Monday, 2 July
Virtual Sales Agents
for Electronic Commerce
Wolfgang Wahlster
German Research Center for Artificial Intelligence
DFKI GmbH
Stuhlsatzenhausweg 3
66123 Saarbruecken, Germany
phone: (+49 681) 302-5252/4162
fax: (+49 681) 302-5341
e-mail: [email protected]
WWW:http://www.dfki.de/~wahlster
Outline
1. Virtual Sales and Shopping Agents
2. Virtual Webpages
3. Life-like Characters as Virtual Sales Agents
4. Virtual Sales Teams
5. Information Extraction Agents for E-Commerce
6. Wrapper Induction and Programming by Example
7. Ontological Annotation of Webpages
8. Encoding Rule Knowledge for E-Commerce Agents
9. Conclusions
© W. Wahlster
Intelligent Agent Technology is a Prerequisite
for Advanced WebCommerce
Virtual Web
Pages
Shopbots for
Automated
Comparison
Shopping
Text Analysis
and Generation
Multimedia
Presentation
Planning
Information Extraction from
HTML/XML Documents
Advanced
WebCommerce
User Modeling and Language Generation
Coordinated Text & Graphics Planning
Machine
Translation
Intuitive,
Multilingual
Access
Multimodal
Interfaces
Robust Dialogue
Understanding
Advanced Speech
Synthesis
Dialogue with
Virtual Sales
Agents
© W. Wahlster
One-to-One
Marketing
Five Generations of Internet Applications
2000
Embedded
Internet Agents
WWW
Research Net
1
Mobile
Internet
Services
EMail
2
3
4
5
Every Car has
Internet Access
a homepage,
via WAP and UMTS
Agents are main
devices
Internet users,
Ubiquitious Computing
© W. Wahlster
t
What are
Virtual Sales
Agents?
l appear as life-like characters
l plan interactive behavior
autonomously
l can initiate interaction
ACTIVE
l understand the user‘s
requests
l answer clarification
questions
l allow mixed initiative
dialogs
INTER- INTERNET
ACTIVE AGENTS
REACTIVE
PROACTIVE
© W. Wahlster
l anticipate the user's needs
l adopt the user's goals
l provide unsolicited comments
l respond immediately
to interruptions
l criticism and
clarification questions
l direct manipulation
Intelligent Web Services
Consumer
buys
l Information
l Goods
l Services
Web Sites
© W. Wahlster
Netbot
l Intelligent Parallel Retrieval
l Information Extraction and
Summarization
l Personalized Presentation
l Matchmaking
l Teleshopping Assistance
l Telemarketing Assistance
l Translation Services
l Data Mining Services
Provider
sells
l Information
l Goods
l Services
Knowledge about:
l Usage Patterns
l User Models
l Consumer Profiles
XML-based Negotiation between Shopping
and Sales Agents
Customer
Shopping
Agent
Negotiation
based on
the Exchange
of
XML
Documents
- Call for Bids
- Offer
- Criticism
- Alternate Offer
© W. Wahlster
Sales
Agents
Companies
Virtual Market Places with Human and
Machine Agents
© W. Wahlster
Performance Ranking of Comparison
Shopping Agents
Performance Ranking Agent
Ranked List
of all Shopping Agents for
a Product Category
Benchmark Problem
Jango
Pricescan
Buybuddy
Cadabra
MySimon
Comparison Shopping Agents
© W. Wahlster
Roboshopper
Three Generations of Web Sites
First Generation
Second Generation
Interactive Web
Sites
Static Web Sites
Fossils cast in HTML
Virtual Webpages
JavaScripts and Applets
Netbots,
Information Extraction,
Presentation Planners
Database Access and
Template-based Generation
User Modeling,
Machine Learning,
Online Layout
Dynamic Web Sites
© W. Wahlster
Third Generation
Adaptive Web Sites
The Idea of Personalized Netbots
e.g. MetaCrawler
Softbots
Personal
Assistants
Indices,
Directories,
Search Engines
Mass
Services
WWW
Traveller’s Netbot: Tries to achieve traveller’s goals (finding and executing plans)
l
l
l
l
checks availability
finds best price
uses personal preferences (e.g. frequent flyer programme, seating preferences)
lets the traveller know, when seats become available (active help)
© W. Wahlster
What is a Virtual Web Page?
Virtual Memory, Virtual Relation, Virtual Reality...
A Virtual Web Page
l is generated on the fly as a combination of various media objects from
multiple web sites or as a transformation of a real web page.
l looks like a real web page, but is not persistently stored.
l integrates generated and retrieved material in a coordinated way.
