Sign Oregon Assignment of Partnership Interest Now

Document type sign assignment of partnership interest oregon now

good evening everyone I am kriti party I'm a Senior Product Manager with Amazon Web Services and today I will be talking about how you can use Amazon text rack for business process automation and document understanding presenting with me today is Adrienne Lam who leads AI projects at change healthcare in this session it will give you an overview of Amazon text rack its features its benefits we will cover some new functionalities and new features we will dive deep into some of the technical details and finally we will see how Tex rack is being used a change healthcare what is Amazon to extract and why is it important for document understanding let's start at a high level of what we call the AWS machine learning stack at the first layer you have the machine learning frameworks and infrastructure that enable you to push the boundaries of what is possible with machine learning today next you have machine learning services our sage make a platform that enables scientist to train and manage machine learning models and finally you have AI services which allow customers like yourself to incorporate machine learning in your applications and this is where we are today Amazon text rack is a machine learning service that enables you to extract text and data from your documents for document understanding Amazon text reads these documents just like a human being does without requiring any configuration or any templates be it financial documents healthcare documents government documents legal documents and many more you can read these documents using our simple api's just like any human being would now I don't need to tell this group why reading documents is important for business but let's spend a few minutes talking about why it is challenging to read documents at scale well first it is slow and error-prone it is difficult to capture relationships that exists between the various data elements in your document it is difficult to capture the different types of data that exist within your documents there are variations between what are seemingly similar forms but are not really from a document understanding perspective and finally there is distortion between the time a document was created and how it was captured because of these challenges the traditional approaches force major trade-offs if you want human beings to read all your documents it is expensive it takes time and it is still prone to errors you can also use traditional software techniques what is known as optical character recognition OCR which works fairly well for simple documents but it outputs only words and numbers and if there is any structure present in your document then it does not do so well there are also extensions to OCR that apply business rule and visual templates to isolate information on a page but these templates are brittle they don't work when the form changes over time they don't work if there is a variation in the document type and they don't work when the image is captured at an awkward angle additionally there is significant management and development overhead to maintaining these templates Amazon text rack does it better Amazon extracts text and structured data from virtually any document the three different type of data elements that we extract our text we extract tabular data that is organized into rigid rows and columns and then form elements that is also structured data but not organized in rows and columns let's take a quick example this is a seemingly intimidating document which is common in the insurance industry let us see the different kind of data elements that is present in this form first we have dense text text that explains the purpose of this form how to fill it and what to do with this form once you have filtered next we have table data that is organized into strict rows and columns and structured like a table we also have forms that is also structured data but not organized in a tabular fashion instead it is a key value pair Amazon text rot extracts all these type of information from the document without requiring any configuration or templates let's see what Ryan Anderson CEO of file vine which is a management software for legal professionals across the globe has to say about text right Ryan thinks to extract is fast accurate and scalable and file vine has been successfully able to use text rack to sir the needs of sophisticated legal professionals across the globe in fact we are thrilled at the early adoption of Amazon text rack in the six months since we became generally available customers from various industries be it healthcare legal financial public sector have been able to use text rack and build really cool applications with it later today you will also hear Adrienne who's here from change healthcare talk about their own text rack use case we are also really encouraged by the vibrant ecosystem of partners be it independent software vendors or system integrators that are helping customers meet their business objectives with text at all of these partners and customers are able to take benefit of Amazon extract to extract text quickly accurately and retain the flexibility of working across various document types they are able to reduce the amount of manual effort that had previously been spent on processing these documents they have been able to reduce the cost their organizations had to incur in processing these documents and they are able to do all of that without requiring any prior machine learning experience so what's new in Amazon to extract since we launched six months ago I can tell you that the team has been busy and I'm excited to talk today about some of our quality updates and some functional improvements in the service first I want to talk about our tables API here is a document with an embedded table