Hackathon: A Simple Chat

Bot
Manish Shrivastava
Why chat bot?
• (Most?!!) Important AI problem
• Inherently challenging
• Saving grace
• Can be solved
• At different levels of difficulty
• With different set of capabilities
• The complexity of the problem is dependent on the domain
• But, general, open-domain chatbot is WAYYYY too difficult right now
Purpose
• AI/ML is about
• Data Preparation
• Collaboration
• Uniqueness of solution (Bonus)
• Chatbot
• Cannot train data heavy methods in one day
• Purpose of the chat bot is
• Get a taste of AI
• Build meaningful tools in a short time
Complexities
• Even in a simplistic setting
• Each task requires independent treatment
• Knowledge/expertise hungry
• In all cases
• Conversational Natural Language Understanding
• Limited scope
• NLU
• Can be made simple depending on task
Why is NLU challenging?
• Intents
• Multiple ways to say the same thing:
• “Find me inexpensive restaurants with nice ambiance”
• “I am looking for an inexpensive restaurant that has relaxed
atmosphere”
• “Locate cheap restaurants good for informal dinner”

An NLU module often needs to map many different surface texts onto
the same meaning
Saturday, 24 March 2018 ML/AI 5
NLU approaches
• Rule-based
• Internal representation frames
• Rules define how to extract semantics from a string/syntactic tree
• Statistical
• Internal representation: intent and/or semantic tags
• Train statistical models on annotated data
• Classify intent
• Tag domain-specific concepts

Saturday, 24 March 2018 ML/AI 6

Natural Language Understanding
•With Frame model, there are many ways to handle the
meaning of sentences
•For dialogue systems, most common is “Frame and slot
semantics”.

Saturday, 24 March 2018 ML/AI 7

Intent Classification
Intent: get_shows

What is playing in Lincoln Center

What movies are showing at Angelica Film center tonight
List movies at Film Forum after 7pm tomorrow
…

Intent: get_restaurants

Find inexpensive restaurants in Chelsea

Sushi restaurants in the Village
Brunch in Brooklyn Heights
…

Approach: supervised classification (SVM, CRF, Decision Tree, etc.)

Saturday, 24 March 2018 ML/AI 8

Concept Labels
Intent: get_shows

What is playing in Lincoln Center/VENUE

What movies are showing at Angelica Film Center/VENUE tonight/TIME
List movies at Film Forum/VENUE after 7pm tomorrow/TIME
…

Intent: get_restaurants

Find inexpensive/PRICERANGE restaurants in Chelsea/NEIGHBORHOOD

Sushi/CUSINE restaurants in the Village/NEIGHBORHOOD
Brunch/CUSINE in Brooklyn Heights/NEIGHBORHOOD
…

Saturday, 24 March 2018 ML/AI 9

Components of a Chatbot Framework
• Dialog Manager
• Managing dialog state
• Natural Language Understanding
• Intent states
• Understanding the intent of a users input
• Conversation FSA (usually a tree)

