December 2, 2004TDT-2004 Adaptive Topic Tracking at Maryland Tamer Elsayed, Douglas W. Oard, David Doermann University of Maryland, College Park Gary Kuhn National Security Agency
Outline Results System design Interpreting the results Next steps
Non-Adaptive Topic Tracking No score normalization Cost= Bottom left is better
Adaptive Topic Tracking No score normalization, unjudged treated as firmly off-topic Cost=0.2438
Adaptive Topic Tracking Cost= One-pass score normalization, unjudged treated as firmly off-topic
Non-Adaptive System Design TDT-5 Evaluation EpochTraining Epoch Compute log-odds ngram weights Compute story scores
Log-Odds Term Weights
Computing Story Scores
Non-Adaptive System Design TDT-5 Evaluation EpochTraining Epoch Compute log-odds ngram weights Compute story scores
Adaptive System Design TDT-4TDT-5 Evaluation EpochTraining Epoch Compute log-odds ngram weights Compute story scores Extended Training Epoch Normalize Story scores Compute Normalization factor
Interpreting Non-Adaptive Results Lack of normalization probably hurt! What can we say about the effect of incomplete judgments?
Interpreting Adaptive Results Normalization hurt! –One-pass design is the problem DET has limitations –Changing the threshold changes our topic model! –Threshold selection is now a critical path item How does judgment density affect the results? Not normalized Normalized
Next Steps Further explore normalization –Implement continuous renormalization –Tune parameters on devtest data Decide between TDT-5 and TDT-4 –Is incomplete judging harmful? Define richer training sets –Explicit queries –Many known on-topic/off-topic training stories –Models of (imperfect) behavioral feedback
Our Favorite Quote of the Day “It takes time to get the implementation correct” [Yiming] We had 30 days from project initiation to non-adaptive submission