as we explained a few minutes ago table is structured information that is organized into rigid rows and columns what's exciting about text rack is you're able to extract these tables or this tabular information without having to define where in the document to find the table how many rows and columns it has where are its boundaries etc now this document which looks simple at the first pass is not really first of all the scan isn't perfect the boundaries are not obvious and before we evaluate how text rack does on this output let's first think about how we would label this table so a human being would label this table to have six rows and seven columns now let's see how to extract does the original output that you would have gotten from text track six months ago would have correctly identified the table but as you can see it is not perfect there are some additional rows and columns that have been inserted that did not need to be there there is certain misalignment between column headers and cell values and that occurred due to the complications of having mode cells within the table with the quality update now to extract gets it right it has identified the table correctly there are no extra columns or rows inserted and the column headers are correctly assigned to the cell values let's look at another complex example this is the same insurance form that we had picked at a few minutes ago you can clearly see that it has tabular information present but which cell is aligned with which column is not immediately obvious unless you spend a lot of time on it here is how to extract used to identify this table it identified both the primary tables but it missed a section the right note extract was right in determining that the section on the right doesn't really belong with the table on the left but it did not identify the section on the right as a table in itself and this is the best part about using a managed service like Amazon tech strap you can count on the continuous improvements that the service team is going to continue to make and see and rest assured that the underlying models that you use in your applications will continue to get better over time now the extract identified both the tables more accurately and more completely so within the same document now to extract provides you more information next I would like to talk about Tech's racks integration with Amazon's Augmented artificial interests that was announced in Andy's keynote earlier today augmented artificial intelligence or Amazon a2i as sometimes we call it allows you to combine human judgment with the AI output this allows you to apply nuanced human judgment when you see fit and still get the benefits of AI and machine learning how is Amazon's augmented AI service use well the first step is that you have to define the workflow you have to define the business conditions that should trigger a human evaluation you have to define who are the personas and the identities that will be doing the human evaluation and then you need to define instructions or rules by which they will be reviewing your documents the second step is to determine what kind of documents you want this human judgment to be applied to specifically within text track you can set multiple business rules that will trigger a human evaluation the first example is setting a business rule based on confidence course let's say you want to send your document for human evaluation if the confidence score threshold is below 90% as a user I know that the document I am processing is complicated and that if text track provides a comparatively lower confidence score then I might want second pair of eyes on it send it for human evaluation the second set of business conditions that you can set for human evaluation is based on important keys within your forms now this is where you are able to take advantage of the knowledge you have a priori about your documents to specify that if a certain important key or field is missing from the text fact output then there must be something wrong and it should be sent for human validation as a user I know that an adjusted gross income should be present in this form so if it is not it should be sent for human valuation lastly you can use a random sampling of your document to be sent for human evaluation for your own QA purposes and all of these thresholds and limits are customizable so you can set these based on your use case this is very powerful for some of our customers here is Chris from UK national healthcare service who says that the combination of a2i and Tech's track has enabled NHS to actually take advantage of the power of machine learning for the first time and that it is a game changer for them I'm also excited to talk about some of our other functionalities in the recent past and the near future on October 10th we announced that Amazon text rack is now hip eligible which means that US healthcare businesses that have to comply with HIPAA regulation can now process their pH I or protected health information workloads on Amazon Tech's track this was launched October 10th on next week we are also going to be increasing the file size limits of a synchronous API to ten megabytes of course you can also use our asynchronous API that offers a size of 500 megabytes apiece and then we continue to expand our regional footprint on Monday next week we will launch Sydney in Northern California making it a total of seven regions that Amazon text rack is available in today with more to come in the new year now let's take a peek under the hood of Amazon text rack and answer some of the questions that our customers ask us they take their POC to development and then production the first is how to make sense of the JSON output because there's a lot of it the output from text rack which is in the JSON format is extensive and covers a lot of information as an example the single page that