l can be tailored to a particular user profile and adapted to a particular
interaction context.
l has an underlying representation of the presentation context so that
an Interface Agent can comment, point to and explain its components.
© W. Wahlster
Virtual Webpage Retrieved from 5 Different Servers
© W. Wahlster
AiA: Information Integration for Virtual Webpages
PAN Travel Agent Andi
Car Route Planner
Yahoo
News
Server
Yahoo
Weather
Server
Hotel
Guide
© W. Wahlster
Gault Millau
Restaurant
Guide
The Generation of Virtual Webpages with
PAN and AiA
Hotel
Agent
Trip Data
Address
Netbot
PAN
Map
Agent
AiA
Pictures and
Graphics
Presentation
Planner
Pieces of Text
Components
of virtual
Webpages
Coordinates for
Pointing Gestures
Input for Speech
Synthesis
Icons for
Hyperlinks
Persona
Server
Constraintbased
Online
Layout
Weather
Agent
© W. Wahlster
Train & Flight
Scheduling
Agent
Major Event
Agent
Virtual
Web
Presentation
Persona as a Personal Travel Consultant
© W. Wahlster
The Combination of Retrieved and
Generated Media Objects for Virtual Webpages
Multi-Domain
Problem Specs
Multiple
Data Sources
Information Structures
l Relations, Lists
l KR Terms
© W. Wahlster
NETBOT
Retrieved
Results
Distributed
Information
Media Objects
l Texts, Sounds, Videos
l Pictures, Maps,
Animations
The Combination of Retrieved and
Generated Media Objects for Virtual Webpages
Information Structures
l Relations, Lists
l KR Terms
Select Canned
Media Objects
Design New
Media Objects
l Icons,
Clip Art
l Frames,
Sounds
l Graphics,
Animation
l Text,
Speech,
Mimic
Select & Design
© W. Wahlster
Retrieved
Results
Media Objects
l Texts, Sounds, Videos
l Pictures, Maps,
Animations
Coordinate
Media Objects
l Temporal
Synchronization
l Spatial
Layout
Transform
Media Objects
l Clip, Convert,
Abstract
l Zoom, Pan,
Transition
Effects
Reuse & Transform
Operational Models of Referential Semantics for
Robots and Netbots (Wahlster 1999)
Robot
Netbot
Set of Subsumption
Relations in an Ontology
Set of Subsumption
Relations in an Ontology
“Screw”
“Departure Time”
Set of Recognizers
Physical
Objects
© W. Wahlster
Screw 1
...
Screw N
Set of Wrappers
WWW
Objects
DT 1
...
DT N
The Role of Ontological Annotations for the Generation
and Analysis of Virtual Webpages (Wahlster 1999)
Webpages
with Ontological
Annotations
Webpages
without Ontological
Annotations
Information
Extraction
Agent
Presentation
Planner
Virtual Webpage
Presentation
Agent
Persona
© W. Wahlster
With Ontological Annotations in: SHOE, OML,
XOL,OIL, DAML
and Persona Annotation in
PML
Information
Extraction
Agents
TriAS
A Natural Language Agent for Finding
Pre-Owned Porsche Cars
Boxter, not red, must
have AC, less than 20k
© W. Wahlster
Towards Mobile and Speech-based
E-Commerce Using UMTS Phones
l UMTS phones
(Wireless Application
Protocol for Cellular
Phones)
l WML as a markup
language for interactive
content
l Mobile access to virtual shops allows price comparisons during real shopping
l Multimodal dialog: Voice In (Speech) - Web Out (Graphics, Hypertext)
l Voice input using advanced speech understanding technology
l Easy to use: customers simply say what they want
© W. Wahlster
Enhanced ECommerce through Personalization
System is able to flexibly
tailor product presentations
to the individual user and the
current situation.
An animated character
serves as “Alter Ego”
of the presentation system.
Personalized Presenters at DFKI
© W. Wahlster
Personified Agents Increase the User's Trust in
the System's Presentation
Experimental evidence for effects of modality on the user's trust (van Mulken, 1999)
The system gives recommendations, which turn out to be wrong in some cases.
How much does a user trust the system's advice depending on the modality of
a presentation?
1.0
0.8
0.7
0.6
0.5
Self-animated Persona,
Speech,
Speech, Gesture, Facial Graphical
Expression, Pointing Highlighting
© W. Wahlster
Text,
Graphical
Highlighting
Impact of the modality of a Presentation on the
User's Trustfulness
Result:
Persona > Speech > Text
Conclusion: If the presentation is more human-like,
