[REPEAT] Better Analytics Through Natural Language Processing (AIM405-R) - AWS re:Invent 2018
•Download as PPTX, PDF•
1 like•5,775 views
Natural language processing holds the key to unlocking business value from unstructured data. Organizations that implement effective data analysis methods gain a competitive advantage through improved decision-making, risk reduction, or enhanced customer experience. In this session, learn how to easily process, analyze, and visualize data by pairing Amazon Comprehend with services like Amazon Relational Database Service (Amazon RDS), Amazon Elasticsearch Service, and Amazon Neptune. We also share real-world examples of how customers built text analytics solutions with Amazon Comprehend.
Recommended
More Related Content
What's hot
What's hot (20)
Similar to [REPEAT] Better Analytics Through Natural Language Processing (AIM405-R) - AWS re:Invent 2018
Similar to [REPEAT] Better Analytics Through Natural Language Processing (AIM405-R) - AWS re:Invent 2018 (20)
More from Amazon Web Services
More from Amazon Web Services (20)
[REPEAT] Better Analytics Through Natural Language Processing (AIM405-R) - AWS re:Invent 2018
- 2. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. BetterAnalyticsThrough Natural Language Processing Nino Bice Product Manager AWS AI A I M 4 0 5 Ben Snively Solutions Architect AWS AI
- 3. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 4. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Analyticswithalldata
- 5. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 6. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. AmazonComprehend D i s c o v e r i n s i g h t s a n d r e l a t i o n s h i p s i n t e x t Entities Key Phrases Language Sentiment Syntax Grouping
- 7. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 8. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Customclassification NEW
- 9. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Automated training Text Label I am calling about my credit card Loyalty My points aren’t being applied correctly Loyalty I really need to shut the service down Account We are moving and we don’t want this anymore Account
- 10. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Customentities Person Organization Part Account_Action NEW
- 11. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Automated annotation and training Entity:Account_Issues Cancel order Large bill Escalate to manager No connection
- 12. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Comprehend structured output PERSON ORG LOCATION Policy Make Dave Amazon Seattle 44-331 Zen Steve Facebook San Francisco 55-433 Theta Pre-trained entities Custom entities
- 13. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 14. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Removing ML/NLP complexity
- 15. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Dmytro Dolgopolov Senior Director Content Services and Analytics FINRA Matt Cardillo Senior Director Application Platforms FINRA
- 16. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. of storageevents per day 30+pb135 Billion Up to Monitoring 99% Equities & 65% Options in the US Reconstructing Trillions of Market Nodes & Edges Investor protection Market integrity
- 17. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Business challenges—MatterContent • High volumes of unstructured content • ~1M documents each year from stock brokers and investors to be reviewed • Documents contain incredibly useful information but mining is a challenge • Overlooking important information • Risk-based approach itself presents risks • Numerous features of interest • Finding information about the Who, What, Where, When and How is labor intensive
- 18. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Business challenges—MatterContent John Doe alleged he did not understand how the two fixed annuities sold to him by William Alex Smith worked and did not understand the impact. At that time William Smith was employed by Company, Inc. Investor John Doe Broker William Alex Smith, ID 12345
- 19. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Our solution Commix Command Center for People and Organizations
- 20. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Oursolution:Project3CPO
- 21. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Architecture <bucket>/input Data Preparation Entity Matching Amazon Comprehend <bucket>/matched
- 22. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Dataflow Feature Sets (collected from document) Name Search (millions) Proprietary Reference Data Comparator
- 23. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Benefits • Benefits to FINRA technology • Extract individuals and organization • Match extracted entities to FINRA records • Flag individuals of interest • Enable higher level analytics • Benefits to our customers • Bring information to the user instead of the user mining for information manually • Flag individuals of interest to find bad actors that could otherwise go undetected • Reclaim hours toward high value, strategic efforts over repetitive tactical document reviews Net Result - Regulatory reviews made
- 24. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 25. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Comprehend use cases Content personalization: Customers are using Amazon Comprehend NLP output to understand related documents based on entities, phrases or even topic similarities for trends analysis, to drive content personalization, and recommendations Semantic search: Customers using Amazon Comprehend to index entities for boosting and ranking search results. Intelligent data warehouse: Customers are using Amazon Comprehend to query unstructured data in relational databases, processing data within the data lake (S3), and then inserting it back into the data warehouse Social analytics: Customers are using Amazon Comprehend to ingest, process, and analyze trends from entities and sentiment from social media posts across Twitter and Facebook. Information management: Customers are using Amazon Comprehend for indexing and finding related content for enterprise information management and various internal business processes including compliance and IT.
