Eugene Klyuchnikov, Business Intelligence Lead, TourRadar
~You ask, we don't answer (yet). How to build a perfect ML-based question answering model which doesn't work.~
5. Question Answering System
● Generalizable to different types of questions
● Oh, it must be awesome
● Autonomous doesn’t require human race to operate
● Trustworthy no hassles
● Scalable to infinity and beyond
8. AWS Lex
Question
MS QnA
Analyze
response
Show Lex
Show QnA
Sorry we could not
process your question.
Score > 50
Lex is able to
recognize the intent
RUAVA
Reisen- und Abenteuervirtuellerassistent
10. AWS Lex
Question
MS QnA
Analyze
response
Show _
Show QnA
Sorry we could not
process your question.
RUAVA
I know
everything!
I’m not sure...
Score > 50
Lex is able to
recognize the intent
11. AWS Lex
Question
MS QnA
Analyze
response
Show _
Show QnA
Sorry we could not
process your question.
Lex is able to
recognize the intent
(of course!)
RUAVA
I know
everything!
I’m not sure...
12. RUAVA
● Bad results
○ Intents were very sensitive to new / inaccurate utterances
○ Intents were hard to train (what keywords should we use to trigger intent?)
○ Lex intents were overtaking responds with incorrect answers
○ QnA Score > 50 was a hard limit
● Improvements
○ Train better, automate training phase
○ Use more data
○ Focus on one platform…
○ … or use a completely different approach
13. Completely Different Approach
cosine similarity
TF-IDF, lemmatization
Natural Language Toolkit
Answer from
the closest question
becomes an answer
to original one
14. Completely Different Approach
cosine similarity model
TF-IDF, lemmatization
Natural Language Toolkit
cover by tests,
create api,
ensure security,
dockerize...
add logging,
monitoring,
alerting...
PROD
18. Happy End
cosine similarity
TF-IDF, lemmatization
Natural Language Toolkit
Answer from
the closest question
becomes an answer
to original one
19. Happy End
cosine similarity
TF-IDF, lemmatization
Natural Language Toolkit
Answer from
the closest question
becomes an answer
to original one
20. Happy End
cosine similarity
TF-IDF, lemmatization
Natural Language Toolkit
Answer from
the closest question
becomes an answer
to original one
Internal
knowledge base /
onboarding tool
21. Happy End
● Automating customer support roster
● Working on recommendation systems
● Personalizing SERPs
● Doing reviews sentiment analysis / language detection
● Generating the product descriptions
● Predicting the intent to book
● Segmenting customers