An interesting paper in that it seeks to prove that NED approaches that rely on exploiting existing tracking technology will invariably exhibit poor performance. Systems that use tracking technology for NED follow the mantra - every time a new topic is found and tracked by a topic tracking system, it is equivalent to finding a new event. So, the NED system is only as good as the tracking system it is built on.

Given tracking error rates, the lower and upper bounds on NED error rates were derived mathematically. These values were found to be good approximations of the true NED system error rates. Thus, given any tracking system, the performance of a NED system built using it could be quite accurately predicted. However the error rates of NED systems obtained from even the best tracking systems’ error rates were not acceptable.

Since tracking and filtering using full-text similarity comparison approaches are not likely to make the sort of improvements that are necessary for high-quality NED results, the paper concludes that an alternate approach to NED is required.

-- GiridharKumaran - 26 Mar 2003
to top