recommendations are more readily followed
For l decision support systems
l tutoring systems
l recommendation systems
l virtual sales agents
personified interface agents have a clear advantage:
They increase the user's trust in the information presented
by the system
© W. Wahlster
PPP’s Persona Server implements a generic
Presentation Agent that can be easily adapted to
various applications
Visual Appearances
Behaviors
l
l
l
l
Presentation Gestures
Reactive Behaviors
Idle-time actions
Navigation actions
Hand-drawn
Cartoon
Bitmaps
Persona
Server
Auditory Characteristics
l Sound effects, auditory icons
l Voice: male, female
Video Bitmaps
Generated
Bitmaps
from
3D-Models
© W. Wahlster
Use of a Life-like Character for
Electronic Commerce
Digital Assistant Selector
© W. Wahlster
DFKI‘s PET-Technology: Flexible Realization of
Virtual Sales Agents
Agent in its
own Frame
simple implementation
limited presentation
behavior
© W. Wahlster
Sales Agent on
Webpage
advanced presentation
behavior
complex
implementation
Sales Agent
on Desktop
very advanced presentation
behavior
download of sales agent
Classification of Persona Gestures
Gesture Catalogue
Talking Posture 1
• cautious, hesitant
• appeal for compliance
• avoids body-gestures
© W. Wahlster
Talking Posture 2
• active, attentive
• self-confident
• uses body-gestures
Context-Sensitive Decomposition of Persona
Actions
High-Level
Persona Actions
take-position (t1 t2)
Context-Sensitive
Expansion
point-to (t 3 t 4)
r-stick-pointing (t 3 t4)
move-to (t1 t2)
(including Navigation Actions)
Decomposition
into
Uninterruptable
Basic Postures
r-turn (t 1 t21)
r-hand-lift (t3 t 31)
r-step (t 21 t22)
f-turn (t22 t2)
r-stick-expose (t 31 t 4)
Bitmaps
...
© W. Wahlster
...
...
...
Extensions of the Representation Formalism
Distinction between production and presentation acts
(i.e. Persona- or display acts)
Explicit representation of qualitative and quantitative constraints
Production Act
Presentation Act
Introduce
CreateGraphics
S-Show
S-Position
Elaborate-Parts
S-Wait
Label
Label
S-CreateWindow
S-Depict
S-Point
S-Speak
S-Point S-Speak
Qualitative constraints: Create-Graphics meets S-Show, ...
Metric constraints: 1 <= Duration S-Wait <= 1, ...
© W. Wahlster
PET: Persona-Enabling Toolkit
Objective:
l Enable non-professional computer users to populate their
web pages with lifelike characters
PET comes with:
l a set of characters and basic gestures
l an easy-to-learn Persona markup language
Developer’s PET will include:
l a character design tool which enables users to build their
own characters
Technical Realization:
l Based on XML and Java
© W. Wahlster
The Persona Markup Language
Specification of
Persona actions
Features:
– XML-based
– easy to learn
© W. Wahlster
<html>
<head>
Specification of the
<title> Persona Test </title>
character to be used
</head>
<body>
<persona bitmap=“cartoon” ...>
<uselib url= .../>
<do name=“greet”/>
<do name =“speak” args=“hello”/>
</persona>
</body>
</html>
Functional View of PET
Bitmaps
Webpage with Reference to
Java Applet
URL of Webpage
with Persona Tag
<html>
<head>
<title> Persona Test </title>
</head>
<body>
<persona bitmap=“cartoon” ...>
<uselib url= .../>
<do name=“greet”/>
<do name=“standard”/>
<do name =“speak”
args=“hello”/>
</persona>
</body>
</html>
PET Application
Server
PET
Parser
PET
Generator
Persona Scripts
waitscreen 4
gesture greet 0 0 null
gesture laugh 0 0 null
...
Persona Engine
Behavior Monitor
Audio Data
© W. Wahlster
<html> ...
<APPLET
archive=“personaplayer.jar”...
</APPLET>
...</html>
Event
Handler
Character
Composer
The Bidirectional Control Flow on
Persona-Enabled Webpages
Triggers
actions of the Persona
l Text Input
l Speech Input
l Menu Input
l Direct Manipulation
Input
Web Persona
© W. Wahlster
Triggers
operations on elements
of the webpage
l Mouse Clicks
l Mouse Movements
Sending Interface Agents to Clients:
Plug-Ins or Applets?
Plug-Ins
l
l
l
l
l
Add features (character players) to browser
Download triggered by user
Requires disk space on client
Unrestricted access to client
Less appropriate for WebCommerce,