Saturday, 24 March 2018 ML/AI 10

Task at hand
• Build chatbots to handle two different domains
• Perform action at the end of the chat
Grading Scheme
• For each of the two tasks:
• 5 marks for dialog state population and understanding
• We will run your chat bot 6 times, randomly varying the input sentences each time. The
chatbot should correctly reach action state with all the relevant information at least
• 3 times for 5 marks.
• >0 times for 3 marks
• Else 0 marks
• 5 marks for information seeking from the group created DB
• 2 marks for 100 diverse entries in the DB
• 3 marks for Correct Query formulation for at least half the runs that reach action state
else 0/3
Skills
• A skill is the ability to gather information required to perform a task
and to actually perform that task
• “With a custom Alexa Skill, you control the requests made by the user, as well
as the words they use to trigger the Skill.” –Amazon
• Each of the tasks provided are a skill
• A simple yet complete task
Requirements
• Some parameters are required for a chatbot skill for a particular task
• What are the “attributes”
• What are the variables required to perform the task? (Source, Destination, time etc…)
• What are the entities?
• City names, store names, book names etc.
• What are the Responses?
• What should the bot say?
• If destination is not set : “Where do you want to go?”
• Possible user inputs
• Bot cannot understand text: We need to prime it for understanding.
• Dialog state
• Some of the above would be dependent on where you are in the conversation
• We have tried to keep it as flexible as possible.
Expectations
• Each team needs to do the following things
• Understand the code (naturally!!)
• Provide necessary entities
• Provide necessary attributes
• Provide the prompts for each attribute
• For each intent list of possible user inputs
• Simplistic assumption
• Create a DB for the task ( ex schema : {model, brand, size , ram }
• Using the attributes fire a query to the DB to fetch the relevant tuple(s)
• Select * from mobile_phones where model=$model AND…
Skills to be developed:
• Two skills
• Common Skill:
• Restaurants booking : Book a restaurant based on cuisine , cost type (cheap,
medium, expensive), location ( east, west, north, south). Final action booking
Skills:
• Buying a laptop. The user should • Buying a mobile phone. The user
be asked questions about Brand, should be asked questions about
RAM, Screensize, Hard Disk Brand, Size, Accessories and
Size/Type, OS and other such other parameters. The bot
parameters. The bot should give should give a final selection
a final selection
Skills
• Buying Vegetables/Fruits. The • Library. After getting
bot should ask which items and information about author, title,
for each item get additional subject and so on the bot should
details. For example, California pick the book
Grapes, Green Grapes …Simla
Apple, Washington Apple, Fuji
Apple ...
Skills
• Jukebox. After getting choices of • Cab Booking: Assuming that the
Genre, Style, Artist, Album etc., starting point is fixed as HYD
the bot should pick the song Airport, the bot should gather
information about your
destination, number of people,
luggage quantity and pick a cab
for you
Skills
• Movie Choice: The bot should • Baby Name Selection: The bot
gather your preferences: should find out the parameters
language, Actor, Genre and the such as Ethnic Group, Region,
date you want to go (today Religion, Gender, Starting Letter,
tomorrow, coming Tuesday) and Other Numerology constraints
suggest the movie and offer the and offer a few choices
Movie hall choices
Skills
• Hotel Room Booking: After • Doctor Appointment Booking:
gathering parameters such as Gather Speciality, Hospital,
Star, With/Without Restaurant, Location, Time of Day, Date
Location, Tariff the bot should information and check if a
offer the choice Doctor is free or not
Skills: SMALL CHANGE
• Ticketing System: The bot should • Ticketing System: The bot should
help automate generating help automate status checking
issues/tickets in an issue tracking issues/tickets in an issue tracking
system, by asking questions system, by asking questions
about the department, issue about the department, issue
type, necessary information and type, find status of a ticket in the
finally a description of the issue tracking system.
problem and generate a ticket in • Resolved, pending etc…
the issue tracking system.
Points to consider
• Consider only discrete values
• Range, regular expressions – Bonus
• Provide enough samples
• You HAVE to have at least two attributes before action state
excluding action and greeting states
• Spelling mistakes handling not present in the current version-Bonus
• Other Bonus tasks
Regarding DB
• DB can be text file (JSON, XML, CSV etc) or MySQL DB dump
• Submission needs to have the DB file
• As of now, there are no DB file provided in the shared code
• DB is a listing of items with their individual attributes
• Sample tuple: {Mantra Restaurant, chinese, cheap, north}
• Provided config file are not the same as DB
• Read each and every one of the config files carefully
• Understand their functionality
Bonus tasks
• Changing an attribute value already given
• Range, regular expressions etc for getting attributes
• Spelling mistakes handling
• Graceful restart at any stage of the chat
• Adding word2vec for intent identification (replacing ngram match)
• Add a branch based on an incoming attribute value
Logistics
• Groups 1 through 40 in TalentSprint
• IIIT-H (NOT 105); rest
• Support available from 9:00am – 6:00pm
• Submission: A zip file containing
• A text file called Team.txt containing the team member names
• All data, configuration and code files
• To be submitted on the LMS by 8pm tomorrow
• Lunch
• At IIIT Mess and TalentSprint
• Hackathon work on personal machines.