1 of 13

HOW DO INTELLIGENT ASSISTANTS UNDERSTAND AND ANSWER QUESTIONS?

Unit 2, Module 2.3

1

Day 3 - How do intelligent assistants understand and answer questions?

(Query) Syntax and Question Answering

Google queries; Berkeley parser; mention schemas

Question Answering
Representation for the questions - Berkeley Neural Parser

Parts of speech: nouns, verbs, prepositions
Syntactic relationships
How BNP works - Not a purely syntactic parser it uses meaning of words and statistics

Semantic parsing
Sources of knowledge
Limitation: Systems are a collection of techniques

Question: How does Google understand and answer questions?

Demo -

Try asking Google these questions:

*What do wolves eat?

*Where do wolves live?

*What sounds do wolves make?

*How long do wolves live?

*Do wolves have families?

*Where did President Obama grow up?

*What are the biggest cities in Georgia?

*What are the smallest towns in Georgia?

Conversion: Convert feets to yard?

Flights:

Facts: Presidents

Q. Let’s explore how Google might make sense of these queries:

Use The Berkeley Neural Parser to parse sentences and observe the parse tree.
Uses the parse to figure out the primary intent / type of query of your question
Google knows about a bunch of different types of queries (uses specialized reasoning engines - use different techniques to understand the query, if they don’t understand than they do keyword search)

Here are some query engines

Definition
Current Information modules (Time, weather, temperatures, movie schedules, bus schedules, train schedules, flights)
Conversions (units, currencies, distance, time)
Restaurant reservations
Flights
Facts - Knowledge graph queries
Note: some of these modules have a way of failing -> perhaps there is a typo
Google has thousands of engineers trying to salvaging it.

Knowledge boxes, knowledge panels or similar queries
Otherwise a generic keyword search

Student Activities

Sentence -> formula -> module computes

Convert 5 feets to ? meters

Ask a non-sensical questions - COnvert 5 feet to years - suggests a corrections for years to yards and then when you force it then it sends generic information

Show a schema related to one of the project

Explore Schemas (Part of the deeper dive)

Schema.org - restaurants

Deeper Dive

What is everything google knows about restaurants? Look at the schema at schema.org

What is a restaurant? Different conceptual understandings - a business, location, chain vs individual store, brand, -Take away
When you design a knowledge graph you have to decide how much complexity you want to include. Humans make decisions about what is important to encode and for what purpose.

Identify areas of ambiguity in human speech that makes it complex
Explain how computers/AI overcome the ambiguities in human speech

2 of 13

Overview of Topics

Question Answering
Representation for the questions - Berkeley Neural Parser
Parts of speech: nouns, verbs, prepositions
Syntactic relationships
How BNP works - Not a purely syntactic parser it uses meaning of words and statistics
Semantic parsing
Sources of knowledge
Limitation: Systems are a collection of techniques

2

3 of 13

How do you think Google understands and

answers questions?

3

Discussion

4 of 13

Google Query Scavenger Hunt

Units conversion

Ask Google: “Convert 5 Dollars to Euros” or “Convert 5 Miles to Kilometers”

Pick your own conversion problem: ____________________

What does Google return?_____________

Facts about famous people or places

Where was famous person born?

Pick your own fact: ____________________

What does Google return?_____________

Look up restaurant hours
Look up weather
Look up airline flights, bus, or train schedules
Word definitions: define “craven”
Movie times

4

https://developers.google.com/search/docs/basics/how-search-works

Link to student assignment

5 of 13

Let’s explore how Google might make sense of these queries.

5

Google knows about a bunch of different types of queries (uses specialized reasoning engines - use different techniques to understand the query, if they don’t understand than they do keyword search)

Here are some query engines

Definition
Current Information modules (Time, weather, temperatures, movie schedules, bus schedules, train schedules, flights)
Conversions (units, currencies, distance, time)
Restaurant reservations
Flights
Facts - Knowledge graph queries
Note: some of these modules have a way of failing -> perhaps there is a typo
Google has thousands of engineers trying to salvaging it.

Knowledge boxes, knowledge panels or similar queries
Otherwise a generic keyword search

Student Activities

Sentence -> formula -> module computes

Convert 5 feets to ? meters

Ask a non-sensical questions - COnvert 5 feet to years - suggests a corrections for years to yards and then when you force it then it sends generic information

Show a schema related to one of the project

Explore Schemas (Part of the deeper dive)

Schema.org - restaurants

6 of 13

Processing the Query:

Tokenizing - Breaking the sentence into words and punctuation
Part of speech tagging
Parsing the request to determine the meaning of the user’s question
Creating a search request that reflects the user’s intent

6

7 of 13

Tokenizing the Query

Ask Google: “What are the largest cities in Georgia?”

Break the sentence down into words and separate out punctuation.

7

What

are

the

largest

cities

in

Georgia

?

8 of 13

Part of Speech Tagging

Assign a part of speech to each word.

8

What

are

the

largest

cities

in

Georgia

?

WP

VBP

DT

JJS

NNS

IN

NNP

.

WH-pronoun

Verb: present tense

Determiner

Adjective: superlative

Noun: plural

Preposition

Proper noun

Terminator

9 of 13

9

This noun phrase is what the query is about

This prepositional phrase modifies “cities” by specifying a location

This noun is the class of thing we’re asking about

This tells us that we have a “what”-type question.

This adjective modifies “cities” by specifying a size

Parsing the request to determine the meaning of the user’s question

10 of 13

Creating a database search request that reflects the user’s intent

The query generator knows that the “size” of a city most often refers to its population size. It creates the following query:

“Find cities where population size of city is large and location of city is Georgia.”

10

FIND x WHERE� class(x) = “city” AND� population_size(x) is large AND� location(x) = “Georgia”�SORT BY population_size

11 of 13

Google’s Response�What are the largest cities in Georgia?

11

Specialized Google Response �Image results of cities ordered by population sizes

Information captured from a website

Questions people also ask

General list of websites with relevant content based on keywords “largest cities in Georgia”

12 of 13

Google has specialized modules for handling some common types of queries

Unit conversions (measurements, currencies, distance, time)
Weather
Airline, train, and bus schedules
Restaurant hours, locations, phone numbers, websites
Movie showtimes and locations
Biographical information about famous people

12

13 of 13

Lesson Recap Activity

Complete the assignment using Google: Google Search Activity

13