the future need to be asking the question. How much maturity do we need to realistically think about or would it be responsible to realistically think about uh before we think about deploying this tax maybe we should have that conversation have that strategic talk about how deep we wanna stick the tendrils just to get quote unquote this one application up and running. It is that something you wish. More procurement folks would pause and think about. Or maybe there another way you would like to to nutshell this idea. I think i think it's It's a framework. What is what is. Also i think helping this discussion right now is that the discussion is not around. Only like one point solution fairly muted and something that is undefined. Because i can do hundreds of things but right now. I don't know exactly. What are these under the things i think that right now there is enough maturity in visibility on also amer. Jay has great content around practical implementation. Right so let's say. The approach new financial enterprise. Now is not that you will there with one point. Sushi can go there with already in the moment when you ask. For this kind of alignment you can say eight. This is one but you can already see in your eyes on without any major kind of are not talking about things not define Duke three four two three four possible implementation that you should actually account for if you do this initial step. So that's i think. What is changing the more visibility on. It's not really pure low hanging fruit but a lot visibility on practical implementation that can point you in the moment when you're thinking about the first solution saying hey. I already see in a reasonable timeframe. If i do this. Investment to completely leverage it also on investment to on use case to use case raiders. Because for and i think this is something different from what it was just a few years ago when you know this scenario is not not likely. Got it again and again. I think you're right. I think there's there's been a sea change in terms of thinking. Part of that is around what we refer to as executive a fluency that is to say does the leadership even understand that. Ai maturity is i think the fact of the matter is for many many years. That answer was a square and firm. No pretty much across the board and and so nobody wanted any maturity investments because it was all extra gobbledygook that the regular. It vendors didn't ask for but but now we've got hopefully more mature folks in part because of smart vendors like yourself we're educating people and hopefully in part because of folks tuning into our show reading materials on deployment. Ai are y. maybe luca. We're lucky we did our part here today. So i I appreciate you being able to join us and enlighten us. A little bit as to the practical realities of nlp and glad to have you on the program again. Look thank you. Thank you very much and you know whoever's listening and they want to reach out to me for any genetic question we don't have to speak only about about this but just unethically about this kind of adoption i think. Can you know. I think we can have a good exchange. Our way slopped. Dr people who actually are facing practically so thank you very much. That's all for this episode of the financial services. Podcast if you like what you're hearing here be sure to leave us. A five star review on itunes because this is our newer and are smaller. Podcast every review really matters a lot plus we take all of our review quotes and look over them every single monday. He would emerge so when it comes to determining our editorial calendar the first point that we check in on is what folks have to say about our podcast and any lincoln comments get our actual written article so your comments really do matter in crafting a great show. We appreciate your listening. And we'd appreciate your thoughts again. Drop us five star review on apple podcast. Easy to find the in financial services podcast. Otherwise we'll catch you next month for next episode. Thanks much for being here.