Intelligent Support Case Routing using Google NLP API and TensorFlow (Cloud Next '18)
Hello. Everybody thank, you for being here. It's a pleasure. Priya. And I are really excited to talk smart case routing systems with you today. It's. A combination, of some of the the better techniques. That we've encountered, the better approaches, the better the best tools and also lessons, we've learned from. Building case routing systems and natural language processing through, the years at Springer Mel, let. Me look at the agenda here we're, gonna start by introducing ourselves spring, ml in case routing briefly, then, we're gonna tackle, the the meat of the matter the, three main models, that that that, build this smart, case routing system and finally. We'll open up for questions and. And. We'll also be around afterwards if you have additional questions so. Let's. Pretend ourselves do want to start for you yes sure hi everyone I'm Priya dyd I'm, a data scientist, at Spring ml I've been a spring ml for about six months prior. To joining spring ml I work for Capital One for about ten years I'm. Very passionate about deep learning and I've been working. Through it for the last three years I also, write blogs very frequently, and you can find me on medium. Thank. You, I'm -, manual immunity guy and I'm also a data scientist I'm the VP of data science at Spring ml and I've, been at spring about practically, since the beginning I think a few months shy since the company was created about three three. Years ago and. I. Also, AM so, prior to that I was a data, scientist in healthcare for four years and prior to that a quantitative developer, on Wall Street for six years and I'm very passionate. And interested about applied data science I'd like to see how we we, take these all, these models that are coming our way every, day and see how we can apply them to make our lives to everyday life easier. Right so of course curing, cancer would be great right but it's it's not an easy thing, to tackle but like. Cased routing these kind of things we have the tools today to do it and to do it well so that's very exciting. So. Briefly, about, spring ml we're a company that brings. AI and machine learning to, companies. Who either don't have the staff the department, and know-how or need. Extra hands, and we will kind of take. A step back with them look at what, they're trying to do look at their data if it can be achieved and then dive. In with them as a team so, that we, work in tandem and that they have a sense of ownership and, then they can take it on they can extend it and they can sell it internally if need be and we're also happy to form longer term relationships. I, mentioned. That we build quite a few case routing systems and that's true over. The years we've built them and it was what's interesting about a case. Writing and natural, language processing in general is that his has evolved, enormously what, we're gonna present today is very different than what we started actually there's probably a very little that intersects, and. We. We've, done a lot of case, routing demos internally, we've done them for Google and we've, also a partnered, with Google for other customers, the biggest customer we worked, with the CBRE, it's, a very large commercial.
Real Estate company they're, all around the world and you, can imagine the flurries of case, of cases going back and forth issues questions. Comments. So you could have on one end you know please change in this country change the light bulb here there you can have always a great, investment. Opportunity there, or you, know here, there's a building that's flooded you know we need the basement is flooded we need you know to call the maintenance crew so all these things go back and four sisters in a very large company so automating. Case routing for companies at that size but even smaller companies is a huge win that's what we're gonna talk about today. So. That, we're all on the same page what, is case routing so case routing is whenever you have a department, a team, or even. A single person that that. Is in charge of taking, you know either external, queries or questions or internal, ones and routing, them case routing right routing them to where they need to go in some cases they can even answer them right and. I, don't know show. Of hands who's been on call with their job. Yeah. I wasn't caught when I was in healthcare I wasn't a call and it was, really tough job very stressful, because I remember like health, care never sleeps machines. Go down on, the weekend or at night and I. Wouldn't, you, know I wouldn't know who to turn to where, to get help but people I would normally ask help they were either sleeping or playing golf right so it was it was very stressful and and, that's why I having a personal agenda to, work, on this to fix this this issue and why. Would you want to made this well you know the obvious reasons what you're removing, that middle layer like the people on call like me who don't want to be on call you're removing that and the customer, or the client - what would, the employee who has a question they know that's gonna go to the right Department immediately. You. Get 24/7, support you can handle multiple languages right we model this in. English, but you can very well model in any language you want so once you have one language down it's just a question, of getting enough historical, data to model other languages, you, have you, can do auto reply right very simple order replies like somebody, saying when, is when, is the office open and you say oh this country has a national holiday try, this you, know this state what is that easy wins you could automate that and, more important I think is the centralized digital, trail you're, gonna start getting a. Warehouse. Of all your your queries all your questions and so. Who who's asking what type of question, where are their problems, what department tackles.