Guides
l Agents integrated in 3D environments
l Appropriate for Entertainment
Examples:
l Extempo's Jennifer James
(Hayes-Roth et al. 98)
l PFMagic's virtual petz
Applets
l
l
l
l
l
l
l
Java animation code sent over the net
Automatic loading
Requires no disk space on client
Restricted access to client
Appropriate for WebCommerce, Guides
Agents integrated in 2D environments
Less appropriate for Entertainment
Examples:
l DFKI's Web Persona
(Müller et al. 98)
l ISI's Adele (Johnson et al 98)
New in AiA/PAN: Balanced combination of Applets and Servelets
Efficient distribution of client-side Java and server-side Java for driving the
Interface Agent
© W. Wahlster
Porsche 9 11 & Boxter
© W. Wahlster
Persona Active Elements (PAE)
l
Active Images
An active image starts a persona action when clicked.
<ACTIVEIMAGE SRC=“image” HREF=“url” NAME=“image name”
STATUS=“status message” ALT=“tooltip” CACTION=“persona action onClick”
OACTION=“persona action on MouseOver” ...>
l
Addressable Objects
An addressable object is an object which can be addressed and
manipulated by Persona via its name and its position.
<PDIV DIVNAME=“name of the element” DVFRAME=“frame name”
TOP=“anchor-y” LEFT=“anchor-x”>some HTML elements</PDIV>
© W. Wahlster
A Virtual Sales Agent for OTTO – World’s
Largest Tele-Ordering Company
© W. Wahlster
DFKI’s Ecommerce Presentation Planner has
been extended to accommodate for various
target platforms through the introduction of a
mark-up language layer
Presentation Planner
PETPML
PET
Persona
Player
© W. Wahlster
Agent
Script
MS-Agent
Controller
SMIL
SMIL
Player
WML
WMLBrowser
Simulated Dialogues as a
Novel Presentation Technique
 Presentation teams convey certain rhetorical
relationships in a more canonical way
l Provide pros and cons
 The single presenters can serve as indices which help
the user to classify information.
l Provide information from different points of view, e.g.
businessman versus tourist
 Presentation teams can serve as rhetorical devices that
allow for a continuous reinforcement of beliefs
l involve pseudo-experts to increase evidence
© W. Wahlster
Presentation Teams for Advanced ECommerce
I recommend you this SLX
limousine.
© W. Wahlster
Underlying Knowledge Base
 Representation of domain
l FACT attribute car_1 consumption_car_1
 Value dimensions for cars adopted from a
study of the German car market
l safety, economy, comfort, sportiness, prestige, family
and environmental friendliness
l FACT polarity consumption_car_1 economy negative
 Difficulty to infer implication of dimension
on attribute
l FACT difficulty consumption_car_1 economy low
© W. Wahlster
Example of a Dialogue Strategy
Question:
How much gas does it consume?
Answer:
It consumes 8l per 100 km.
Negative Response:
I’m worrying about the running
costs.
Dampening Counter:
Forget about the costs.
Think of the prestige!
© W. Wahlster
Header:
(dampening_counter ?agent ?prop
?dim)
Constraints:
(*and*
(positive ?agent)
(pol ?prop ?other_dim
positive))
Inferiors:
(Speak ?agent
(“Forget about the ” ?dim “!”))
(Speak ?agent
(“Think of the ” ?other_dim “!”))
Multiple Interface Agents for User-adaptive
Decision Support
User-Adaptive Search Planning
...
Spare parts for
this car are rather
expensive!
© W. Wahlster
...
weighted propositions
Multiple Decision Support Agents
But, it’s
fast!
MAUT (Multi-Attribute Utility Theory) - I
 formalism for the evaluation of structured objects
 basic idea: identification of relevant dimensions for
the evaluation of an object
n
 total evaluation v ( x ) of an object x : v ( x )   w i v i ( x )
i 1
 evaluation v i ( x ) of the object x regarding the
dimension d i :
v i ( x )   rai v ai ( l ( a ))
a  Ai
 definition of the relationships between dimensions
and attributes within the ontology
 decision support: connection between currently
focused data and user preferences
© W. Wahlster
MAUT (Multi-Attribute Utility Theory) - II
hotel
user´s interest on
the dimension
0.5
cheapness
sportiness
dimension
relative weight of an
attribute for the dimension
0.8
0.1
0.4
0.2
culture
0.5
0.5
...

