The paper proposes a framework for improving organizational performance by actively analyzing face‑to‑face interactions using unobtrusive wearable sensors (MIThril and the Sociometer). It outlines how sensor streams can capture four categories of information—interaction features (speech prosody, body language), participant identity, context (location, calendar data, peripheral video), and conversational content (noisy keywords). These data are aggregated and modeled with a tractable probabilistic “influence model,” a simplified coupled hidden Markov model that estimates how individuals influence each other's turn‑taking dynamics, providing a proxy for information flow without full speech understanding. To enrich the low‑level signals, the authors integrate commonsense knowledge from OMCSNet, inferring fine‑grained conversation topics (gists) via probabilistic linking of keywords to semantic relationships. Experiments show that incorporating contextual cues (e.g., location) markedly improves gist identification. The paper then sketches several applications: automatic expert and collaborator discovery, real‑time collaboration tools, and ad‑hoc conversation patching that can route knowledgeable participants into ongoing discussions. Finally, privacy concerns are addressed through user‑controlled data posting, delete/mute buttons, and situational trade‑offs. The authors argue that such a “social intelligence” layer can transform static org‑charts into dynamic, data‑driven representations of human capital.
Topics
Vision of Social Intelligence via Wearable Sensing
The overarching proposal to embed unobtrusive wearable devices in the workplace to capture and analyze face‑to‑face interactions for organizational benefit.
MIThril and Sociometer: Hardware Platforms for Interaction Capture
Description of the modular PDA‑based MIThril system and the badge‑like Sociometer that record audio, acceleration, IR proximity, and other sensor streams.
Four Pillars of Interaction Data: Features, Participants, Context, Content
Breakdown of the types of information extracted from sensor streams—prosodic and body‑language features, participant identification via mutual audio information, location/context via IR/RF/Bluetooth and calendar cues, and noisy keyword extraction for content.
Influence Model: Tractable Probabilistic Framework for Dynamics
A simplified coupled hidden Markov model that uses pairwise influence parameters (α) to predict state transitions and infer directionality of information flow from turn‑taking dynamics.
Commonsense Reasoning with OMCSNet to Infer Conversation Gists
Use of the Open Mind Common Sense semantic network to map noisy keywords and contextual cues to fine‑grained topic gists via probabilistic linking and ranking.
Application Scenarios: Expert Locator, Collaboration Tools, Ad‑hoc Patching
Illustrations of how the sensed and modeled data can power automatic expertise profiling, real‑time knowledge‑base queries, and dynamic insertion of relevant participants into ongoing discussions.
Privacy‑Preserving Mechanisms and Trade‑offs
Discussion of user‑controlled data posting, delete/mute buttons, and situational contexts where privacy concerns may be outweighed by operational needs.