Transcript 2006 Paula Matuszek

Artificial Intelligence
Paula Matuszek
What is Artificial Intelligence

Definitions
– The science and engineering of making intelligent machines,
especially intelligent computer programs. It is related to the
similar task of using computers to understand human
intelligence, but AI does not have to confine itself to methods
that are biologically observable. (McCarthy, 2002)
– The exciting new effort to make computers think ... machines
with minds, in the full and literal sense (Haugeland, 1985)
– The automation of activities that we associate with human
thinking, activities such as decision-making, problem solving,
learning ... (Bellman, 1978)


Strong AI and Weak AI
Turing Test
©2006 Paula Matuszek
What Methods Does AI Use?

AI can also be defined in terms of what
kinds of methods it uses
– Search
– Knowledge Representation
– Inference
– Logic
– Pattern recognition
– Machine Learning
©2006 Paula Matuszek
Typical AI Domains







Games
Natural Language Processing
Planning
Perception
Robotics
Expert Systems
Intelligent Agents
©2006 Paula Matuszek
So when WILL we decide that
computers are intelligent?
©2006 Paula Matuszek
How Do We Know When
We're There?

Some requirements I think any test we
use must meet:
– Whatever test we use must not exclude the
majority of adult humans. I can't play
chess at a grand master level!
– Whatever test we use must produce an
observable result. "Isn't intelligent because
it doesn't have a mind" is perhaps a topic
for interesting philosophical debate, but it's
not of any practical help.
©2006 Paula Matuszek
What can AI systems do?
Here are some example applications

Computer vision: face recognition from a large set
 Robotics: autonomous (mostly) car

Natural language processing: simple machine translation

Expert systems: medical diagnosis in narrow domain

Spoken language systems: ~1000 word continuous
speech

Planning and scheduling: Hubble Telescope experiments

Learning: text categorization into ~1000 topics

User modeling: Bayesian reasoning in Windows help

Games: Grand Master level in chess (world champion),
checkers, etc.
©2006 Paula Matuszek
What can’t AI systems do yet?








Understand natural language robustly (e.g.,
read and understand articles in a newspaper)
Surf the web
Interpret an arbitrary visual scene
Learn a natural language
Play Go well
Construct plans in dynamic real-time domains
Refocus attention in complex environments
Perform life-long learning
©2006 Paula Matuszek
AI Uses in Information Science




Retrieval
Ontologies
Intelligent Agents
Text Mining
©2006 Paula Matuszek
Challenges and Possibilities

Information overload. There’s too much.
We would like
– Better retrieval
– Help with handling documents we have
– Help finding specific pieces of information
without having to read documents

What might help?
– Statistical techniques
– Natural language processing techniques
– Knowledge domain based techniques
©2006 Paula Matuszek
Retrieval


Find correct documents, with high
precision and high recall.
AI used extensively for:
– Determining relevance: heuristic rules
capture human intuition about importance.
Improves precision
– Using domain models: using domain
models/ontologies with synonyms and
classes improves recall.
©2006 Paula Matuszek
Retrieval: Some Current Directions

Intelligent spiders
– Can't cover all of the web; it's too big!
– Determine relevance as documents are retrieved;
spider only those with high relevance
– Goal is to improve precision AND recall

Intelligent disambiguation
– When you search for "bank" do you mean the
financial institution or the side of a river?
– Use ontologies to find multiple meanings
– Scan for related words to choose meaning

Semantic web
– Add meta-information as you create web pages.
Intelligent data instead of intelligent tools.
©2006 Paula Matuszek
Ontologies



Definition: An ontology is a formal description
or specification of the concepts and
relationships in a domain.
Synonyms, hierarchy of terms, richer relations.
Example: cat
– Synonyms: pussy, feline, kitty
– Is a: mammal, pet
– Subclass: Persian, Siamese, tabby
– Has characteristics: carnivorous, purrs
©2006 Paula Matuszek
Ontology: Another Example

Example: Panadol
–
–
–
–
–
–
–
–
–
–
Broader term: chemical drug substance
Narrower term: acetaminophen tablet
synonyms: Tylenol, acetaminophen, paracetamol
Preferred term: paracetamol
Trademarked in country: UK, US, EU.
Company-holding-trademark: SmithKline
Ingredient-in: Contac
USAN: acetaminophen
BAN: paracetamol
Therapeutic class: analgesic agent, antipyretic agent
©2006 Paula Matuszek
Intelligent Agents





