Conversational commerce: emerging architectures for smart & useful chatbots. Dynamic Talks Seattle 3/26/2019
Download as PPTX, PDF1 like365 views
Victoria Livschitz presented on conversational commerce and emerging architectures for voice and chatbots. She discussed how voice commerce is the next big trend after mobile, with smart voice devices entering many households and massive investment from tech companies. Her company Grid Dynamics is experts in digital transformation through technologies like conversational commerce, which involves deep, branched conversations for tasks like discovery, selection and recommendation. They demonstrated a conversational commerce prototype called Grid Genie to showcase the technical feasibility of current conversational technologies.
Conversational commerce: emerging architectures for smart & useful chatbots. Dynamic Talks Seattle 3/26/2019
- 1. Conversational commerce: Founder and CTO Victoria Livschitz Emerging architectures for smart & useful voice and chatbots
- 2. 2 About the speaker • Founder & CEO of Grid Dynamics from 2006, transition to CTO role in 2015 • Principal engineer at Sun; lead architect of SunGrid, world’s first public cloud, 1997-2006 • HPC engineer at Ford Labs 1994 - 1997
- 3. Privileged and Confidential 3 About the company: experts in digital transformation thru emerging technologies Digital commerce technologies Big data & real time analytics infrastructure & applications Re-platforming from legacy/on-premise to microservices on the cloud ML applications: NLP, AI, voice, image recognition, predictive analytics Open Source Cloud-ready Scalable Automated
- 4. 4 Thesis: voice commerce is the next big thing after mobile Voice devices entering households • 9M smart voice devices shipped in Q1 2018 • 50% household penetration expected by 2022 • IOT makes voice control an expected feature Massive investment • From tech giants – over $1B IBM Watson, etc. • From VC - over $700M in 2017 • 20 AI acquisitions – siri, api.ai, viv Technology advances by leaps and bounds • Speech2text approaches human quality • Deep learning-based NLU / NER enters mainstream • Plethora of AI platforms on the market Better interface for many digital interactions • Reduces engagement barrier • Perfect for search, Q&A • Interactive clarification of intent
- 5. Conversational commerce: • Deep, branched conversation • Discovery, selection, recommendation • Large result sets, ambiguity, comparisons • Deep integration with search, catalog, ontology • Customer-trained NLU • Smart dialog management to arbitrate and orchestrate multiple models Other “deep” conversational applications 5 Types of CUI experience Infobot: • Fact-oriented • Few intent-reply patterns Shallow, linear conversation Virtual assistant: • Task-oriented • Few intent-entity-reply patterns Smalltalk: entertainment, empathy, human mimicry
- 6. Other “deep” conversational applications Conversational commerce: • Deep, branched conversation • Discovery, selection, recommendation • Large result sets, ambiguity, comparisons • Deep integration with search, catalog, ontology • Customer-trained NLU • Smart dialog management to arbitrate and orchestrate multiple models 6 Types of CUI experience Infobot: • Fact-oriented • Few intent-reply patterns Shallow, linear conversation Virtual assistant: • Task-oriented • Few intent-entity-reply patterns Smalltalk: entertainment, empathy, human mimicry
- 7. Grid Genie: Conversational demo by Grid Dynamics R&D project from Grid Labs
- 8. 8 Why Grid Genie project? • Demonstrate technical feasibility of current state-of-the-art • Explore the limits of current technologies • Learn to develop, test, deploy CUI applications • Develop reference designs and tooling for all facets of conversational application development • Accelerate development of new CUI applications
- 9. 9 Conversational demo design principles • 100% open and free: written using open source technologies in pluggable architecture • Multiple devices: supports Google Home, Alexa, Siri, etc. via device adapters • Multiple channels: voice-based, text-based, web-based via channel adapters • Multiple skills: specific dialog skills can be added (promo, checkout, order tracking) • Multiple models: mix, match and use best-of-breed ML/NLU models for specific tasks • Integrated with backend services via layers of platform services and enterprise adapters • Containerized: for seamless on premise or cloud deployment
- 10. 1010 Conversational commerce blueprint dialog skills domain skills channeladapter enterpriseadapters enterprise services ontology recommend catalog search data mining / ML dialog manager context intent API NLU API NLG API search discovery factoid recommend notifications smalltalk checkout loyalty order promo
- 11. 11 Let’s chat about cameras • Diverse category of products: about1,000 items in catalog • Wide range of subcategories, from cheap point- n-shoot to high-end professional cameras • Technical products with vastly different features • Some folks know exactly what they want; others don’t know much about them Why cameras? What we are looking to proof? • Device interoperability • One dialog manager arbitrating between multiple (pluggable) skills • Taxonomy-based domain knowledge Better dialog thru better taxonomy • Know the difference between advisory & order taking • Fluidity of conversation, keeping it natural, human-like
- 12. Grid Genie Demo
- 13. 1313 Conversational commerce blueprint dialog skills domain skills channeladapter enterpriseadapters enterprise services ontology recommend catalog search data mining / ML dialog manager context intent API NLU API NLG API search discovery factoid recommend notifications smalltalk checkout loyalty order promo