Those Answers which, one are struggling to answer those questions all that is gold for a company right you can really learn a lot about your company about what's going on with your cost and learning about your customers, as well, so. These are there's three main, models, that we're going to talk about but, basically the brains behind what, we're going to present, there. Will be a convolutional, neural network that's gonna do the. The. Predicting, of what department, to route a case to there, is the. Language at the Google language API that we're going to leverage as much information as we can from from there to get additional. Information and finally. Finding. Similar, cases so you got a case that came in you want to know what is a similar case and how is it resolved, that's great hints for somebody who's handling that the caseload and, as. A high-level architecture before, we dive in into CNN's this. Is what we this is what we build in the past this is what we would build today - you, have the incoming case that comes in and, you. Have the App Engine so the App Engine is a service, instance in the cloud it's a great - a great type of, instance. Compute engine for this type of project especially if you have cyclical, caseloads, right so, if you have more cases, during the year these things will size, automatically, well what will balance will will scale and you, pay only for what you use it's, also kind of the the choreographer, the the, manager of the data you will do some data cleaning. Asset to the the, cloud Emily the the mounted machine learning engine in the cloud which, which. Is hosting our convolutional, neural network and Priya. Will talk more about the data set but, it's about financial. Consumer complaints, right we're using an open-source data set and it, can route, it into one of seven departments consumer, loan vehicle, loan bank at Cameron student loan mortgage credit card or payday loan so these departments are all you know related, and all within the financial industry and so, the. The. CNN what will tell you what Department to route it to so now we, know where to route it the. App engine will then also pass the information to the language API where we're going to get all sorts of information we can like sentiment we'll know if it's a the tone of the case things. Like the. Organization involves. And now. We can populate that information, in our in our in our case reports and we're. Also going to query bigquery, with a huge bottomless. Data. Warehouse where I was reading it can you know it could query like 35 billion rows and seconds with no indexes, so we're very powerful, tool perfect, for you know storing. All of your knowledge. Base all your cases and here, we're going to pull similar cases and the, end product for the demo is this. This. Smart report we call it right so it, will tell, you what the complaint is give you some some, hints about what's going on there and some suggestions. Of how to resolve it so. This is what the demo right most of the cases we would feed this directly into the customer's CRM right they probably have a tool already so either a rest point or an API right but for the sake of the demo we wanted something that was that that was visual, and I'm, going to pass it to Priya, talk about the cnn's thanks. Manuel so, I'm gonna talk about how we build a convolutional. Neural network, in. Tensorflow to do the case routing, so before I start by. Show of hands can I know how many of you have, built any kind of model in tensorflow before. Okay. That's a lot that's really good to know I've. Kept the talk here high. Level but if you have any specific questions, feel, free to stop me after, the talk. So. The first thing I want to talk, about is the data set that we used to to build this model, consumer. Financial protection, bureau in the US has a publicly, available data set of two. 74,000. Consumer, complaints, across many different products like bought, cash credit, cards student loans etc, when. A consumer complaints, to CFPB, they, have the option of making their complaint public and when, they decide to do that any sensitive, information like their, names or, amounts, or dates are naanum eyes and this data. Is publicly available we, think this is a really good data set for deep learning because, it's pretty big there's, a huge variety here, it, uses natural language, and. What I mean by that is if, you actually read through the complaints you'll see that some companies are very long some, are shard their spelling mistakes some consumers, repeat themselves so it's it's.
