際際滷

際際滷Share a Scribd company logo

Understanding Human Conversations with AI

Rajath D M
Rajath D M

Presentation from my talk about using AI to understand natural language human conversations and derive insights, trigger workflows etc. Also talking about a bit on Symbl's conversational intelligence API platform

1 of 33
Download to read offline
Understanding Human Conversations Rajath D. M., March 2021
Understanding Human Conversations with AI
Understanding Human Conversations with AI
Understanding Human Conversations with AI
Understanding Human Conversations with AI
WALK AND TALK
Understanding Human Conversations with AI
We still talk a lot...
We still talk a lot... And theres a lot of untapped knowledge...
Here comes AI
Here comes AI Good for Object/Pattern/Speech Recognition
But not really!
What about Conversations?
Human to Human Human to Machine
H2H H2M Conversational in Nature Highly Unstructured Flow of conversation not predefined Context and purpose is dynamic Transactional in Nature Pretty Structured Pre-defined flow of conversation Context and purpose is static
Existing Conversational AI Systems Built on the state of the art deep learning systems like GPT3, BERT Mostly Intent-based systems that act on a defined scope Needs training data
We must refocus, working towards developing a framework for building systems that can routinely acquire, represent, and manipulate abstract knowledge, using that knowledge in the service of building, updating, and reasoning over complex, internal models of the external world. Gary Marcus,MIT, Reboot.ai
How can we solve the problem?
Will require General Context-Awareness Model all Aspects of Unstructured Conversation Dynamic Conversation Adaptation Unbiased Information Modelling Conversation specific outcomes Self-calibration Analyzing Open Domain Conversations
How do business start implementing conversational intelligence? Start with Speech Recognition Data-science strategy... Rule-based Domain-specific Intent based Open Domain Choose between... Real Time and Asynchronous Languages and Dialects, Channels, Audio Quality Build intelligence Train & maintain the model Continuous Improvement... Gather training data Reduce biases Feedback loop Data Engineering 1 2 3
Other aspects to consider Training Data: Ensure high quality, Unbiased data Building the Model: Analyze and conduct data science experiments to get to production Too Many Models: (Analyze, Build, Deploy, Scale, Maintain) x number of models User Experience: UX Research, Time to design and develop UI, Maintain the UI Build backend around the AI systems: Build and Maintain for real time streaming / batch Integrating with Communication Channels: Build and Maintain Complex integrations with Telephony and other third party dialers / channels Scaling for other types of conversation - start with one type and scale to more
Symbl: How it works? Voice & Video API Real-time or Async Text API Async Telephony Websocket REST Calibration API Conversation API User Experience API Transcription Speaker Metrics Timeline Contextual Topics Action Items Entities Sentiment Topic Hierarchy Questions Calendar Invites 1 2 3 4 INGEST YOUR DATA INTEGRATE TO YOUR CHANNEL GENERATE INSIGHTS BUILD YOUR DIFFERENTIATED EXPERIENCE
Deep Understanding Best of both worlds: Deep Learning (Semi/Supervised) and Classical AI (Unsupervised) Mathematical & Statistical systems can help with inference Deep Learning systems can help with pattern recognition
Build Differentiated Experiences
Using Pre-built UI Coming soon - Symbl JS Elements!!
Getting Started Join our Developer community on Slack Get your TADHack credits early by email - tadhack@symbl.ai Create an account on https://symbl.ai Retrieve your API Credentials Headover to API documentation
Sales/CRM Intelligence Using Recordings or Real Time Calls Real Time Updates to CRM Topics Follow-Ups Question + Email Context + All Context Aggregated Intelligence for Managers
E-Learning Integration Using Lecture or Session Recordings Index or Clip Video Recordings Topics Follow-Ups Question Sentiments Action Items Search and Navigate Videos by Context
Contact Center Using Lecture or Session Recordings Index or Clip Video Recordings Topics Follow-Ups Question Sentiments Action Items Search and Navigate Videos by Context
Symbl - Conversational Intelligence APIs Eliminate cost and complexity Reduce time to market Built on Secure Infrastructure Developer-first platforms No Upfront Training https://github.com/symblai https://docs.symbl.ai/
Capability specific demo
Understanding Human Conversations with AI
Thank you! Questions?