- 14. 14 Microservices architecture Intent Classifier Dialog ManagerAmazon Alexa Discovery Agent Google Home Facebook WISMO Agent Small Talk Agent Catalog Service Order API Adapter Auth Service Adapter Catalog ETL Order API OAuth2.0 Bulk Catalog API Search Service Order Search Service Shipment Service Auth Service Shipment API Dialog Services Channel Adapters Agents Platform Services Adapters Store Context Platform Storage MongoDB Product Elastic Search NLG Ontology Service Order Indexing Service Catalog Ingestion Public API Shipment Service Adapter Order
- 15. 15 Lessons learned (so far)? • Reference architecture works (by and large) • Closed platforms like Google dialog flow, are easy to start with, but hard to grow with • It is possible to write open, cross-device applications! • But… lack of standards leave you at mercy of (frequently) changing Google/Alexa APIs • So, continuous testing (and full true CICD) is key to spot regressions quickly • Google and Alexa publishing standards help test your application • Creating pluggable taxonomies is hard • Finding good dialog writers is hard • Best practices for writing general-purpose, cross-platform dialog managers are immature • Conversational applications are ready for prime time • It’s fun and practical to write them, although not (yet) cheap or easy
- 16. Q&A
- 17. 17 Conversational platform components Service Description Responsibilities Channel adapters Connect the platform to edge devices Account linking, Auth, SSO, user management Context Context keeps dialog state Shares context between devices and interactions Dialog manager Controls conversation flow maintains conversation context, detects the customer’s broader intent, delegates tasks to an appropriate dialog agent Pluggable dialog agents Implement narrow conversation flow execute particular use case, communicate with NLU, enterprise adapters to fulfill the actions, use NLG to form natural language response Dialog services NLU/NLG/NER service adapters and implementations Provide NLG/NLU/NER capabilities Enterprise adapters Connect agents to data and services Provide access to enterprise APIs Domain ontology Capture domain knowledge deep understanding of business domain, its terminology, slang, and the relationship between terms
- 18. • Neo4J Implementation: 18 • BabelNet • WordNet • DBpedia Content sources: Domain specific ontology aperture f-stop camera telescope diaphragm f-number PART_OF PART_OF SYN SYN SYN
- 19. CNN based model for intent classification Privileged and Confidential 19 Intent classification Can you suggest me an action camera? CLASS PROB ACCEPT 0.01 DENY 0.001 NEXT 0.05 PREVIOUS 0.012 REQUEST 0.011 RESET 0.03 … … WISMO 0.05 DISCOVERY 0.81 SELECTION 0.1 ORDER 0.05 SMALLTALK 0.02 … … suggest recommend action GloVe/fasttext word embedding Matrix representation CNN: Conv layers Pooling Fully Connected
- 20. 20 To learn more • Subscribe to our blog: blog.griddynamics.com
- 21. Appendix
- 22. BiLSTM based model for entity tagging Privileged and Confidential 22 Named entity recognition: BIO tagging BIO tags (Beginning, Inside, Outside) I need diving camera under two hundred bucks GloVe/fasttext word embedding Matrix representation softmax BiLSTM BiLSTM BiLSTM BiLSTM BiLSTM BiLSTM BiLSTM BiLSTM RNN O O B-purpose B-device_name B-relational B-quantity I-quantity B-unit I need diving camera under two hundred bucks O O B-purpose B-device_name B-relational B-quantity I-quantity B-unit
- 23. Privileged and Confidential 23 Named entity recognition: tag grouping unrelated related related related relatedto predict Linear XGBoost model is trained to make binary classification for token pairs I need diving camera under two hundred bucks O O B-purpose B-device_name B-relational B-quantity I-quantity B-unit
- 24. About Grid Dynamics Founded in 2006, Grid Dynamics is an engineering services company built on the premise that cloud computing is disruptive within the enterprise technology landscape. Since that time, we’ve had the privilege to help companies like Microsoft, eBay, PayPal, Cisco, Macy’s, Yahoo, ING, Bank of America, Kohl's, among others, to re- architect their core mission-critical systems, develop new cloud services, accelerate innovation cycles, increase software quality, and automate application management. Grid Dynamics has multiple locations in the USA and Europe, and employs over 1,000 expert engineers worldwide. Privileged and Confidential 24
Editor's Notes
- #5: https://www.canalys.com/newsroom/google-beats-amazon-to-first-place-in-smart-speaker-market Reduces engagement barrier: talking is easy, even kids can interact with voice bot IOT voice devices: https://www.globalme.net/blog/best-voice-controlled-devices-market https://www.microsoft.com/en-us/research/blog/microsoft-researchers-achieve-new-conversational-speech-recognition-milestone/ Auire high quality intent: one of the problems of keyword search is that people are too lazy to type their intent. They prefer keywords. Voice channel allows for longer phrases with more content.
- #6: Only few players will survive long term Strategies: Amazon - cloud ecosystem and retail focus, Google - strong ML capabilities, IBM with deep insight and vertical focus on health and law, Microsoft focuses on employee and enterprise
- #7: Only few players will survive long term Strategies: Amazon - cloud ecosystem and retail focus, Google - strong ML capabilities, IBM with deep insight and vertical focus on health and law, Microsoft focuses on employee and enterprise