Perfect To use that for deep learning I also. Feel that the. Database is very rich in financial. Terms so it's, possible to actually use this data. Set for other use cases like. Routing customer, service calls. So. Now now, that we have the data set you can think of this data set as you know individual, complaints, and complaints, can be long or short so I've taken a sample complaint, here that, says you decline my transaction, when I was shopping and this is all words but. Neural networks love numbers, and maths so we do pre processing that. Converts, these sentences. Into numbers, the. First thing we do is we break. The, sentence into individual, words and so that would be you. Decline, my transaction. Etc, and. Then we map. These words into integers so we could say that you corresponds, to 23, and decline corresponds, to 45, and so on and this can be arbitrary. You know you can choose any of them you just need to save this mapping for later so if you want to convert the integers back towards you can do that. So. Now we have you know our complaints. Represented. As integers. So. When we feed them into the neural network we. Have to, assign. A tensor, to hold them and it's, tensorflow. Likes the tensors to be fixed size so you can specify them them in advance so this means that it, would like all, complaints, to be of the same length but in your data set different. Complaints can be different lengths so to solve this we, we. Choose a length and then we so everything is it's that line so if it's longer than that we truncate it if it's shorter, than that then we pad it so let's, say in this hypothetical example, we. Want everything. To be twelve words long so this is four short so we add four zeros at the end which, correspond, to a pad token in. The model, but. Just for reference you know the actual actual, data set was much longer so the, complaints, were were. Padded to 600, words ok, so awesome and so as you. Know after, this point we have you know our data set is all complaints, of same then all of them represented, by integers. But. If you think about it all these integers are unique like so we have all words, represented, uniquely but that's not really how our languages, words are not unique ones, are related to each other some words are synonyms some, are antonyms some.
Words Are very likely to be in similar contexts, for. Example a king and queen are very likely to be talked about in the same context, much, more so than king and aii right, so. To get over that or to give the model our understanding. Of English. Language we, convert. Each word into a vector and this. Can be you know a vector of any size we ended up using 128. Dimensional, vector and. So what, happens is that words, that are similar, or are likely, to occur in similar contexts have, word vectors that are closer to each other in the vector space and. So this allows us to give the. Model understanding. Of semantics. So. To train these word vectors we can use two techniques we can train. These word vectors as part, of training the neural networks, so we initialize. Them randomly and then, we back use back propagation to, train them as we, you know build, the model and that's what we did in this case and, this allows the vectors to have, understanding. Of financial terms the, other way is that you can actually use pre trained word vectors and these are available on the Internet I think, love is a good source of these pre trained vectors and. These pre-trained vectors have been trained, on huge corpus of you know English language like Wikipedians, and stuff so they are, also good at generally, understanding, what context words are used in so. Manuel will later show you scales way use the pre-trained word vectors, okay. So. Now we have you know we took our complain and we converted everything into numbers so we're good to use a neural, network. So. In this case we build a CNN for text classification I. Think. A, lot of you might have heard of CNN's they have been quite popular in computer, vision they. Have been used very successfully on, images, for detection tracking, classification. Etc, but more, recently CNN's, have actually, been used very successfully for text classification as well and this might sound a little bit unintuitive so. Part of the goal of this slide for me is to actually explain, to you the, architecture, on a high level and also give, some intuition around why this actually.