10

10
10

10
evaluation function
0

0
tennis
© W. Wahlster
0
1

0

0
0
1
golf
1
0
0
10000
price
MAUT - Example
hotel
user´s interest on
the dimension
0.5
sportiness
dimension
relative weight of an
attribute for the dimension
0.8
0.1
0.4
cheapness
0.2
culture
0.5
0.5
...

10

10
10

evaluation function
0

0
0
1
tennis

0

0
0
1
golf
10
1
0
0
10000
price
 hotel 1:
l tennis
l golf
l price: 5000 DM
 hotel 2:
l tennis
l price: 2000 DM
 v ( h 1 )  10  0 . 8  10  0 . 2   0 . 5
 0  5  0 . 5   0 . 4  6

v ( h 2 )  10  0 . 8  0   0 . 5
 10  0 . 5  8  0 . 5   0 . 4  7 . 6
© W. Wahlster
Research Topics: Multiple Interface Agents
 Interactive Presentation Teams
 Corpus-based Approach to Gesturing
 Empirical Evaluation of Presentation
Teams
© W. Wahlster
Non-Interactive Presentation Teams
I recommend you this
SLX limousine.
© W. Wahlster
Characteristics of the Interactive Presentation
Scenario I
 Character-Centered Approach
l Story is not defined by a script, but by the character‘s
role, personality, status, attitude etc.
© W. Wahlster
Characteristics of the Interactive Presentation
Scenario II
 Open Architecture
l New agents can join at any time
 Auto-Progression
l Story unfolds no matter whether the user actively
participates or not
 Handling of Barge-ins
l Agents may interrupt each other at any time
 Computer-Moderated Dialogue
l Meta-agent makes sure that all agents follow an
agreed-upon interaction protocol
© W. Wahlster
System Architecture for Miau Multi-Party
Dialogue Scenario
Dialog Management
Goal Board Dialogue
Protocol
JIMPRO
Agent
Agent
Server
Handler
Jam
Jam
BDI
Client
BDI
Client
Spin: Templatebased NL Analyzer
© W. Wahlster
Jam
BDI
Client
Multi-Agent Dialogue Control
Is it my turn?
What can I
do for you?
Competence 
Status 
Personality 
...
© W. Wahlster
Layer Model for Multi-Party Conversation
Improvisational Setting
Characters’ Roles, Status, Goals etc.
Domain-Specific Communication
Sales Strategies etc.
Character-Character & User-Character Communication
Multimodal Dialogue Strategies (Speech, Gestures, etc.)
Client-Server Communication
Low-Level Synchronization and Coordination Strategies
© W. Wahlster
From Script-Based Approaches to Interactive
Performances
M e ta p h o r
s crip te d th e a tre
S c rip tin g T im e p rio r to p re s e n ta tio n ,
o fflin e
© W. Wahlster
im p ro vis a tio n a l th e a tre
d u rin g p re s e n ta tio n
o n lin e
S tru c tu rin g
P rin c ip le
p lo t-ce n te re d
c h a ra c te r-ce n te re d
d ra m a tic e le m e n ts
S c rip t
P ro d u c e r
s e p a ra te s ys te m
com ponent
in vo lve d c h a ra c te rs a n d
user
T e c h n ic a l
R e a liza tio n
c e n tra lize d p la n n in g
com ponent
d is trib u te d re a c tive
p la n n e rs
Evaluation of Presentation Teams
Does the number of agents in a presentation team affect
l the user’s ability to decode and retrieve information, and/or
l her affective judgement of the dialogue?
© W. Wahlster
Formulation of Hypotheses
 According to the levels of processing theory [Craik & Lockhart, 72],
the additional spatial and episodic organization of discourse
elements implemented by multiple agents should support the user’s
retrieval performance.
 Based on earlier experiments by [Nass 96], we expect that agents
are rated more competent if specific areas of knowledge are
assigned to them.
 We also expect that the affective rating of the product increases
with the observed competence of the presenters.
© W. Wahlster
Settings for the Scheduled Experiment
P a rtic ip a n ts
A llo c a tio n o f
u tte ra n c e s
© W. Wahlster
M o n o lo g u e
D ia lo g u e
O n e A g e n t:
T w o A g e n ts : T h re e A g e n ts : T h re e A g e n ts :
B u ye r/S e lle r: B
B u ye r B
S e lle r S
B : Q u e stio n s & B : Q u e stio n s
S : A n s w e rs
A n s w e rs
M u lti-P a rty
D ia lo g u e ,
s tru c tu re d
M u lti-P a rty
D ia lo g u e ,
ra n d o m
B u ye r B
S e lle rs:
S E : E co lo g y
S C : C o m fo rt
B u ye r B
2 S e lle rs:
S1, S2
B : Q u e stio n s
S E : A n s w e rs E
S C : A n s w e rs C
B : Q u e stio n s
S1  S2:
A n s w e rs
Dialogue Examples for the Experiment
Monologue
Dialogue
I wonder, how fast this
car might be?
-Oh, I see 200 km/h.
Multi-party
dialogue,
structured
1. Maximum
speed is 200
km/h.
I wonder,
how fast
this car
might be?
2. Built-in servo-steering
keeps the car easy-tohandle.
© W. Wahlster
I wonder, how
fast this car
might be?
Multi-party
dialogue,
random
I wonder,
how fast
this car
might be?
-Maximum speed is
200 km/h.
1. Built-in servo-steering
keeps the car easy-tohandle.
2. Maximum speed is 200
km/h.