you see here on the left has more than 22,000 lines in its JSON output now this amount of information is very powerful for customers so long as you know how to read it and make sense of it let's focus on the yellow section highlighted on the left here and the 32 lines corresponding to that yellow highlighted section these are the same 32 lines from the previous slide and let's see what is the different type of information that is represented here the first is the content itself the text and the numbers in this case the the value is 54 next is the confidence score associated with that content and finally the have the structure of the blog type of the content the relationship between various blog types and the geometry of that content within that page finally within the structure is the block ID that establishes the relationship of that block with the rest of the page text rad organizes its content into a hierarchy of block there is a one block for each content type for text the page is related to line is related to word for tables the page is related to tables is related to cell and the cell is related to the words within it and for forms the page has key value pairs and the text associated with the key and value respectively so within each block you get the content the relationships between the different blocks and the geometry that allows you to locate it on the page as powerful as the text rack JSON output is our customers have asked us for a simpler output depending up their need so we have taken the JSON output from text tract and transformed it into providing three different type of simpler outputs that you can also use depending upon your application needs the first is a simple text file that contains the words and numbers that were extracted byte extract the second is a CSV file that contains the tables and the key value pairs that was that was extracted by text track and finally you have what we call a searchable PDF which is a flat image of your document superimposed with an invisible layer of text extracted by text drag so when you do a controller for a command F it finds and highlights the third string on the page itself these scripts allow you to use to extract in a simpler way and use it within your applications you can find these scripts in the link that is on the page or also on the QR code which is on the bottom left now let's pause and talk about confidence course there's a lot of vite assigned to the confidence score but we don't really spend a lot of time talking about what they mean and how they are derived so that the confidence score from text rat which is assigned to each block type ranges from zero to one and is derived as a combination of signals from the multiple and various machine learning models that are at play as an example the confidence score associated with the word text tract is made up of the individual count of confidence course of each characters that make up the words what do these do the confidence scores are designed in such a way to enable our customers to choose between a trade-off between the best precision and recall depending upon the use case the customers can choose high precision low recall which means that less content is detected but also means that there is less noise or they could choose the latter or the other alternative which is lower precision high recall which means a more content is detected but there could be some noise this trade-off is what is represented by the confidence score output by Amazon to extract a higher confidence score means higher precision means even though the content detected might be Lord is also less noise or a lower confidence score means that more content is detected by contrast but may have some noise so this is how we intend you to interpret the confidence score the confidence score the way it is intended to be used also depends upon the content type it is associated with so for text the confidence score answers the question was this content extracted accurately and for structured entities like tables and key value pairs the confidence code represents if text tracked was able to correctly establish the relationship between these entities and identify its position on the page now as promised I'd like to invite Adrian on the stage and talk about how they are using text tracked at change healthcare thanks.thanks Creedy alright so as mentioned I'm Adrian I work at change healthcare as a lead data scientist so healthcare is really expensive I know you feel it I feel it as well and projections don't make it seem any better in fact it's expected to rise to ap roximately six trillion by 2027 or a fifth of our GDP and out-of-pocket spending will also increase as well to add on top of this if you look at how much we spend per capita the u.s. spends fundamentally the most per human on healthcare and we actually if you look at rankings of healthcare we are nowhere near the top so fundamentally what's happening is that we spend more per person for worse healthcare so now I know what you're thinking why is healthcare it's so expensive fundamentally we believe there's a lot of waste in the system and this leads to extra costs we estimate that to be about 30% or trillion dollars where is this waste coming from well health care is still an exceptionally manual process there's a lot of humans doing things that are very repetitive tasks they're making sub-optimal decisions as they're trying to navigate our healthcare system which is really complex so let me give you example I'll dive a little deeper into what's going on in the backend of the healthcare system so if I get sick and I ended up in the hospital for a few days the hospital needs to write up what's called a medical claim and that claim is what gets sent over to the payer and that's how the hospital gets reimbursed for services rendered writing up that claim is a surprisingly complex process they