Upload
firstmark-capital
View
321
Download
1
Embed Size (px)
Citation preview
Data Science for Customer Service
Sameer Maskey
About • Founder & CEO, Fusemachines • Adj. Asst. Professor, Columbia University • PhD Computer Science • Teach Data Science, Machine Learning/Statistical
Methods for Natural Language Processing (NLP) in Columbia University
• We sell Customer Service Automation Platform
We're sorry all a/endants are
currently busy. Please con8nue to hold and
your call will be answered by the next available agent. Thank you for your pa8ence
42+ Billion Customer Service Interactions a Year
Not close to reality Close to reality
Technology Hasn’t Caught up for Call Center Automation
Easy for humans to learn language; Machines are terrible in learning language
True Conversational Machines Not a Reality Yet…
Can We S8ll Automate Customer Service?
Customer Interactions Social Web In Person Mobile Live chat Email Telephone
Contact Center
Browse community forum
Browse company website
Browse FB page
Tweet
Email a service agent
Visit an in store sales agent
Receive info via sms
Navigate an IVR via a smartphone
Online chat forum
Call customer support
Data Generated in Customer Interactions
Social Web In Person Mobile Live chat Email Telephone
Data
Data
Data
Data
Data
Data
Data
Data
Data
Data
Contact Center
Data
Data
Data
Data
Data
Data
Data
Data
Data Data
Can We Use Generated Data to Enhance Customer Service?
Social Web In Person Mobile Live chat Email Telephone
Data
? ?
?
??
??
?
??
Contact Center
Data ?
Automation From Data Social Web In Person Mobile Live chat Email Telephone
Automatically rank new piece
of knowledge from
community forum Data
Data
Data
Data
Data
Data
Data
Data Data
Data
? ?
?
??
??
?
??Based on
context pre-empt and display different help options
Predict the best next step/response based on tweet
Predict the product purchase
Rank the channel preference
Provide best automated response
Predict which agent to transfer
Data ?
Automation From Data Social Web In Person Mobile Live chat Email Telephone
Contact Center
Historical Purchase/ Order Data Context
Data
Historical Usage Data
Tweets and post
Web forum text
Email Text Speech
Automatically rank new piece
of knowledge from
community forum
Based on context pre-empt and display different options
Predict the best next step/response based on tweet
Predict the product purchase
Rank the channel preference
Provide best automated response
Predict which agent to transfer
Automation From Data Historical Purchase/ Order Data
Context Data
Historical Usage Data Tweets and
post
Web forum text
Email Text Speech
Automatically rank new piece of knowledge
from community
forum
Based on context pre-
empt and display different
options Group the
customers for next best action
Predict the product
purchase
Rank the channel
preference Provide best
automated response
Predict which agent to transfer
Unstructured Data Structured Data
Historical Purchase/ Order Data
Context Data
Historical Usage Data Tweets and
post
Web forum text
Email Text Speech
Automatically rank new piece of knowledge
from community
forum
Based on context pre-
empt and display different
options
Predict the product
purchase
Rank the channel
preference Provide best
automated response
Predict which agent to transfer
Ranking Classifica8on Clustering
Group the customers for
next best action
3 Fundamental Problems
• Data to Scores - Ranking
• Data to Classes/Labels - Classification
• Data to Clusters - Clustering
Data Science for Customer Service
• Data to Scores – Rank answers to provide end users – Rank channel preference – Rank new piece of knowledge in community forum
• Data to Classes/Labels – Predict answer type machine should respond with – Predict next best action for customer service representative – Predict which agent to transfer to
• Data to Clusters – Cluster customers based on various features – Cluster topics from text data
• Data to Scores – Rank related questions based on context – Rank answers based on the question input
• Data to Classes/Labels – Predict answer type (number vs list, etc) – Predict dialog act (statement vs rhetorical question)
• Data to Clusters – Cluster topics from text data
Data Science in Fusemachines Customer Service Automation
Platform
Data to Scores Ranking of related questions based on context
Rankings
x
z
y
Data to Classes/Labels
Generative Classifier
Given a new data point find out posterior probability from each class and take a log ratio
If higher posterior probability for C1, it means new x better explained by the Gaussian distribution of C1
p(y|x) = p(x|y)p(y)p(x)
p(y = 1|x) ∝ p(x|µ1,
1)p(y = 1)
• Learn from a few samples of data on how to tag topic labels of answers
Classes
Data to Clusters • Automatically cluster text at various granularity –
sentences, passages and documents • Use the cluster label as kayak like filters for
customer service reps to find answers quickly
Clusters
Data Science for Customer Service
• Data generated in customer interactions can be used to improve customer service automation using machine learning algorithms
• Data Science can be applied across all data sets that is generated in customer-company interaction points
• Full Automation (Conversational Robots) not a reality but we hope that these small steps will take us closer to that possibility
Thank you
@sameermaskey @fusemachines
Facebook.com/fusemachines
linkedin.com/company/fusemachines