Personalized Sales Dialogues with
Presentation Teams in the Miau System
© W. Wahlster
Information Extraction Agents
• Information Filtering
• Information Retrieval
• Information Integration
identify relevant
documents
– ...
– identify and extract relevant
pieces of information
– transform them into canonical form
wrappers
wrappers
• operational descriptions of a target concept
• abstract from concrete occurrence within document
• robust against modifications
© W. Wahlster
The Trainable Information Agents Framework
(Bauer, Dengler)
combination of "classical"
problem-solving methods
and information agents
specifications
results
Application
info requests
User
preferences/heuristics
info
domain ontology
requests
training
planning knowledge
user preferences
Browser
InfoBroker
improved dialog guidance
info requests
PBD dialog
© W. Wahlster
info or script
query
Info Extraction
Trainer
planning, optimization,
Web site annotations
and execution
Overall Architecture
Catalog
Objects
Stereotypes
DB
Interface
Argument
Generator
Result
Abstraction
Strategic
Planner
User Model
Refinement
Ontology
© W. Wahlster
Conflict
Res. Str.
UMs
User Modeling
Mediator
Query
Generator
Interaction Management
Search
Interface
Ontological Reasoning for Decision Support:
Topic Maps
 specification for Topic Maps: XML Topic Maps (XTM) 1.0
 description of complex relationships with associative
knowledge structures
 goal-oriented search and navigation in large data sets
l basic technology for the construction of knowledge structures
l key for knowledge management
 flexibility by separation of maps and information resources
 key concepts: topics, associations, and occurrences
© W. Wahlster
Domain Theory
 description of the general world knowledge
(logical) representation of objects, their properties, and
 the relations among themselves
inference mechanism
l near($sport, $hotel, $city, 0-5 km)
l is_a(“Munich”, city)
l is_part_of(“Munich”,”Bavaria”), is_part_of(“Bavaria”,”Germany”)
 is_part_of(“Munich”,”Germany”)
© W. Wahlster
Ontological Reasoning for
Decision Support: Topic Maps - I
 objective: generating suitable queries for the
l object catalog
l Web information search
in a conflict case
 specifications for Topic Maps: ISO/IEC standard
13250 (1999), XML Topic Maps (XTM) 1.0
 description of complex semantic relationships
with associative knowledge structures
 goal-oriented search and navigation in large data
sets
l basic technology for the construction of knowledge structures
l key for knowledge management
© W. Wahlster
Ontological Reasoning for
Decision Support: Topic Maps - II
 flexibility by separation of topic maps and
information sources:
l application of the same topic on different sources
l application of different topic maps on the same source
 difference between pure XML-documents and topic
maps:
l for application of topic maps no annotation of the available
data necessarily
l XML-document: view of the author on the knowledge
structure
l topic map: view of the user on the knowledge structure and
the associations with the available data
© W. Wahlster
Ontological Reasoning with Topic Maps
Example - I
Interests of user A
10%
30%
60%
Wellness
Culture
Sport
Interests of user B
Sport
© W. Wahlster
 conflict situation:
l user A found a
suitable hotel in the
catalog
l the hotel doesn’t
offer sport
possibilities for user
B
 objective: generation
of a suitable Web
query in order to
receive information
about the surrounding
area of the hotel
Topic Map
sport
sport
offers
offers
sport
sport
club
fitness
fitness
center
center
concentrates on
football
tennis
tennis
diving
teach
diving
diving
certificate
certificate
issues a certificate
diving
school
school
Ontological Reasoning with Topic Maps
Example - II
Topic Map
 situation: user found
different suitable hotels in
the area of:
l Costa Brava
l Costa Blanca
l Costa Dorada
l Cote d’Azur
 objective: generation of a
goal-oriented query for
the object catalog in order
to find additional hotels
© W. Wahlster
Spain
Spain
France
France
Costa
Costa
Brava
Brava
Costa
Costa
Dorada
Dorada
Mediter ranean
ranean
Atlantic
Ocean
Ocean
Ocean
Ocean
Costa
Costa
Blanca
Blanca
Cote
Cote
d’Azur
mild
climate
climate
Ontological Reasoning for
Decision Support: Topic Maps - III
 key concepts:
l topic and topic type
l topic association and association type
l topic occurrence and occurrence role
 additional concepts:
l scope and theme
l facets
 topic map as knowledge
structure
Information pool
© W. Wahlster
Programming by Demonstration - I
wrappers for information extraction from (static) HTML
documents




