In just a year, the AI @ T-Mobile team has gone from creating models exclusively in keras or tensorflow with no supports to fully embracing Rasa - and not just for it's chatbot functionality. In this talk, I will go over how we've used Rasa to stand up a beefy customer service chatbot in a company that highly prioritizes human-to-human interactions - as well as the surprise lift our data science team has found from leveraging Rasa models outside of chatbot use cases.
Presented by T-Mobile Sr Machine Learning Engineer, Heather Nolis at the 2021 Rasa Summit (https://rasa.com/summit/).
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
… or how I learned stop worrying and love the chatbot framework | Rasa Summit 2021
1. PAGE1
… or how I learned stopworryingand love the
chatbot framework
formypals at RasaSummit2021
Heather Nolis
MachineLearningEngineer
AI@T-Mobile– @heatherklus
10. PAGE10
Some CustomersPreferSelf-Service
One third of care callsopt-in to a bot experience
Messaging care volume continuallyincreases.
More customers prefer messaging each year.
The onlywaywecould trulybelistening toourcustomersis to builda chatbot– forthose whowantit.
13. PAGE13
A taleof 10 intents, a Self-AssistBotStory
In-HouseTensorflow topic model
88 intents
Hierarchical, defined taxonomy
General topics
Runs on a 10-message window
2,000 utterances per intent (at minimum)
Self-Assist Bot Ask
10 new intents
Overlapping
Highly specific
Runs on a single message
No labeled data
No data labeling support
14. PAGE14
Our topic: General Payment
Intent they wanted: Pay My Bill
Things thatare the topic general payment but are not “pay my bill”
I won’t pay my bill because I don’t understand it.
Checking to see if my payment hasgone through.
I want to change my payment method.
If we treated our topics as intents,we risked showing nonsense
responses to customers.
15. PAGE15
Shop for a device
vs
Add a line and get a new
device too
vs
Add a line but bring your
old phone
“Iwannabuy aCoolPhone”
“Upgrademy phoneto the CoolPhone
“Buya CoolPhoneformy sister’sline”
“Buya CoolPhoneformy sisterandadd the lineforher”
“Iwanttoadd my sistertomy accountbut Iwantthe CoolPhonebogo”
“Mysisterneedstobeadded tomy accountbut isbringingherown CoolPhone”
“We can’t wait monthsfor you to add new intents.”
16. PAGE16
Idea:
Try Rasa
Why Rasa?
Solid machinelearning
open source(wecanchecktheircode)
uses the sameframework(Tensorflow)as our internaltopic model
Lesstraining data
Reuseour custom embeddings
Extensibleinto further bot functionality
Problem:
Time boxedto 4 hours devtime
(training timenot included)
18. PAGE 18 |AI @ T-MOBILE
So what’s different with Rasa?
19. PAGE19
The modelsare different…
BespokeTensorflow Topic Model
Runs on a window of messages
2,000 utterances (minimum) to bootstrap an
intent
About 80% accuracy
Rasa NLU Model
Runs on a single message
About 100 utterances to bootstrap
an intent
83.1% accuracy
20. PAGE20
…but so isthe pace.
BespokeTensorflow Topic Model
2Yearsin market
Hundreds of production releases
<10 model releases
+2 intents
Rasa NLU Model
5months in market
43 production releases
19 model releases.
+28 intents
21. PAGE21
Visibilityleads totrust.
With Rasa X, visibility comes out-of-the-box.
Immediately review the impact of releases in realtime.
Allow stakeholders to review conversations, building trust in our
systems.
Allows stakeholders to suggest improvements directly to mygit
repo– without knowinggit.
22. PAGE22
The burdenof initial data is lessened.
Fast intent creation leads to
rapid experimentation.
Intent:Broken
Canreleasesmallintent “stubs”andquickly iteratewithlive
conversationreviews
Reporting available out of the
box.
Topic modelaudit… stillongoing.
ProdAccuracy is King–butcross-validationmetricshelptarget
areasfor incrementalimprovement.