Integration of document detection and information extraction

We have conducted a number of experiments to evaluate various modes of building an integrated detection/extraction system. The experiments were performed using SMART system as baseline. The goal was to determine if advanced information extraction methods can improve recall and precision of document detection. We identified the following two modes of integration:I. Extraction to Detection: broad-coverage extraction1. Extraction step: identify concepts for indexing2. Detection step 1: low recall, high initial precision3. Detection step 2: automatic relevance feedback using top N retrieved documents to regain recall.II. Detection to Extraction: query-specific extraction1. Detection step 1: high recall, low precision run2. Extraction step: learn concept(s) from query and retrieved subcollection3. Detection step 2: re-rank the subcollection to increase precisionOur integration effort concentrated on mode I, and the following issues:1. use of shallow but fast NLP for phrase extractions and disambiguation in place of a full syntactic parser2. use existing MUC-6 extraction capabilities to index a retrieval collection3. mixed Boolean/soft match retrieval model4. create a Universal Spotter algorithm for learning arbitrary concepts