code repair
canonical representation by HTML parse tree
training dialog with the user
characterization of document units in terms of
(generalized) HTML concepts
 generation of HyQL scripts
© W. Wahlster
Programming by Demonstration - II
new requirements caused by dynamic documents
 DHTML, JavaScript/JScript, ...
 documents only virtually available
 canonical representation by DOM
 flexible characterization of document units
 more powerful extraction wrappers
 more flexible training dialog
© W. Wahlster
Programming by Demonstration - III
 enhanced reasoning capabilities of learning agent +
 DOM representation
 multiple views of document and training task
l HTML objects, semantic concepts, spatial relationships, ...
l flexible user-agent interaction
l user chooses preferred characterization
l agent translates into canonical representation and
l picks "most natural" interaction style with user
 cooperation with MERL (COLLAGEN architecture)
© W. Wahlster
PbD-based Wrapper Construction for Information
Agents
HTML document
dynamic HTML,
JavaScript, CSS,...
pure HTML
PbD Agent
browserspecific
DOM
static HTML treestructure
 Generally, no equivalence between current DOM of browser and
limited DOM of wrapper
 DOM of wrapper is base for appropriate PbD heuristics and data
model for wrappers written in HyQL
 Problems concerning visual and structural correspondence of objects
 Strictly server-based wrapper evaluation
© W. Wahlster
New approach to wrapper construction
HTML document
dynamic HTML,
JavaScript, CSS,...
browserspecific
DOM
PbD Agent
 wrapper component and PbD Agent
become integral part of browser
 direct access to current DOM
Additionally,
 specialized data structures for additional
reasoning, e.g. spatial information of DOM
objects, semantic categories associated with
DOM objects
 server-side processing of wrappers by hosting
the enhanced browser control component
 debugging of wrappers easily supported
General architectural decision:
„A wrapper becomes part of the page which is the resource for its operation“
© W. Wahlster
Sample Wrapper Generation using PbD
H1
$1
$2
Query for all pairs ($1,$2) where
$1:image H1:bold text
$2:simple text
© W. Wahlster
Hypothesis:
• context/style
$1 is image
$2 is simple text
H1 is bold text
<ask for color relevance>
• spatial relation
H1 above $2
$1 left to H1
$1:width << $2:width
• topology
$1 meets H1
H1 meets $2
Apply the same wrapper to a new source
© W. Wahlster
Apply the same wrapper to a new source
(2nd case)
© W. Wahlster
PAN-Video
© W. Wahlster
High Degree of Parallelism of Queries
© W. Wahlster
Knowledge about a Webpage Shared by User
and Agent
common part (usable
for communication)
structural
Learning Annotation
Agent
© W. Wahlster
visual/semantic
Naive
User
procedural
Example - Ontology
Train_Connection [ from =>> Location;
to =>> Location;
travel_date =>> Date;
time =>> Time;
depart_time =>> Time;
arrive_time =>> Time;
cost =>> Price;
travel_duration =>> Duration;
info_url =>> URL;
... ]
© W. Wahlster
Query Planning
 states: information states
l concepts / attributes and instantiations
 operators: querying schemes
l preconditions (´+´) and effects (´-´)
to time arrive_time travel_duration
< +, +, +, +, -, -, -, u, - >
from travel_date depart_time cost info_url
© W. Wahlster
Query Planning
Top
State
Ontology
City.value = München
City.language = German
...
String
...
Operators
CityName1
Language
+
–
+
–
< String, Language, String, Language >
City
value
babelfish
language
" c  pre op  $ i : S 0  i :: c   c
© W. Wahlster
.
.
.
c  S 0  Int ( op )
Query Plan Visualization
Features
l alternative queries
l past states
l future states
l state descriptions
l PBD requests
l accept / reject PBD request
l assessment of plans
l expected completion time
© W. Wahlster
Three Levels of Mark-up Languages for the Web
Content
OIL/M3L
Structure
XML
Form
HTML
WWW Document
Content : Structure : Form = 1 : n : m
© W. Wahlster
M3L Integrates Three Language Families
Frame Languages
Concept Languages/
Terminological Logics
Object-oriented Modelling
Primitives
Formal Semantics
Subsumption, Inferences
M3L
Web Languages
XML and RDF Syntax
© W. Wahlster
RuleML: Ontology Extensions for
Rule Knowledge
 Rules in the Web have become a mainstream topic since
l inference rules were
l marked up for E-Commerce
l identified as a Design Issue of the Semantic Web
