evolves as they spread through a corporate network. Cataphora has developed a fuzzy search algorithm to detect them, but Schon admits the task is complex. Creating an algorithm that organizes sentences into text blocks, for example, often forces researchers to make inflexible choices about boundaries, using punctuation, length limits, paragraph breaks, or some other scheme. That, in turn, could cause a program to overlook a document whose author formats things differently, such as not breaking the text into paragraphs very frequently or using unconventional punctuation.

Cataphora has also developed a proprietary set of ontologies that cover human resources-related topics, marketing issues, product development, and more to examine various subject-specific communications. One way in which they are useful, Schon explains, is for studying the relationships between people and topics. If an executive is central to communications about product development, marketing, and finance, but marginal to those about sales, it’s likely that she or he is out of the loop when it comes to the newest sales tactics. Ontologies can also identify communications related to particular tasks, such as hiring and performance reviews. From there, engineers can statistically determine what the “normal” procedure is, and see when it is and isn’t followed. Thanks to the training corpus Cataphora has built over time through its clients, these ontologies perform quite well. Yet to detect communication that is specific to a particular industry, location, or research group and whose names can be idiosyncratic, “we may need to examine the workflow and develop more specific ontologies,” says Schon.

Further analysis helps identify how employees influence each other at work. Aral, for example, correlates his electronically derived network topolo-gies with traditional accounting and project data, such as revenues and completion rates, to try to understand which factors enhance or diminish certain outcomes. “The old paradigm was that each employee had a set of characteristics, like skills or education, which he or she brought to a firm,” Aral explains. “Our perspective is that employees are all connected, and that companies build a network of individ-

microsoft
examines internal
communications to
identify so-called
“super connectors,”
who communicate
frequently with
fellow employees
and share information
and ideas.

uals who help each other.” In a study of five years of data from an executive recruiting firm, Aral found that employees who were more central to the firm’s information flow—who communicated more frequently and with a broader number of people—tended to be more productive. It makes a certain amount of sense. “They received more novel information and could make matches and placements more quickly,” Aral notes. In fact, the value of novel information turned out to be quite high. Workers who encountered just 10 novel words more than the average worker were associated with an additional $70 in monthly revenue.

Yet Aral’s conclusions also point to one of the more challenging aspects of this type of research. If a position in the corporate network is associated with increased productivity, is it because of the nature of that position or because certain kinds of people naturally gravitate toward it? “You always have to question your assumptions,” admits Aral. New statistical techniques are needed, he says, to more accurately distinguish correlation from causation.

Large-scale data mining presents
another challenge. IBM’s SmallBlue,
which grew out of research at its Wat-
son Business Center, analyzes employ-
ees’ electronic data and creates a net-
worked map of who they’re connected
to and where their expertise lies. Em-
ployees can then search for people
with expertise on certain subjects and
find the shortest “social path” it would
take to connect them. SmallBlue is an
invaluable tool for large, international
firms, and IBM has used it to connect
its 410,000 employees since 2007. Yet
indexing the 20-plus million emails
and instant messages those employees
write is not a trivial task—not to men-
tion the 2 million blog and database
entries and 10 million pieces of data
that come from knowledge sharing
and learning activities. It is the largest
publicly known social network dataset
in existence, and the project’s founder,
Ching-Yung Lin, says IBM worked hard
to design a database that would hold
different types of data and dynamically
index the graphs that are generated.

Further Reading

Aral, S., Brynjolfsson, E., and van Alstyne, M. Information, technology and information worker productivity. International Conference on Information Systems, Milwaukee, WI, 2006. Manning, C.D., raghavan, P., and Schütze, H. Introduction to Information Retrieval. Cambridge University Press, new York, 2008. Mikawa, S., Cunnington, S., and Gaskis, S. Removing barriers to trust in distributed teams: understanding cultural differences and strengthening social ties. International Workshop on Intercultural Collaboration, Palo Alto, CA, 2009.

Wasserman, S. and Faust, K. Social Network Analysis: Methods and Applications. Cambridge University Press, Cambridge, 1994.

Wu, L., Lin, C.-Y., Aral, S., and Brynjolfsson, E. Value of social network: a large-scale analysis on network structure impact to financial revenue of information technology consultants. Winter Information Systems Conference, Salt Lake City, UT, 2009.

Leah hoffmann is a brooklyn, ny-based technology writer.