(Slide 1)

User Modelling in Expert Man-Machine Interfaces:
 A Case Study in Intelligent Information Retrieval
 Georgio Brajnik, Giovanni Guida and Cario Tasso
 Presented by: Dawn Chappelle Jones

Introduction

UM-Tool is designed to aid users in an environment such as on-line
searching. They do this by tailoring the system to each user and by
doing it in a non-intrusive way.

(Slide 2)

An expert interface supports co-operation between man (user) and
machine (target system).

(pic1.gif)

Providing:
  support
  assistance
  training 

They first give their definition of an expert interface, which is an
intelligent intermediary that can support co-operation between man
(the user) and machine (target system) in an interactive
problem-solving environment.

The benefits of an expert system would be:

(SLIDE 3)

Benefits of an Expert Interface
  Extend usability
  Improve:
   - quality of interaction
   - performance from machine
   - degree of satisfaction from user

Their main criteria, however, is that the expert system must be
capable of modeling and remodeling dynamically in a non-intrusive way
the users it is interacting with.

In their study they focus on the issue of modeling user
characteristics rather than the issue of investigating the mental
models of the users.

(SLIDE 4)

The Behaviourist View:
  1. Exploited user stereotypes
  2. Poorly organized stereotypes
  3. Poor user classification
  4. User modeling component interaction

In other words, they focus on the behaviourist view rather than the
cognitive view, paying particular attention to the following: It is
their belief that stereotypes can be used to reclassify users, as well
as initially classifying them.

 They also use a different classification system for stereotypes that
we will get into later.  Classes also overlap so that users can fit
into more than one class.  It's also not a stand-alone system - it
interacts with all other parts of the system.

They are experimenting with a prototype tool called UM-Tool (User
Modeling Tool) for creting, maintaining, and using individual and
explicit user models within an expert interface in the domain of
on-line retrieval.  They chose this domain because it isn't brainless
work - there is a need for a system that can assist users in searching
well and cost-effectively (GTEL, for example).

(SLIDE 5)

Why We Need an Intermediary
  
 Dialog Database       			$/hr
  
 CHEMSEARCH       			$354
 Derwent World Patents Index     	$210
 Paperchem        			$168
 Current Biotechnology Abstracts    	$162
 PTS PROMT       			$126
 ERIC         				$ 38
 
Their approach to the design of a good user model sets them apart:

(SLIDE 6)

Setting UM-Tool Apart: Approach
 
 What?
  Hierarchy of stereotypes
  Information about the current user
 How?
  Generically useable knowledge
 What next?
  General framework

Concept and Benefits of a User Model

They describe a model as an abstraction of reality that lets you focus
on relevant characteristics.  So a user model is one that represents
characterstics of a user that are relevant to the goal of supporting
an effective and graceful man-machine interaction.

They classify user models in the following way:

(SLIDE 7)

Concept:
 - Implicit vs. explicit
 -Given vs. inferred
 -Static vs. dynamic
 -Canonical vs. individual

The benefits of such a system are:

Benefits:
 -Economy of interaction
 -User acceptability
 -Effectiveness/efficiency of use of target system

In other words, the user has to do less, the user likes it, and the
user likes his bill at the end.

Architecture of an Expert Interface

The architecture of a user modeling system consists of:

(SLIDE 8)

(pic2.gif)

The dialog manager which works between the user and the interface, a
support problem solver which simulates the user and works with the
target system and the user modeler which doesn't have a direct link
with either the user OR the target system, but manages the user
models.

The purpose of the user modeling system is to gather information about
the user, either single items or clusters of items - both linguistic
or conceptual. It does this externally by dialog inspection and direct
questioning or internally by inference or re trospection (this is
Fig. 2, page 170).

UM-Tool

The UM-Tool is an experimental system which is a generic, application
independant tool for user modeling. The things that they took into
consideration when designing UM-Tool were:

(SLIDE 9)

Design standpoints:
 1. Classifiability of user population
 2. In sufficiency of the classification criterion
 3. Stability of the user population
 4. Incompleteness of user information
 5. Incrementatility of user modeling
 6. Incompleteness of user models
 7. Plausibility of user models
 8. Nonmonotonicity of user modeling

Hence:

(SLIDE 10)

Their concept is:

 -Explicit
 -Inferred
 -Dynamic
 -Individual

The main goal of the UM-Tool is the construction, maintenance and
exploitation of uer models, specialized data structures devoted to
storing information about individual users currently accessing an
expert interface.

(SLIDE 11)

The Model Manager:
-Responsible for:
 -interaction
 -classification
 -operations
-Utilizes:
 -stereotype knowledge database
 -user model database
 -session history database

See Fig. 5, p. 172.

Knowledge Representation

(SLIDE 12)

(pic3.gif)

The UM-Tool gathers information about users in a variety of ways -
aquisition, inference, retrospection or inconsistancy (172-3). It
takes these values either structured or atomic, and put them into
slots in the general framework of a stereotype.

The generic stereotype is shared by the entire user population.  Class
stereotypes are specific classes of users. As user may belong to
several classes and be described by more than one stereotype.

(SLIDE 13)

The Modeling Process

1. Acquisiton of information abou the user
 -select acquisition method
 -execute selected method

2. Insertion of information into user model
 -check consistancy
 -resolve inconsistancies
 -write value into slot
 -execute inference methods

3. Determination of active stereotypes

An expirimental expert interface called IR NLI which assists
non-technical users in the access to on-line bibliographic systems in
the field of computer science.  

(SLIDE 14)

Architecture if IR-NLI System

-User provides statement
-IRES: 
 -analyzes
 -gathers details for comprehension
-Databases are selected
-Suitable approach is taken
-Appropriate tactics are adopted
-Search strategy is produced
-User:
 -sees retrieved documents
 -evaluates
-If user is unsatisfied, search is repeated

The architecture of this system mirrors the general organization of an
expert interface in Fig 1, p. 169.

(SLIDE 15)

Evaluation

Advantages:
 -ease of modification
 -hierarchy of stereotypes
 -overlapping stereotypes
Problems:
 -definition of stereotypes
 -inconsistancy
 -efficiency

Top-down is the best way to figure out how to represent knowledge
about the domain you are working with. It's also the natural way to
work within a hierarchy.  You also have to be careful when you pick
stereotypes because everything depends on them. If you pick them
wrong, the whole system doesnt' work.

(SLIDE 16)

Related work
 Grundy
 Real Estate Agent
 Monstrat
 I3R
 Impact

The authors feel that none of these are as good as the UM-Tool becuse
they either don't care about deactivation of stereotypes, the user
models aren't saved, there is no model revision, they don't _use_
stereotypes, or there are no ways to resolve conflicts.

(SLIDE 17)

Discussion topics:

1. Does it make sense to develop these systems when there are people
that can carry out the function hjust as well (or better!) in the same
amount of time?

2. Could they have picked a better set of design principles that could
do the job more efficiently?

3. Have they really addressed what they consider to be the benefits of
an expert interface.

4. How well can this be scaled up? Would it be efficient for a large group of users (eg. GTEL)?


Dawn Chappelle Jones
Georgia Institute of Technology

gt9167a@prism.gatech.edu