Works So, maybe. Let's start with a small sentence. Here I like, this movie very much so this is seven. Words and each, word is, represented. By a five dimensional word vector so you can think of this as a seven by five matrix and, at. This point you can see that's somehow like an image because it becomes 2d and. Then we choose different. Filters so we choose filters, of size. Two words three words and four words and they're represented, by different colors here. These. Filters will, do the convolution, operation on the sentence, one. Way of thinking about this convolution, operation is, the, filter of two words will look at two adjacent words three words will look at three, words and then four would look at four and so this is kind, of like a bag of words so the neural network is trying to look at adjacent words and trying to extract meaning from the. Sentence and, then it creates these feature maps which are these one-dimensional. Vectors. That have summarized. Everything and you can see them as different vectors. Of different colors here, and. Then finally since all of these feature maps are different lengths to, actually. Use them in, the neural network we do sort of like a hacky fix where we do, a 1d, max pooling we take the maximum value from. Each of these feature. Maps and convert combine, them to, create this. Colored vector at the end so. The, way I think of all of this operation is the neural network has or the CNN has gone. Over the sentence and extracted, all the relevant information and combine, that into that colored vector which is the output from this stage. Okay so I want to summarize the, model once more before I move on so we started, with draw text of, complaints. We did, pre-processing, that converted, the text into integers then. We converted all the word integers, into word vectors we, built a CNN in. Our actual CNN we use four different filter, sizes of two words see words four words and five words we, do the concatenation, and. Then the final layer in the model is a fully connected layer with. Sigmoid, activation and. What this does is it. It. Assigns, it. Basically. Assigns. Each complaint into one of the seven departments assigning, them different probabilities, and we can just pick the department, with the highest probability as the prediction from the model so. As. You can see this is a relatively, simple architecture. And it's amazing, you know how well this works. So. We were able to get 85%, accuracy, in. Correctly predicting, Department, on the test set which is which. Is the set that the model wasn't exposed to at the time of training, some. Of other other big benefits, here where the, model is actually quite quick to Train so I took, only about 15 minutes of training time on a single GPU. Since. It's so quick to train it's very, easy to do tuning. Of hyper parameters, or testing our different architectures. It. Transparently, fast at inference time as well so we're able to use a CPU, at, inference, time and that's what we will show in the demo as well right. Now, it. Can also be retrained periodically. So. Let's say you rollout with this model and you use it for three months and you collect new. Real data in the field you can easily plug that back in and retrain the model, and. Then finally I think a more subtle but actually, pretty important point is because this model uses deep learning you don't really need to do any complex feature engineering the model. Itself extract features which means that you, don't need to be an expert in financial, services and you also don't need to be an NLP expert, to use this this. In my opinion makes deep learning very versatile, where same. Architecture, can be applied on different. Problems in different domains legal. In financial, services in, you know wherever, you want to apply them. Okay. So. We, have talked a lot about the model so now I want to actually do. A demo of it let's, see if I can switch screens. Okay. This is awesome so what we did is we designed. A web demo so this demo, allows us to interact with the model more easily. So. In this first page of the demo we have the. Customer's name phone number and address which is arbitrary. At this point because the actual data had this information anonymized, and. Then in this section right, here is where we can put in. Complaint. And then when we when we run it this complaint is run against the CNN. At the back, so. The complaint that I have chosen is an interesting one it says I'm, in the process of refinancing.
With Quicken Loans it's. Been over 45 days and I still don't have a closing date they. Keep giving me the runaround and it's been nothing but a headache, so. What, I personally find interesting about this complaint is there is no mention of any department, here and it's actually a very small complaint, as well so. If you're not familiar with the financial, services domain it might be hard for you know to know what product is being talked about here and and, that I think is the full benefit. Of deep learning is just by looking at a whole bunch of complaints, it has, understood, what are the key terms, involved. With each product, and it's then able to use that so, now. When I hit. Submit. The, neural. Network runs, in the backend and. Actually. Assigns this to, the market department which is correct. So. Besides. Actually predicting, the department, which is what. The case routing, would help with we, also, have, here, what, we call the smart report so, this report has a lot of interesting, insights so, from. The text in the complaint we are able to create this word cloud which has the keywords in the complaint we're, also able to give. A sentiment, index sentiment, score to this complaint we. Can extract organizations. And categories. That are being talked about as, well. As we are, able to look into our historical. Database for. Other similar complaints, and present, that to the agent who's actually working this complaint and. That can be pretty useful right so if you are if you're working. This complaint you might want to know what, what was done in the past and you can apply a similar resolution. So. In the next part of this presentation, we, want to go and talk, about how we actually, created. This smart report switchback. Okay. So I'm, gonna now talk about how we use the Google language, API so, before, I start like by show of hands can I know how many of you have, used, the Google language API before. Okay. That's good to know so, those. Of you who haven't I would actually encourage you to try. It out it's pretty, simple to use so there is the. The home page for the natural language API has. Sort of a section, where you can use, it easily you don't really need to code in Python to, do this so. That's what I want to do now is I want to actually. Showcase. The, the web web page where you can use the language API so this is you know I click, right into invent into the natural language API. Home page and within, here there is the option to try, the API so I picked, sample, sentence that says. ABC. Bank has associated my, number with someone else's delinquent, account they're, calling me every hour and will not remove my number from the other person's, account, so. This person is clearly stressed. So. When I hit, analyze. Essentially. What would happen is that the. Language API, would go and extract. Information. And. Share that with me across four tabs so there's entities, sentiment. Syntax. And categories. The. Entities, portion of the API looks. At things, like hey what's, an organization. What's a person what's an event and. Stuff like that and presents, that information. The. Sentiment. Sentiment. Tab gives me a sentiment, score for the entire, text. As well, as for each individual, sentence so it. Has correctly assigned this sentence. A slightly negative sentiment, which is true and. Then for, the each individual, sentences, there is a sentiment score what I find very very interesting, here is that if you look at the second sentence which, says they're calling me every hour and will not remove my number for the other person's account there's, actually no real big negative, word here but, the model has correctly understood, that the overall sentiment, here is pretty negative and I would have actually done the same I would have assigned this sentence.
