Challenges & Design Patterns for Conversational AI fileIncomplete sentences Grammatical errors...

Post on 30-Mar-2019

228 views 0 download

Transcript of Challenges & Design Patterns for Conversational AI fileIncomplete sentences Grammatical errors...

Challenges & Design Patterns for Conversational AI

Peter Skomoroch, Head of Data Products

Introductions

Peter Skomoroch@peteskomoroch

• Co-Founder and CEO of SkipFlag, Enterprise AI startup acquired by Workday

• Co-Host of O’Reilly AI Bots Podcast

• Principal Scientist and early member of the data team at LinkedIn

• Machine Learning and Search at MIT, AOL

Challenges & Design Patterns for Conversational AI

• Didn't understand request

• Wrong interpretation of request

• No results found

• No memory of past conversations

• No knowledge of user’s identity

• No grasp of slang, typos, jargon

• Entity disambiguation errors

Common Scenarios in AI Conversations

Credit: @JamieSkella

• Rule Based Bots & Heuristics

• Slot Filling & Intent Classification

• Generative Models

• Retrieval Based Models

Common Approaches to Conversational AI

Rule Based Bots & Heuristics

Slot Filling & Intents

Generative Models

Smart Reply: Automated Response Suggestion for Email (Kannan et al)

Retrieval Based Models

https://rajpurkar.github.io/SQuAD-explorer/

Narrow Domains vs. Unconstrained Conversations

Knowledge Graphs & Conversational AI

SkipFlag: A Knowledge Base That Builds Itself

• Smart Knowledge Base

• Expert Identification

• Instant Answers

Content is auto-organized into a Knowledge Graph

Entity Understanding and Linking

Job DescriptionKnow python and django, and have some experience with docker

PythonHigh-level programming language

DockerComputer program

DjangoSoftware

Fact Extraction from Text with Linked Entities

Workday was founded by David Duffield, founder and former CEO of ERP company PeopleSoft, and former PeopleSoft chief strategist Aneel Bhusri. It is an on-demand (cloud-based) financial management and human capital management software vendor.

<Workday, Inc.> <founded by> <David Duffield>

<Workday, Inc.> <founded by> <Aneel Bhusri>

Good Training Data is Often the Bottleneck

Credit: @mrogati

Entity Understanding Training Data

Common Crawl: ~4B pages monthly

Challenge: Workplace Dialogue and Internal Jargon

Product ManagerJob Title

Agora ProjectInternal Project

Site AnalyticsInternal Team

Workplace Conversations

● Short messages

● Incomplete sentences

● Grammatical errors

● Alternating speakers

● Meandering topics

● Internal jargon

● Overlapping chat conversations

Conover et. al., “Pangloss: Fast entity linking in noisy text environments”, KDD 2018 (to appear)

Question Answering Training Data

• Don’t assume building a bot for a messaging platform is easier than an app. If you are training conversational AI, it’s much harder.

• User retention issues will cause most bots to fail, unless platforms let them be ambient and contextual.

• Distribution and discovery are still challenging on messaging platforms. You need users to get the conversation data flywheel going.

• Google and Alexa Assistants are becoming a higher level discovery layer that delegates requests to 3rd party skills or conversational agents.

Parting thoughts: Platform Level Challenges

Q&A