- 26. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. AWS textanalyticsworkload Amazon Kinesis Amazon ES Amazon Redshift Amazon EMR • Boosting • Rich Filtering • Grouping, Trends • Joining, Correlating • Clustering • Graph, Search • Near real-time • Alerts Amazon S3 Social Media, Support Amazon Aurora
- 27. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Customerfeedback analytics
- 28. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Customerfeedback analyticsarchitecture
- 29. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 30. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Dataflow: RSS News Feeds EntitiesRSS News Metadata Named Entity Property Graphs
- 31. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 32. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Dataflow: RSS News Feeds Entities Key Phrases Sentiment Syntax
- 33. Thank you! © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 34. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Editor's Notes
- Amazon Comprehend is a natural language processing service that uses machine learning to find insights and relationships in text. Amazon Comprehend identifies the language of the text; extracts key phrases, places, people, brands, or events; understands how positive or negative the text is; and automatically organizes a collection of text files by topic. Our customers are using Amazon Comprehend to identify key topics, entities, and sentiments in social media and news streams, and to enhance their ability to access and aggregate unstructured data from the vast document libraries that exist within their organizations. Hotels.com has thousands of customer views and comments that are submitted by people who stay at the properties. It’s historically been difficult to find what matters in all this data. By using Amazon Comprehend, Hotels.com is able to uncover the unique characteristics that people like or don’t like about each hotel. Consequently, the company is better able to make recommendations to their users.
- Our customers want the accuracy of Comprehend to now support their own organization labels, terms and phrase Finance Insurance Manufacturing What are most important use cases for customization? Organizing documents based on their content Analyzing documents looking for business-specific terms and phrases
- Our customers want the accuracy of Comprehend to now support their own organization labels, terms and phrase Finance Insurance Manufacturing What are most important use cases for customization? Organizing documents based on their content Analyzing documents looking for business-specific terms and phrases Introducing Comprehend Custom Classification and Entity Types
- When our journey we began to the public cloud 5 years ago, high water mark was ~1/2 of what it is today We are FINRA – non-profit SRO with oversight from the SEC We regulate nearly the entire equities market and the majority of the options market State the mission We have massive amounts of data and we perform complex reconstructions to make sense of market data In the past five years, developed core competency for managing structured data in the cloud We do not think of our infrastructure as “fixed” any longer We’ve developed a robust data management solution We employed innovative partitioning strategies on our highly skewed data We leveraged multiple query engines based on our varying usage scenarios So, we’ve done a lot around our structured data…what’s next??
- We are working to develop the same level of competency around our unstructured data FINRA manages a backlog of case work that we refer to as matters. Working on these matters involves lots of unstructured content. Including: form filings; documents (about 1M each year); email correspondence; other data sources as reference Unearthing key features like who, what, where, when, and how is time consuming, error prone, and painful for our regulators. Some of our more complex matters have thousands of files associated. So, how do can our analysts know with confidence that they have not missed something?
- In this simple example… This is where Comprehend comes into play. We building text analytics solutions using Comprehend leveraging Entity Recognition Text Classifier To help identify features important to our regulators out of the piles of content received from filings, information requests, tips, and so on We want to go further then this simple example. We’ll be considering Comprehend’s custom capability around entity recognition, to deal with terms specific to our business of regulation. So, this is what we’re working on…
- We called this capability - Commix Command Center for People and Organizations True to form, like any organization with governmental oversight, we’ve become really good at coming up to flashing names for our capabilities… which always seems to conveniently turn into an catchy acronym
- Now Comprehend enables us to build these solutions. Here you see one of applications used by our users that streamlines the process of reviewing filings usually containing several multipage documents. A big part of such a review is to identify “bad actors”. Those are individuals that FINRA knows had some past infractions and filings that they are associated with will require additional level scrutiny. This easy to read dashboard helps our investigators to quickly identify “bad actors” without any of reading hundred page documents. Investigators can now quickly navigate thru the document and assess the level of risk associated with this filing.
- This is high level architecture of our solution. There are many operations we have to perform in order to extract the wealth of the information found in the documents. There are many regulatory insights we can get from the documents. Once you do it though it creates the new opportunities. It enables machine learning using Sagemaker, link analysis using Neptune, improves enterprise search using Elastic Search – the list goes on and on. Sorry Nino but this is why we want to use Comprehend. It’s like a salad that you have to it before you can have dessert. So this purple broccoli in the middle is very important part of our diet.
- Let’s look under the hood to see how we use entity extraction to match individuals with FINRA records. First we leverage Comprehend to collect various features found in the documents. Those are the features describing the individuals. For instance, name, email, address, employment, etc. Using proximity, context and other techniques we combine those features into feature sets – one per individuals identified in the document. Then we use full text search to limit the number of candidates. Our reference data contains information about millions of individuals. We take in consideration all the name variations. In the next step we use custom comparator that limits the number of candidates to a few and provides the level of confidence. Now we can surface the candidates of interest to the investigator for review. In the past they had to perform a manual search which would be very time-consuming especially for people with common names.
- Benefits of Comprehend Superior Entity and Locality recognition Extract multiple features to increase the probability of entity matches More accurate document categorization through topic modeling Key phase extraction is very beneficial to FINRA’s use case to quick discern signal from the noise Match extracted entities to FINRA records