@unpublished{alberola:tracking,
    author = {Carlos Alberola and George Cybenko},
    title = {Multiple Hypothesis Text-based Tracking of Land Vehicles},
    year = {1997},
    copyright = {the authors},
    group = {agents, actcomm},
    note = {Submitted to IEEE Intelligent Systems},
    keyword = {information retrieval},
    abstract = {This paper deals with tracking based on natural language
messages. Of particular interest are messages generated by multiple human
observers over extended periods of time. The tracking problem is to associate
messages that are about the same object or phenomenon, create the most
probable tracks based on those associations and finally make inferences about
the situation. A key technical innovation of this work is the use of mature
radar signal processing tracking methods, originally used for correlating
radar sensor reports, in the domain of natural language processing and
understanding for tracking applications. We believe that our results can be
useful in several applications that involve tracking and correlation of
natural language reports based on text semantics. Such applications include:
1) military situation awareness using human observations (such as land
vehicle tracking); 2) maintenance planning using text-based reports; 3)
analysis of computer intrusion detection reports; 4) customer service
monitoring and tracking. \par In this paper we describe our present
implementation, called TextTrack. It uses the multiple hypothesis tracking
(MHT) paradigm developed for multiple object tracking. TextTrack has a simple
natural language frontend which does rudimentary parsing and analysis of
short messages or reports. Messages are then assembled into maximum
likelihood tracks as computed by an MHT algorithm. These tracks can be
analyzed to estimate, for example, an entity's identity. We develop both
Bayesian maximum likelihood and fuzzy logic approaches to the problem. In
addition to presenting the basic ideas behind our system, we also develop
analytic convergence results for one tracking scenario under very weak
assumptions.}
}