A Pretty, negative sentiment, score. So, the API also has very very rich information about syntax so this has things like part of speech like what, works are now in verbs adverbs, etc. It. Also has dependency, map which which, can show you visually how words are related to each other so you can use this to very. Quickly understand, what are the verbs in your sentence, and then do, analysis. After that and. Then. Finally. It also has categories, so for us it's. Very generic it says these complaints are related to financial services and banking which is true but, I think this feature can be very useful if your data set has complaints, of many different types and then, it can actually tell you what. The relevant, categories are quite, accurately and it would be useful. Okay. So now I want, to hand it back over to Manuel, who will walk, you through how we use the language API, in our demo. Thank You Priya. So. Priya gave us a good overview of what, the Google language API is all about and I, want to kind of dive a little bit deeper into how we used it in the smart report it's, a very powerful tool I. Think. These. Are pre-trained, models in the cloud so, whenever, you can I know use, anything that's pre trained whether an image modeling transfer, learning in in using pre trained word vectors and NLP or, things like tools, like. These. Pre trained models. Google. Offers many of them not just a language as vision, this video, there's translation. There's transcription, all sorts of really cool tools whenever, you can push that work on to other people who specialized on that it's a win that allows you to focus. You. And your team on the use case on the business use case and leverage these, tools that work really well and are probably harder, for us to replicate as well, and it, frees, up a lot of time on our end so the. Smart case report this portion of the report is a, fuel. Despaired by the language. API, and. I'm. Gonna walk you through a few, so the, the word cloud is interesting, right it's been used a lot in educational, tools but it's also very powerful in terms of highlighting. Word. Frequencies, which words are used the most is very visual, you don't have to worry too much about a waste present it's gonna present it for you they'll show the largest words, you. Know in a larger font and everything else you know smaller but, it doesn't have to be frequency it can be anything you want right you can program this you could do for example part, of speech which, priya was showing us like the action verbs i want, i need you, know that. Would be also useful in terms of quickly figure out what a piece. Of text is about, they're. Very easy to implement you can program them in all sorts of languages through, rest calls through through python it's just a call to the API and, oh. And, also the sentiment. Index right it's a number between I think negative one and one right and here. We presented. In a linear scale with emoticons with, from very unhappy, to very happy you, know in in a fraction of a second you quickly know the tone of the email so very powerful especially if you're, doing customer service. The. Topics so, those, also refer, to as categories, there, is a category, in a subcategory currently. There is I think about last time I read it about 700, categories, it supports in the financial world does a really good job at, differentiating, it because as you saw our seven departments are pretty closely related it still tells, the difference between all sorts of things very powerful, and a, nice thing about using. The these. Cloud-based pre-training. Models is it's a 700, right now and you. Can go to sleep and when you wake up in the morning suddenly it's a thousand categories, or ten thousand categories right so you automatically, benefit from better pre trained model without having to worry about it because you're just calling the API. The. Organization's, also really cool you'll pull all sorts of.
