MAS837: Collaborative Web Browsing Experiment

In this experiment, I played the role of the Assistant. I was with Lee Campbell, who was the Observer, and Andy Dahley, who was the Searcher. We successfully found answers (or near-answers) to five of the questions (third through seventh); searched for, but were unable to find, answers to the first two; and did not have time to search at all for the last two.

Responses to Questions for the Assistant

How hard was it to provide assistance?

At first, it was difficult to be of much help. The roles of Assistant and Searcher were not clearly defined; I did not know what kinds of help the Searcher expected, or needed; and it was difficult to determine when to observe, and when to take the initiative by making suggestions. I felt that I was more in the way than of help. As the experiment progressed, however, I gained knowledge of what the Searcher wanted, by several means: (1) explicit verbal commands, (2) observation of the Searcher's actions, (3) positive and negative feedback regarding my actions and suggestions, and (4) induction applied to the information gained from the previous three methods. By the end of the experiment we were collaborating relatively efficiently.

Could the searcher have provided any kind of input or advice that would have helped you give assistance?

Of course. The Searcher could have explicitly stated his preferences to me, in regards to search strategy, when he wanted help, and what types of help he wanted. Also, explicit feedback on each one of my actions would have been quite useful. An Assistant is only as good as its knowledge of its user (and its domain); and direct instruction is the quickest means of acquiring that knowledge.

Could you have done a better job giving assistance if you had more time to work on the question yourself?

Yes. By browsing beforehand and in parallel with the Searcher, I could have done two things: (1) explore adjacent routes in the current search path, say by browsing unselected but promising links on the page being viewed and recommending them to the Searcher, if they prove interesting; and (2) attempt completely different search strategies (e.g., subject index vs. search engine), and suggesting switching strategies if they seem to be more likely to give a result than the current strategy.

Would you have an easier time giving assistance to the same person next time as a result of your experience?

Yes. By building on my knowledge of the Searcher's preferences and strategies, the learning period discussed in the first question could be avoided, and we could begin at a higher level of performance. Also, I gained some domain knowledge (discussed below) which would help me assist the Searcher in similar tasks.

Did you learn anything from assisting this person that you could apply if you were assisting someone else?

In the session, I was able to gain some domain knowledge (new search options for Alta Vista, etc.) so that if I were to assist another in browsing, I could make better recommendations to him/her. As far as applying the knowledge of how to assist my Searcher to another user, it would seem that most of it would be entirely subjective and would have to be relearned for another user.

How much of your assistance could have been provided by a computer agent (assuming current technology)?

Some, but not all, of my assistance could have been replicated by current technology. A rough division can be fashioned between assistance based on knowledge of the structure of the search space and the Searcher's event stream, and the content of the search space. The former kind of assistance could be realized today, roughly by using a base of domain-specific knowledge (how to use search engines, lists of "meta-indices", etc.), a model of when to provide assistance and how, and suitable means of utilizing and acting on that information. Meanwhile, the latter kind of assistance can only be realized in a limited way, given the current state of natural language understanding. In our experiment, assistance was given at both levels, with both kinds being of help.

Also, much of our interaction assumed full speech generation and understanding by the Assistant. Both of these capabilities are beyond our reach today; more realistic would be the use of artificial (i.e. GUI) interfaces, or more restricted languages.

If the Searcher were another person (but not in the same room), how could a computer have facilitated interaction between you?

Two possibilities come to mind. One is to try to replicate the scenario of the experiment, by having the Assistant monitor the actions of the Searcher remotely, and receiving commands and offering suggestions by two-way communication. It would be useful in this setup not only to be able to see the Searcher's browser window, but the Searcher himself; body language and facial expression offer valuable clues as to the level of success or frustration the Searcher is experiencing.

But the advantages of parallel browsing I outlined above suggest an alternative. The Searcher and Assistant could both browse simultaneously, and each could see the other's window as well as their own. The Assistant could recommend links merely by pointing at them in the Searcher's window; and if the Assitant's path seems more promising than the Searcher's, he could swap control of the two windows and take over from there. Of course, there would be additional means of communication for discussing strategy and other matters.