l transformation rules were used for document
generation from central XML repository
 Rule interchange is becoming more important in
Knowledge Representation (KR), especially for Intelligent
Agents in E-Commerce
© W. Wahlster
The Rule Markup Initiative
 The Rule Markup Initiative has taken initial steps
towards defining a shared Rule Markup Language
(RuleML) for interoperation between companies
 RuleML permits forward (bottom-up) and backward
(top-down) rules in XML for
l deduction
l rewriting
l further inferential-transformational tasks
© W. Wahlster
From traditional XML Representation to RDF-like
Representation of RuleML Rules
 XML: N-ary, positional representation of rules;
overspecification for non-sequential parts
 RDF: Binary, labeled representation of rules with
nodes for resources and labels as explicit role
names; Seq container needed for sequential parts
 RuleML:
l Sequential parts from XML
l Labeled parts from RDF
© W. Wahlster
Recommender Rule: Forward Markups
Challenge hypertext as one XHTML paragraph:
<p>If you want to review rule principles, you may look at
<a href="http://www.cs.brandeis.edu/...">Rule-Based Systems</a></p>
Original RuleML markup with XHTML in body/head
(English premise and semiformal conclusion):
<imp>
<_body>
premise
<p>You want to review rule principles</p>
</_body>
<_head>
<p>You may look at
<a href="http://www.cs.brandeis.edu/...">Rule-Based Systems</a>
</p>
conclusion
</_head>
</imp>
© W. Wahlster
Recommender Rule: Backward Markup
Further formalized RuleML markup (still unanalyzed English
relation and individual-constant names):
<imp>
<_head>
<atom>
<_opr> <rel>may look at</rel> </_opr>
conclusion
<var>you</var>
<ind href="http://www.cs.brandeis.edu/...">Rule-Based Systems</ind>
</atom>
</_head>
<_body>
<atom>
<_opr> <rel>want to review</rel> </_opr>
<var>you</var>
<ind>rule principles</ind>
premise
</atom>
</_body>
</imp>
© W. Wahlster
An Example coded in RuleML 08
A person owns an object if that person buys the object from a merchant
and the person keeps the object.
imp--------------------------------*
*
head *
body *
*
*
atom-----------------and---------------------------------------*
|
|
|
|
opr *
|
|
|
|
*
|
|
|
|
rel
var
var
atom--------------------------atom-----------------.
.
.
*
|
|
|
*
|
|
.
.
.
opr *
|
|
|
opr *
|
|
.
.
.
*
|
|
|
*
|
|
own person object
rel
var
var
var
rel
var
var
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
buy person merchant object
keep person object
<imp>
<_head>
<atom>
<_opr><rel>own</rel></_opr>
<var>person</var>
<var>object</var>
</atom>
</_head>
<_body>
<!-- explicit 'and' -->
<and>
<atom>
<_opr><rel>buy</rel></_opr>
<var>person</var>
<var>merchant</var>
<var>object</var>
</atom>
<atom>
<_opr><rel>keep</rel></_opr>
<var>person</var>
<var>object</var>
</atom>
</and>
</_body>
</imp>
© W. Wahlster
Two-Way Relationship Between RuleML and RDF
 RuleML in RDF:
l RDF graphs and
serializations for
RuleML rules
l RDF triples as facts
l Exemplified in
l Example: Next slide
previous slides
© W. Wahlster
 RDF in RuleML:
described by a DTD
in the RuleML family
Research on Personalized Interface Agents
brings disparate subfields in the area of
intelligent systems together
Planning
User
Modeling
Knowledge
Representation
Image
Understanding
Personalized
Intelligent
Natural Language
Interface
Web
Agents
Understanding
Services
Machine
Learning
© W. Wahlster
Plan
Recognition
Information
Multimodal Retrieval
User
Interfaces
Conclusion
l The generation of virtual webpages can be achieved by plan-based
internet agents.
l Ontological annotations are needed not only for information extraction
agents but also for presentation agents
l Realization procedures and wrappers form an important part of
the referential semantics of objects on the web
see www.dfki.de/~wahlster/ACAI01/
l ECommerce projects of DFKI have shown that research on personalized
interface agents can be transferred to real world applications:
l Dekra (largest European organization of used car dealers):
FairCar as an ECommerce platform with NL access and a comparison
shopping agent for used cars
l DaimlerChrysler: online user modelling in a one-to-one marketing system
for Mercedes cars
l Porsche: Virtual Market for Pre-owned Porsche Cars
© W. Wahlster
URL of this Presentation:
http://www.dfki.de/~wahlster/ACAI01/
© W. Wahlster
Descargar

Kein Folientitel