Definition: a software program which
autonomously gathers information or
performs some task for a user.
Communicative
Capable
Autonomous
Adaptive
©2006 Paula Matuszek
Some Current Intelligent Agent Tasks

Screen out junk mail
– Understand what makes mail junk: Hand-built
rules or machine learning

Shopbots: Find the best price for X
– Know about and access shopping sites
– Know about and understand costing: Price for
items, discounts, shipping fees

News and mail alerts
– Understand what I am interested in
– Watch relevant sources to find those things and
bring them to my attention

Recommender systems
– What movies or books might I be interested in?
– Collaborative systems, faceted or characteristicbased systems.
©2006 Paula Matuszek
Intelligent Agents: The Vision


Lucy calls her brother Pete: "Mom needs to see a specialist and then has to
have a series of physical therapy sessions. I'm going to have my agent set up
the appointments." Pete agrees to share driving.
At the MD office, Lucy instructs her agent through her handheld browser. The
agent
– retrieves information about Mom's prescribed treatment from the doctor's agent
– looks up several lists of providers
– checks for the ones in-plan for Mom's insurance within a 20-mile radius of her
home and with a rating of excellent or very good on trusted rating services
– finds a match between available appointment times (supplied by the agents of
individual providers through their Web sites) and Pete's and Lucy's busy schedules.


The agent presents a plan. Pete doesn't like it – too much driving, and at rush
hour, and has his agent redo the search with stricter preferences about
location and time. Lucy's agent, having complete trust in Pete's agent in the
context of the present task, supplies the data it has already sorted through.
A new plan is presented: a closer clinic and earlier times—with warning notes.
– Pete will have to reschedule a couple of his less important appointments.
– The insurance company's list does not include this provider under physical
therapists: "Service type and insurance plan status securely verified by other
means. (Details?)"


Lucy and Pete agree and the agent makes the appointments.
Pete asks his agent to explain how it had found that provider even though it
wasn't on the proper list.
Example taken from Scientific American article on the Semantic Web, May, 2001.
http://www.scientificamerican.com/article.cfm?articleID=00048144-10D2-1C70-84A9809EC588EF21&catID=2
©2006 Paula Matuszek
Text Mining


Common theme: information exists, but
in unstructured text.
Text mining is the general term for a set
of techniques for analyzing unstructured
text in order to process it better
– Document-based
– Content-based
©2006 Paula Matuszek
Document-Based

Techniques which are concerned with
documents as a whole, rather than
details of the contents
– Document retrieval: find documents
– Document categorization: sort documents
into known groups
– Document classification: cluster documents
into similar classes which are not
predefined
– Visualization: visually display relationships
among documents
©2006 Paula Matuszek
Document Categorization

Document categorization
– Assign documents to pre-defined categories

Examples
– Process email into work, personal, junk
– Process documents from a newsgroup into
“interesting”, “not interesting”, “spam and flames”
– Process transcripts of bugged phone calls into
“relevant” and “irrelevant”

Issues
– Real-time?
– How many categories/document? Flat or hierarchical?
– Categories defined automatically or by hand?
©2006 Paula Matuszek
Categorization -- Automatic





Statistical approaches similar to search engine
Set of “training” documents define categories
– Underlying representation of document is bag of words
(BOW): looking at frequencies, not at order
– Category description is created using neural nets,
regression trees, other Machine Learning techniques
– Individual documents categorized by net, inferred rules
Requires relatively little effort to create categories
Accuracy is heavily dependent on "good" training
examples
Typically limited to flat, mutually exclusive categories
©2006 Paula Matuszek
Categorization: Manual


Natural Language/linguistic techniques
Categories are defined by people
– underlying representation of document is stream of
tokens
– category description contains
– ontology of terms and relations
– pattern-matching rules



– individual documents categorized by pattern-matching
Defining categories can be very time-consuming
Typically takes some experimentation to "get it right"
Can handle much more complex structures
©2006 Paula Matuszek
Document Classification

Document classification
– Cluster documents based on similarity

Examples
– Group samples of writing in an attempt to
determine author(s)
– Look for “hot spots” in customer feedback
– Find new trends in a document collection (outliers,
hard to classify)

Getting into areas where we don’t know ahead
of time what we will have; true “mining”
©2006 Paula Matuszek
Document Classification -- How

Typical process is:
– Describe each document
– Assess similarities among documents
– Establish classification scheme which creates
optimal "separation"

One typical approach:
– document is represented as term vector
– cosine similarity for measuring association
– bottom-up pairwise combining of documents to get
clusters