Entities. That can pull famous. Famous. Locations, it. Can think, locations. In. This case organizations, so really smart, to pull these kind of things and. Quickly. Tell, the the person of the agent has taken care of this who, are the major players in this complaint and. Now. I'll talk about now pulling similar. Complaints. That, were resolved in the past so this is this there are different ways you can approach this problem we thought we had a pretty, clever one here and the. Idea is that you know you have an incoming case and, it in complaint in this in for, this data set and you want to know what is what. Happens well what, have other agents, done in the past and how have they resolved, it right so not only is that going to be a lot of help for the for the agent it's also a step closer to automatic, to to doing auto-replies, right once you know how this works well you can determine well you know I think it can the. The the the machine learning can handle this automatically, that's just fire off a reply. So. You, have this original complaint that comes that imaginary has 20 words right, and we. Want to so, it's hard to you have these 20 words it's hard to kind of fish in an entire knowledge base of historical. Complaints, that we've dealt with in the past and finding a similar one right so we're going to do is going to use pre trained word embedded vectors to boost the vocabulary, right so, what, is a pre, pre I talked a bit about it what is a you, know pre-trained, or D meta vectors so, galuf, is a really popular one they. Have taken either Wikipedia. As a corpus, of the English language or even common crawl which is like a certain, percentage of the entire Internet and modeled. It to figure out exactly how each word relates, to each other word so in the end it's very complex modeling. So that's why we're very happy somebody else did it for us and in, the end you have a matrix with all the the most commonly spoken English words and a long array of numbers and these, dismembers. For each word represents, how one word relates. To another and that's how you can get the the King to queen is you know a man to woman it understands, that there's basically a mathematical underlying. To the English language that's captured, in these pre, trained word vectors and if, anybody has worked, with word Tyvek and has pre trained their own corpuses, it's not easy it takes a lot of time you gotta have the right equipment and you got a tune you gotta you gotta tune it correctly, right otherwise you got to go back and do everything I did that lot in healthcare and it was very uh it was a lot of work and, we didn't have GPUs so it was like sometimes, a week to two weeks of work right and just, to find out that you made a mistake and you have to do it again very frustrating, so.
This Is phenomenal that they make these available the. Glove is from Stanford I think Stanford and Glover working together, Stanford, is maintaining the glove data sets and they're, open source anybody can use them so. Now, so what we're doing we're basically taking every. Word of that original. Complaint let's pretend there's 20 words and here is a well, zoom into one word the word transaction so, then you can say. Using this pre-trained. Matrix. Find, me all the words that are close to, the word transaction, and because, you have this complex matrix, with, that, represents how each word relates. To air to every other word you can use any distance metrics you want Euclidean, distance cosine. Distance it doesn't matter and you basically have created a a an, electronic, very powerful. Thesaurus. Right so instead of having store transaction, you can boost it to buying business, contract, agreement or even a hundreds, of words so, you do that for all your important words not just you know you'll be ignore your stop words all your important words so you take that that case I has 20 words originally, now you can have you know ten thousand words right and you can go fishing basically. A big bag of words and you can go fishing into your knowledge base and finding, the most related, the closely related cases, to, bring, to the surface um, of. Course you've got to do a little you got to weigh the word you can do a little normalization, otherwise the very large historical cases work would take precedence and you don't want that right so there's a little bit more more work involved but that's the overall idea so, and then so, then. You basically would repeat. What I just said you're gonna go with that the huge bag of words instead of you know twenty, now it's a thousand, and pull similar, similar cases so that has three distinct advantages are pretty pretty pretty. Powerful, for. Starters you, can now find cases, you. Can have fun historical, cases that. Don't intersect I have no word at all intersecting, with your original. Date the new incoming case right because you buffer that with a new vocabulary they, could have no word in common and and you can still pull them out the. Second one is there's no training on our end we're not modeling this on our end because, we're we're kind of piggybacking, on the pre trained word vectors right so all that work is pushed on to somebody else we, don't have to do any pre training that save us a lot of time and the. Third advantage which, is related to this is that. Because. We're not doing a pre training we don't have to waste any of our own knowledge based data our own historical, case data for training we, can then use the entire corpus, to, find similar complaints you can going, all the way from the first complaint ever entered in your data warehouse to the word that maybe it was resolved two seconds ago if it's, similar we, can find it and we can pull it so you have access to your entire your, entire four that's pretty powerful, I'm.
