JHU/APL at TREC 2005: QA Retrieval and Robust Tracks

The Johns Hopkins University Applied Physics Laboratory (JHU/APL) focused on the Robust and Question Answering Information Retrieval (QAIR) Tracks at the 2005 TREC conference. For the Robust Track, we attempted to use the difference in retrieval scores between the top retrieved and the 100th document to predict performance; the result was not competitive. For QAIR, we augmented each query with terms that appeared frequently in documents that contained answers to questions from previous question sets; the results showed modest gains from the technique. HAIRCUT The Hopkins Automated Information Retriever for Combing Unstructured Text (HAIRCUT) (3) is a document retrieval system developed at the Applied Physics Laboratory. It uses a traditional inverted index, a unigram language model for its similarity metric (2, 4), and a flexible tokenizer. The tokenizer supports words, stems, character n-grams, word n-grams and phrases. We have focused on language-independent techniques in developing HAIRCUT. It has been evaluated in TREC, CLEF and NTCIR in at least sixteen languages, and routinely performs among the top systems for both monolingual and translingual ad hoc retrieval. Question Answering Information Retrieval Track