Development of a Digital
Assistant for Travel
Booking - Phase I
By: Anthony Leung
Agenda
1. Overview
2. Methodology & Results
a. Proof of Concept Model
3. Development Demo
4. Summary and Next Steps
2
1 OVERVIEW
Product Introduction
Development Plan
Product Introduction
NotSnapTravel wants to create a digital assistant (i.e., chatbot) that
allows users to query the following through conversation:
FLIGHT HOTEL CAR
PRICES BOOKINGS RENTALS
4
Development Plan
2020.01.31 20XX
Phase 1 Phase 2
Create a proof of concept Interface model with an API
model that can extract relevant and deploy it in a suitable
flight information from platform (e.g., Slack).
conversational text.
5
Development Plan
THIS PRESENTATION 20XX
Phase 1 Phase 2
Create a proof of concept Interface model with an API
model that can extract relevant and deploy it in a suitable
flight information from platform (e.g., Slack).
conversational text.
6
2
METHODOLOGY &
RESULTS
Natural Language Understanding
NLP Transfer learning (ULMFit)
Named Entity Recognition & Phrase Matching
Development Toolbox
Experiment/
Prototype Deep learning
environment modeling
Data
manipulation NLP
8
How does the chatbot know
?
what to respond to the user?
NLU: What is the intent when we say stuff?
● Natural Language Understanding:
The ability to interpret text.
SPEECH: “I am looking for one-way flight
from Toronto to Tokyo, leaving on Feb 6”
INTENT: SearchFlightPrice
10
Classification: Associating speech with intents
X_input =
“Looking for CLASSIFICATION
one-way flight MODEL
from Toronto to
Tokyo on Feb 6”
Y_pred = RESPONSE
“SearchOneWayFlight” FETCHER
11
Dataset size of 236 across 8 different intents
Intent Count
SearchOneWayFlight 41
SearchHotelPrice 37
SearchCarRentalPrice 35
Greeting 33
Goodbye 26
GiveThanks 25
SearchRoundTripFlight 24
Help 15
Total 236
12
Challenge: Obtaining an accurate model
with a small corpus of travel queries.
!?
How can we obtain one in a feasible
manner?
NLP Transfer Learning: ULMFit
● Universal Language Model Fine-tuning
(ULMFit) [*]
● A transfer learning approach to text
classification.
● Beneficial for tasks with small datasets
(i.e., it’s fast with great accuracy!).
● Can be perform on fastai.
[*] J. Howard & S. Ruder, arXiv preprint:1801.06146 (2018).
14
NLP Transfer Learning: ULMFit
[*] J. Howard & S. Ruder, arXiv preprint:1801.06146 (2018).
15
Training Datasets for LM and NN
a) LM Pre-training:
● Utilize the AWD-LSTM[**] pretrained model.
b) LM Fine-tuning:
● “General-Chat-Log-Telegram” (13,207 sentences) +
Present dialogues w/o intents (236 sentences).
c) Classification fine-tuning:
● Present dialogues w/ intents (236 sentences).
[**] S. Merity, N. S. Keskar, R. Socher, arXiv preprint:1708.02182 (2017)
16
Overall intent classification accuracy: 91%
Intent Accuracy
SearchRoundTripFlight 100%
SearchCarRentalPrice 100%
Greeting 100%
Goodbye 100%
SearchOneWayFlight 92%
SearchHotelPrice 83%
Help 83%
GiveThanks 78%
17
Chatbot blueprint
X_input =
“Looking for CLASSIFICATION
one-way flight MODEL
from Toronto to
Tokyo on Feb 6”
Y_pred = RESPONSE
“SearchOneWayFlight” FETCHER
18
? How can the chatbot recognize
locations and dates from text?
Named entity recognition (NER):
Recognizing places, dates, etc.
● Utilize spaCy library.
? ● E.g., “Looking for one-way flight
from Toronto to Tokyo on Feb 12”
Which one is origin and
destination?
20
How can the chatbot distinguish
?
between origin and destination?
Phrase Matcher: Match sequence of
tokens based on pattern rules
● Can combine Parts-of-Speech tags and NER labels.
# Pattern: “from [this loc]”
start_loc = [{‘LOWER’: ‘from’},
{‘ENT_TYPE’: ‘GPE’}
]
# Pattern: “to [that loc]”
end_loc = [{‘LOWER’: ‘to’},
{‘ENT_TYPE’: ‘GPE’}
]
22
DEVELOPMENT
3
DEMO
(Subject to Changes)
SUMMARY AND
4
NEXT STEPS
Summary and Next Steps
● Transfer learning (ULMFiT) was used to train a neural network that
can classify speech intent.
● Relevant entities of travel queries are extracted by using NER and
can be differentiated by Phrase Matching.
● Next Steps:
○ Further improve the speech intent classifier by including
more dialogue and fine-tuning hyperparameters.
○ Interact chatbot model with a travel API.
○ Deploy chatbot onto a platform (e.g., Slack).
25
