Our Secret Weapon: The Secretary

1
Our Secret WeaponThe Secretary
Ryan Shaun Baker Human-Computer Interaction
Institute Carnegie Mellon University Pittsburgh,
PA USA rsbaker_at_ cmu.edu http//www.cs.cmu.edu/rsb
aker
2
Situationally Appropriate Interfaces

• This is a huge area
• I will focus on one sub-area
• When do we allow full-scale interruptions?
• Not just a little ignorable hint at the periphery
• But a full scale STOP and pay attention!

3
Interrupting The User

• Sometimes well want to interrupt our user
• Sometimes we wont
• How does our system decide?

4
A model secretaries

• Secretaries already deal with deciding whether or
  not to allow interruptions
• Take this example of a deans secretary denying
  an interruption

5
An intruder

• AA Sorry, do you have an appointment?
• I Oh, no, do I need one?
• AAYes, you do.
• II cant see him right now?
• AANo, you cant.
• IAll right, when can I see him?
• AAUm, when are you available?
• IUm, I have a, how about tomorrow?
• AAOk, tomorrow after noon? 2 oclock?
• IUm, sure. Okay.
• AAOkay.

6
Analysis

• In this situation, the secretary
• Prevented an unwanted interruption
• Scheduled an appropriate time for the business to
  take place

7
Subtle cues about the user

• Theres a lot of value to studying subtle cues
  about what the user is doing
• Heart rate, skin electric valence, pattern of
  speech, gaze, attention, and so on
• But secretaries make a lot of decisions to allow
  and disallow access without this information
• Often they cant even see the user

8
Sophisticated Models

• Secretaries have sophisticated models for when to
  allow interruptions (phone and face-to-face)
• These models are formed in part by explicit
  instructions
• These models also include reasoning which is not
  based on explicit instructions

9
Work Model

• Lets say we could perfectly sense when a user
  was in a meeting with one other person, with each
  talking an equal amount
• This is not enough to be able to decide if we
  should interrupt
• We need some sense of the importance of the
  meeting, and the importance of the interruptor

10
Developing a Model of Importance

• Give diaries to several secretaries who
• Control face-to-face access to a faculty member
• Control phone access to a faculty member
• Ask them to record
• every attempted interruption
• whether the person was granted access if not,
  what action was taken
• every time they were explicitly asked by the
  faculty member to give or deny access
• if they were not going from explicit
  instructions, why they chose to allow or deny
  access

11
Developing a Model of Interruptions

• Why Diaries?
• Because Contextual Inquiries are unlikely to
  generate enough cases in any reasonable amount of
  time
• Because Retrospective Interviews will inherently
  miss a lot of cases, and may be more focused on
  memorable cases than everyday ones

12
Developing a Model of Interruptions

• Also, we will give diaries to a couple faculty
  members and ask them for one day
• Every half hour while they are at the university
• Write down a one-phrase description of what
  activity they are currently engaged in (meeting
  with the dean, teaching class, reading email)

13
Developing a Model of Interruptions

• This will give us a list of the situations these
  faculty members find themselves in
• We dont (at this point) want their reports on
  whether or not they think they would be
  interruptible, because they would be
  differentially interruptible by different people.
  (making them go through a list would probably be
  overkill)

14
Developing a Model of Interruptions

• Given this, develop a system which will
• Support all of the explicit requests our faculty
  members use
• Will have categories for different types of
  interruptors and situations, based on our data
• Will have rules for when an interruption is
  acceptable (in the absence of explicit
  information, and even sometimes going against
  explicit information)

15
Learning

• We will use develop this system using an initial
  model of the situations and categories of
  interruptors, and then taking a set of training
  data (again, from diaries) on when interruptions
  were accepted and not accepted
• Then we will run a learning algorithm (FOCL?) to
  develop a set of rules for when interruptions
  should and should not be allowed