Gonna Do a demo as well so. I have a simple case here I'm having issues my bank I have been late on two payments, and I can't get in touch with anybody to work on some kind of repayment plan I'll be upset if this ends up hurting my credit score so. I'll run this. So. The top, of the right so I saying credit reporting which is correct right that. Sounds about right. The original case is there you have you know the the. The the. Information from whom that was that that came from and you have the the sentiment index Oh more, or less correct it's. Slightly, negative to to neutral and you have you know the the word cloud says issues with the bank and even, the word late so after bad after bad you know exactly what what what the what the case is this, becomes way more useful if the case is very long right because you'll see you'll really see things pop out, it. Says you know topic is finance and I, didn't, I purposely didn't put a bank name in there so it's, just says bank as the organization, so you know that's acceptable here and then we see the similar complaints right so there, are complaints that are extremely, long and what, I like here is that it's not pulling anything too large so we, know that there, is some you know it is kind of for now looking, at not only the similar words but also weighing it correctly and we see you know everything, is related to a repayment, plan so you can't repay my loan can't, repay my loan, dealing. With my lender. Need. Information about my loan are struggling to repay it alone and you'll see you know. This. Is not helping you know it's going to I. Was. Seeing you're. Worried. About you know that credit score being affected, here it is I like my picked one here. Late. So. I mean. They. They're all related to to the query in question right and, they, all very different but they're similar right, so that's that's what that's what's really interesting about this approach and. You. Know you can it's, kind of like there's. Word Tyvek and there's doctor vac I don't know if you've heard of these terms this is more paragraph. To vector a B will, make it into a final product where you can take a paragraph, and find a similar paragraph, by adding extra weighting by uttering all these kind of things um. So. I'll. Close this by saying to things and from, seeing the the, people we, know. Raising. Their hands I think we're all kind of on the same page there, are two things that are really uh that are really great right now a natural language processing - in my opinion I've worked in natural language processing for. The past ten years and I've seen such a progression, to enormous, boons to -. To our work is one, is having, access to these pre trained word vectors right this is huge, gilad is one there are tons of other ones you have them for all major languages and you'll have them for specialized, languages, legal, medical. People, have gone. Through the painstaking, work of modeling, huge corpuses, of different. Languages and made, them available have, released, them in the public domain for us to work so I encourage anybody who needs unless you need really specialized, rare, dialects, or rare technical, language use, them it's probably better than anything we can come up with these. Are usually a PhD, students who work on it they do it really well you know we should leverage that work and it's there for us to use the. Second so. Work perfect was actually a big one right but the pre-trained, word. Vectors even better right because you don't have to do it yourself and the second thing that was really a real boon to to natural language processing is for. A ceding language in - convolutional, neural networks, like Korea talked about it's you. Can CNN, seem to do anything these days time series natural language processing along, with you know images, and and and sound, these kind of things that it does normally it, does it really well and a huge advantage is a, lot. Of the feature engineering is done automatically, by the the neural network you don't have to worry about it so it you know and as we get better at this it's gonna get a lot better than any feature engineering we can come up with us humans. We. Do have a github it's not currently open we're gonna we need to we were focusing on getting this presentation done we're gonna clean up the code and then we'll.
We'll We'll. Publish. Every. One of these parts and maybe like in three separate code. Bases. In. In this github repository, if you want access please email us if you want access before that please email us there's. Also the the, link to the demo is there that, also will be open for a bit we may we may close it after a while and you. Have the the link to spring ml and Prius and my in. My email address so you. Know we you know if you have questions about this it doesn't have to be about case routing or anything natural language processing or even anything CNN's we'd love to help out we'd love to get involved if your organization needs help there we also be happy to get, involved so we're gonna open it up for questions if there's any questions please you. Can ask them and you know and we'll be around afterwards as well.