need someone specialized someone called a DRG coder and what their job is to do they get to read my medical record figure out all the diagnoses that were diagnosed during my stay all the procedures that were done during my stay and they need to code that up so that if I had pneumonia they have to look that up all right maybe that's J 189 it's basically a mapping problem but it gets more difficult when you think of how many diagnoses that are out there so it's a natural language processing problem into classification but it's a 70,000 class problem to add on top of that Medicare comes out with updates every single year where it eliminates codes adds codes updates codes the American Heart Association every quarter comes out with the coding clinic that says hey you can't use these codes this way anymore so hopefully you're starting to see why if we have humans doing this they start making very sub-optimal decisions and the result of that comes out an increased time it takes time for people to go through look through these medical records look up codes understand all the information out there it's increased cost the prea provider has to spend extra overhead just to get paid for things that they have done and if someone does something wrong in the process that claim comes back to the provider and they have to rework that claim it's called a denial it's about at least a hundred dollars per claim that comes back and that just leads to friction in the provider pair relationship so what can we do about this so fundamentally since we have humans doing all this work we have data and because we have data we can start applying AI to all of this so if we have machines up front we can make accurate decision first with all this information we have we can train AI to do the right thing but AI isn't meant to be perfect ai can can't solve all problems but again solve a fair number of them so we can have machines take care of the things that are very straightforward very easy things that we know are not complex have them do it and we can route the things are more complex to humans and we need smart about how we route things to maybe different employees have different specializations will you be intelligent about what kind of work we give to which people and fundamentally we're just hoping to reduce the number of rework that's being done it sucks for hospitals to get denials because they just have to put in extra work just to get their money so who are we who has changed healthcare so I changed kal care we have about like a hundred offices across the u.s. 15 thousand employees we've been around for 20 30 years doing this in total there's about a trillion dollars in healthcare claims that we process per year or 14 billion healthcare transactions we see about 2/3 of the United States healthcare so if you've been to the hospital twice in your life there's probably a 90% chance that we had an effect on that Network transaction for that claim another fun fact if we just cease to exist one day the healthcare system would just come to a grinding halt so we think we're in a pretty good position since we've built up relationships with those providers and pairs and have started generating all this data to really apply AI so I'm going to dive into a problem that we face which is related to auditing medical records so when medical records get submitted payers are interested in auditing to look for fraud waste and abuse but a audit requires validation of multiple elements it's validation of the diagnosis codes that are there the procedure codes that are there the one I'm going to look at specifically are the itemized hospital bills which need to be extracted from the medical records so auditors can go through and validate line by line that services were actually rendered this is what it looks like a couple examples of itemized bills and you can start to see where texts racked might come into the picture so fundamentally they look like tables but there's a little bit of complexities that you might not notice from these tables these are just two examples of I bills but you can see there are pretty different floor mats some have just the line dividers some have just the column dividers some are left aligned some are right aligned if you look really closely at the quantity and charge amount it might be difficult to figure out where that edge of that column is so we have all these itemized bills that are varying structures in terms of what you might see for tables and as mentioned before it's currently an extremely manual process we have an Operations team that their sole job is to turn these tables into Excel files it sounds ridiculous because it is and they use a software they go through they try and detect tables if they can't they get to draw a bounding box over where that table is they get to draw where the rows are they get to draw where the columns are and then they export it over to excel if any OCR errors come up they get to fix that too it's so there Thursday have such a backlog that there's even a process if you want your I bill extracted expedited you have to go through a specialized process through them to get that done fundamentally table extraction is actually a pretty hard problem it's a two-part problem the first part is where's the table or table detection and - what does the table look like like the structure what are the rows what are the columns I've got a table up there on the left and I know kriti showed a couple examples as well you can see some rows they don't expand all the way over some columns don't expand all the way down maybe you'll have some empty rows as well it's difficult there's not that many data sets out there as well there's one from Las Vegas run from Washington one from Dar it's a document analysis conference happens every year every other year sorry the last one was actually this year in Australia that's about 20,000 