chatbot q a encoding and matching for customer service
play

Chatbot Q&A Encoding and Matching for Customer Service Wenxi - PowerPoint PPT Presentation

Chatbot Q&A Encoding and Matching for Customer Service Wenxi Chen wchen@juji-inc.com Juji, Inc. https://juji.io Goal Chatbots for business 2.6 Billion Customer service (2019) Interview State-of-the-art AI to democratize


  1. Chatbot Q&A Encoding and Matching for Customer Service Wenxi Chen wchen@juji-inc.com Juji, Inc. https://juji.io

  2. Goal • Chatbots for business 2.6 Billion • Customer service (2019) • Interview • State-of-the-art AI to democratize AI • Non-IT professionals can use 9.65 • Faster to build Market Study Report, LLC. Billion • Leverage cutting-edge hardware and software • Deep learning + expert system (2024) • NVIDIA GPUs • Automatic chatbot Q&A generation • vs. Writing code to update chatbot Q&A

  3. Problem • Answer user questions • How do businesses cover the questions and their variations? • How to update those questions and their answers?

  4. Solution A process to evolve the chatbot’s Q&As that: • utilizes the state-of-the-art sentence encoding; • refines deep learning models with NVIDIA GPUs; • updates Q&As in real-time by businesses.

  5. Q&A Solution Flowchart During Conversation User ask a question in Business chat creates/adds FAQs System encodes System encodes System generalizes the question question & answers business questions System tries to match user’s System notifies business question with business questions if Q&As can be improved During Conversation Design Chatbot responds to user question

  6. State-of-the-art sentence encoding • Deep learning models: • Bidirectional Encoder Representations from Transformers (BERT) • Universal Sentence Encoder (USE) • InferSent • They capture semantics, and perform well in evaluations • However, public tasks are different from domain specific customer service scenarios • E.g. a statement with its negation can have highly similar encoding

  7. Compute similarity Siamese Network Finetune (loss) Further Further • Identify criteria for domain specific transformation transformation customer service • Negation • Alternative expression Sentence Sentence • Real world conversation data Encoder Encoder • Encode sentence pairs to compute pair similarity loss Sentence 1 Sentence 2

  8. NVIDIA GPUs to speed up the process • GeForce GTX 1080 Ti • Training time reduction • Fast iteration • Continuous update • 30x increase in # sentences encoded per second • Make powerful deep learning model possible in production • Stable performance

  9. Fulfill the promises of conversational AI • Jennifer for COVID-19 resource • https://www.newvoicesnasem.org/jennifer-ai-chatbot • Jumpstart for education • https://activity.jumpstart.com/#/jsaactivity • And more • https://juji.io/gallery/ • Email hello@juji.io

Recommend


More recommend