This is the third post by Jeremy and me in a series on collaborative information seeking. The first was an introduction to the space, and the second dealt with the topic of collaborative intent. This post deals with synchronization of data that underlies the collaboration. While it is possible to collaborate in searching for information without tool support by exchanging URLs or documents directly, more interesting interactions are possible when they are mediated by the search system.
When we first started developing our model of collaborative information seeking we groped around for a way to express the flow of influence among collaborators. We argue that a key aspect of collaboration (as opposed to recommendation, for example) is the ability of the system to make collaborators aware of each others’ actions (and their results) as those actions took place. Whereas in a recommendation system influence flows in one direction—from the “crowd” to the incremental user—in a collaborative information seeking system, influence flows between any collaborators, as appropriate to their roles.
Initially, we labeled this dimension “synchronicity”: two (or more) people working together are made aware of each other’s actions instantly. This is also the terminology that Colum Foley uses: Synchronous Collaborative Information Retrieval (SCIR). But this was confusing because it suggested that people have to work at the exact same time, in lock step, for this approach to be useful.
But synchronized information seeking is not like synchronized swimming. Rather than forcing everyone to repeat the same actions, data synchronization allows collaborators to be aware of what their partners have already done (without having to repeat it) and to incorporate those results into their activities. For example, relevance feedback provided by one person can cause the system to make term suggestions for query expansion to both collaborators.
It is not important, however, to have temporal synchronicity to have data synchronization. Two people can be searching collaboratively even if they take turns interacting with the system, as long as the system makes side effects of one person’s actions available to other collaborators. Without data synchronization it is not possible to have division of labor or the sharing of knowledge.
We also considered “symmetry” as a candidate label for this dimension, but it was difficult to account usefully for the difference between asymmetric roles that happen not to need data from all collaborators and asymmetric systems that preclude the possibility of omni-directional data flow. We wanted to have a clear distinction between systems that made it possible for all participants to influence each other from systems where this influence was fundamentally constrained by the algorithms.
In the end, we settled on “synchronization” as the name for this dimension. Collaboration can either be synchronized or non-synchronized. This characterization seems to capture the flow of influence that is the enabler of division of labor and the sharing of knowledge without constraining us to specific system architectures or role configurations.
When a collaborative information seeking activity is synchronized, there is an implication that the activities of the searchers “co-exist” or are otherwise directly coordinated. Synchronized collaborative information seeking implies that influence (searcher activity data) is immediately available to all participating collaborators, but depending on their workflow they may or may not act on that information the moment it arrives. Influence is synchronized; searcher activity is not.
We have a few examples of synchronized collaborative information seeking: SearchTogether implements synchronization by allowing collaborators to share search results and relevance judgments. Colum Foley’s SCIR system passes relevance judgements back and forth (omni-directionally) between collaborating searchers. Our system, Cerchiamo presents term suggestions derived from relevance feedback of one person to the other, and shows integrated search results to both users. Recommender systems, on the other hand, do not exhibit data synchronization among all participants; they are non-synchronous. Instead, data is aggregated over some large group of people, and then this aggregate is made available to the person receiving the recommendation. However, that person’s actions (whether confirming, ignoring, or rejecting the recommendation) are not fed back to the people whose prior activities informed the recommendation. Thus influence flows only in one direction, and we say that such system is not collaborative, but rather an example of social search.