Everything around the
NLP
@neal_lathia
London.AI
February 2021
(yes, lots of hiring happening)
NLP applications in
our award-winning
customer support
NLP in help article search
⬅ Search query
⬅ Help article search results
⬅ Prompt for information
⬅ Predefined answer with a
direct call to action
⬅ Self-selected chat closure
⬅ The chat first turn (this could
be more than one message)
Automated self-service answers
and in-chat interactions (2019)
Financial Crime
Account Management
Finding the best help & prioritising tasks
with rules & NLP
Other domains
Recommending saved responses
⬅ Customer first turn
⬅ Friendly goodbyes or new
issue?
⬅ Agent first turn, resolving the
issue
Conversation analysis
How do we enable all
of this to happen?
How to get NLP into production
Train a model Build the rest of the system
Everything around the NLP
Conversation
feature store
Labelled dataset
creation
Monitoring &
analysis
Serving infra &
orchestration
Data labelling
Model store
Conversation feature stores?
Conversation feature stores translate
between messy events & structured input
⬅ Customer first turn
⬅ More turns
Labelled dataset creation?
Labels from elsewhere ➡
⬅ Customer first turn
⬅ More turns
Labelled dataset creation structures datasets
for various types of experiments
Labels from elsewhere ➡
Customer first turn Labels
... ...
Customer first turn Agent first turn
... ...
Model store?
Our model store makes it easier to get
artifacts into production & gives us
insights about what we’re shipping
Lovely data
Serving infra & orchestration?
https://qconlondon.com/london2020/presentation/modern-banking-1500-microservices
Orchestrator
Classifier for A
Classifier for B
Classifier for C
Winner!
Rules that determine
eligibility for topic C
Classifier for D
Serving infra & orchestration enables us to
plug & play models, turn some off, and
surround them with business logic
Monitoring & analysis - we track our
models’ performance in the same place we
track our product’s performance
Data labelling enables us to create unique
datasets using our agent’s expertise
Everything around the NLP
Conversation
feature store
Labelled dataset
creation
Monitoring &
analysis
Serving infra &
orchestration
Data labelling
Model store
Thanks!
@neal_lathia
London.AI February 2021

Everything around the NLP (London.AI Feb 2021)