More Related Content

Understanding Human Conversations with AI

  • 8. We still talk a lot...
  • 9. We still talk a lot... And theres a lot of untapped knowledge...
  • 11. Here comes AI Good for Object/Pattern/Speech Recognition
  • 14. Human to Human Human to Machine
  • 15. H2H H2M Conversational in Nature Highly Unstructured Flow of conversation not predefined Context and purpose is dynamic Transactional in Nature Pretty Structured Pre-defined flow of conversation Context and purpose is static
  • 16. Existing Conversational AI Systems Built on the state of the art deep learning systems like GPT3, BERT Mostly Intent-based systems that act on a defined scope Needs training data
  • 17. We must refocus, working towards developing a framework for building systems that can routinely acquire, represent, and manipulate abstract knowledge, using that knowledge in the service of building, updating, and reasoning over complex, internal models of the external world. Gary Marcus,MIT, Reboot.ai
  • 18. How can we solve the problem?
  • 19. Will require General Context-Awareness Model all Aspects of Unstructured Conversation Dynamic Conversation Adaptation Unbiased Information Modelling Conversation specific outcomes Self-calibration Analyzing Open Domain Conversations
  • 20. How do business start implementing conversational intelligence? Start with Speech Recognition Data-science strategy... Rule-based Domain-specific Intent based Open Domain Choose between... Real Time and Asynchronous Languages and Dialects, Channels, Audio Quality Build intelligence Train & maintain the model Continuous Improvement... Gather training data Reduce biases Feedback loop Data Engineering 1 2 3
  • 21. Other aspects to consider Training Data: Ensure high quality, Unbiased data Building the Model: Analyze and conduct data science experiments to get to production Too Many Models: (Analyze, Build, Deploy, Scale, Maintain) x number of models User Experience: UX Research, Time to design and develop UI, Maintain the UI Build backend around the AI systems: Build and Maintain for real time streaming / batch Integrating with Communication Channels: Build and Maintain Complex integrations with Telephony and other third party dialers / channels Scaling for other types of conversation - start with one type and scale to more
  • 22. Symbl: How it works? Voice & Video API Real-time or Async Text API Async Telephony Websocket REST Calibration API Conversation API User Experience API Transcription Speaker Metrics Timeline Contextual Topics Action Items Entities Sentiment Topic Hierarchy Questions Calendar Invites 1 2 3 4 INGEST YOUR DATA INTEGRATE TO YOUR CHANNEL GENERATE INSIGHTS BUILD YOUR DIFFERENTIATED EXPERIENCE
  • 23. Deep Understanding Best of both worlds: Deep Learning (Semi/Supervised) and Classical AI (Unsupervised) Mathematical & Statistical systems can help with inference Deep Learning systems can help with pattern recognition
  • 25. Using Pre-built UI Coming soon - Symbl JS Elements!!
  • 26. Getting Started Join our Developer community on Slack Get your TADHack credits early by email - tadhack@symbl.ai Create an account on https://symbl.ai Retrieve your API Credentials Headover to API documentation
  • 27. Sales/CRM Intelligence Using Recordings or Real Time Calls Real Time Updates to CRM Topics Follow-Ups Question + Email Context + All Context Aggregated Intelligence for Managers
  • 28. E-Learning Integration Using Lecture or Session Recordings Index or Clip Video Recordings Topics Follow-Ups Question Sentiments Action Items Search and Navigate Videos by Context
  • 29. Contact Center Using Lecture or Session Recordings Index or Clip Video Recordings Topics Follow-Ups Question Sentiments Action Items Search and Navigate Videos by Context
  • 30. Symbl - Conversational Intelligence APIs Eliminate cost and complexity Reduce time to market Built on Secure Infrastructure Developer-first platforms No Upfront Training https://github.com/symblai https://docs.symbl.ai/