Assumes you have the corpus in hand
©2006 Paula Matuszek
Document Clustering

Approaches vary a great deal in
– document characteristics used to describe
document (linguistic or semantic? bow?
– methods used to define "similar"
– methods used to create clusters

Other relevant factors
– Number of clusters to extract is variable
– Often combined with visualization tools based on
similarity and/or clusters
– Sometimes important that approach be incremental

Useful approach when you don't have a
handle on the domain or it's changing
©2006 Paula Matuszek
Document Visualization

Visualization
– Visually display relationships among documents

Examples
– hyperbolic viewer based on document similarity;
browse a field of scientific documents
– “map” based techniques showing peaks, valleys,
outliers
– Faceted search results showing document counts
for different categorizations, with browsing

Highly interactive, intended to aid a human in
finding interrelationships and new knowledge
in the document set.
©2006 Paula Matuszek
Content-Based Text Mining

Methods which focus in a specific
document rather than a corpus of
documents
– Document Summarization: summarize
document
– Feature Extraction: find specific features
– Information Extraction: find detailed
information

Often not interested in document itself
©2006 Paula Matuszek
Document Summarization

Document Summarization
– Provide meaningful summary for each document

Examples:
– Search tool returns “context”
– Monthly progress reports from multiple projects
– Summaries of news articles on the human genome


Often part of a document retrieval system, to
enable user judge documents better
Surprisingly hard to make sophisticated
©2006 Paula Matuszek
Document Summarization -- How

Two general approaches:
– Extract representative sentences/clauses: extractive
– Capture document in generic representation and generate
summary: abstractive

Extractive
– If in response to search, keywords. Easy, effective
– Otherwise term frequency, position, etc;
Broadly applicable, gets "general feel“. Current state of art.

Abstractive
– Create "template" or "frame"
– NL processing to fill in frame
– Generation based on template
Good if well-defined domain, clearcut information needs. Hard.
©2006 Paula Matuszek
Feature Extraction


Group individual terms into more complex entities
(which then become tokens)
Examples
– Dates, times, names, places
– URLs, HREFs and IMG tags
– Relationships like “X is president of Y”


Can involve quite high-level features: language
Enables more sophisticated queries
– Show me all the people mentioned in the news today
– Show me every mention of “New York”

Also refers to extracting aspects of document
which somehow characterize it: length, vocab, etc
©2006 Paula Matuszek
Information Extraction


Retrieve some specific information which is
located somewhere in this set of documents.
Don’t want the document itself, just the info.
– Information may occur multiple times in many
documents, but we just need to find it once
– Often what is really wanted from a web search.


Tools not typically designed to be interactive; not
fast enough for interactive processing of a large
number of documents
Often first step in creating a more structured
representation of the information
©2006 Paula Matuszek
Some Examples of Information Extraction

Financial Information
– Who is the CEO/CTO of a company?
– What were the dividend payments for stocks I’m
interested in for the last five years?

Biological Information
– Are there known inhibitors of enzymes in a pathway?
– Are there chromosomally located point mutations that
result in a described phenotype?

Other typical questions
– who is familiar with or working on a domain?
– what patent information is available?
©2006 Paula Matuszek
Information Extraction -- How


Create a model of information to be extracted
Create knowledge base of rules for extraction
– concepts
– relations among concepts

Find information
– Word-matching: template. "Open door".
– Shallow parsing: simple syntax. "Open door with key"
– Deep Parsing: produce parse tree from document


Process information (into database, for instance)
Involves some level of domain modeling and
natural language processing
©2006 Paula Matuszek
Why Text Is Hard
Natural language processing is AI-Complete.
 Abstract concepts are difficult to represent
 LOTS of possible relationships among
concepts
 Many ways to represent similar concepts
 Tens or hundreds or thousands of
features/dimensions
http://www.sims.berkeley.edu/~hearst/talks/dm-talk/
©2006 Paula Matuszek
Text is Hard





I saw Pathfinder on Mars with a
telescope.
Pathfinder photographed Mars.
The Pathfinder photograph mars our
perception of a lifeless planet.
The Pathfinder photograph from Ford
has arrived.
The Pathfinder forded the river without
marring its paint job.
©2006 Paula Matuszek
Why Text is Easy


Highly redundant when you have a lot of it
Many relatively crude methods provide
fairly good results:
– Pull out “important” phrases
– Find “meaningfully” related words
– Create summary from document
– “grep”

Evaluating results is not easy; need to
know the question!
©2006 Paula Matuszek