images 5000 tables we also have table bank which seems like more it's four hundred thousand tables but those were word document tables and latech tables so they're actually fitting a very specific schema if you think about scale in terms of what's a large Dimmick dataset ambe's net is 14 million images so this scale of data sets of tables there's not a lot out there we've gone through different approaches throughout the years for just historically what has been done if you know what your table looks like if there are certain features you can extract automatically you know let the template is you can just go with the rules based approach that's fine there are other image processing techniques that can be done there's a package called Camelot it's open source they look at PDF metadata or they use image processing techniques like Hough transform there's duke outlines so if your table has very distinct lines you can start easily identifying where rows and columns are and as you might imagine it can go into deep learning table Bank has open source models of what it's done is not you're just Jake Joe's mill model they actually use state of the art image net models that have won competitions they open source it it's available to try if you were at it dar this year you might have heard of something called graph neural networks those are something really cool that have been really starting to explore in the document analysis space it first came out if you think about molecules where you have like carbon atoms bound binded to each other it might not make sense as much to represent that as like a number but more is a graph of things connected and documents are the same way in terms of if you look at things across rows things that are above and below each other it's actually meant a little bit better to recognize documents and process them as graphs and when it comes to tables this makes a lot of sense and this is basically where state of the art is that but fundamentally it is still a very difficult problem so where does text rack come into the problem well text rack change healthcare really simplify this instead of going through any complexity we get AI bills they're scanned in for us they just drop them as PDFs we use texture OCR table extract some minor data reform adding and then it's just sent to the OP scene for validation so really streamlined process very simple very quick where does value come in well I've used a couple different OCR technologies in terms of text detection so OCR is also sort of two-part is detecting where's the text and then what is the text in terms of text detection it does a wonderful job it's better than the ones I've used previously so that's great it helps structure very variably unstructured data so data which you have tons of different formats you might not know what it looks like it helps us structure that for us automatically overall it's helping us improve efficiency of just everyday operations it's making people's lives better and more manageable and just making them faster and happier and we're hoping just it'll streamline the entire auditing process so how do we continue to use this what's down the line for us fundamentally there are also different areas we can use this so in other additives there's something called clinical validation and a clinical validation audit we're looking at lab values and saying hey based on these lab values and medical necessity it says if you have this lab value you have this disease if you don't have this lab value it is not this diagnosis so a lot of these lab values they've ever if you've ever had like your blood drawn and you get the results back of like what are your white blood cells your red blood cells those often come back in tables so we can use text track to help us find those tables automatically extract them and format them for us and fundamentally after that we can build models on top of those extracted lab values to just help us streamline quicking the auditing process we've found these values we've confirmed this diagnosis we can move on so ultimately our goal is to make health care faster better cheaper it's a slogan our chief a officer talks about a lot and it's something I actively try and do every day health care at its core is still a very heavily manual process and Amazon extract helps make that easier medical records are messy they come in all sorts of formats and if you can structure that that makes machine learning on it that much easier and ultimately we're just trying to build a better healthcare system and that's the end of what I've got to present back to you Creedy Thank You Adrian that was a great overview of the challenges faced by US healthcare industry and how you are and what change healthcare is doing to help thank you so much now wrap up I just want to leave you guys with a few messages the first is that reading documents at scale is expensive and error-prone and that you can use Amazon text tract to extract text and structured data table and key value pairs from virtually any document without requiring any configuration or any templates you can use Amazon text rack to extract data quickly and accurately with flexibility across different document types reduce any manual effort you have to put in lower your processing cost for document processing and do all of this without requiring any significant machine learning experience we talked a little bit about some of the quality and the feature updates that the team has been working on and my final thought is that getting started with text rack is really easy I encourage all of you to go to our console upload a document and see what text Rack is able to retrieve for you thank you so much for your time today I will be hanging out here with Adrian at the side in case any of you want to come up and chat and please don't forget to leave the feedback in the mobile app thank you [Applause]