...
- TREC Genomics collection is a full-text collection (different from bioCaddie which is descriptive metadata collection). It consists of full-text HTML documents from 49 journals published via Highwire Press. Hence, each document's text is much longer.
- Topics used in TREC Genomics collection are common queries and quite similar to bioCaddie original queries.
Eg:
<200>What serum [PROTEINS] change expression in association with high disease activity in lupus?
<201>What [MUTATIONS] in the Raf gene are associated with cancer? - Relevant judgements contain judgements for different passages of a document (RELEVANT or NON-RELEVANT). Some documents can be divided into multiple passages of different length and can have different judgement for each passage. However, in our baselines run, we use the judgement for the whole document; hence if a document has one or more relevant passages, it is considered RELEVANT.
Eg:
200 10090921 10160 2221 RELEVANT
200 10090921 12404 720 NOT_RELEVANT
200 10090921 13147 1084 NOT_RELEVANT
200 10090921 59180 515 RELEVANT
200 10090921 101717 349 RELEVANT
Document number 10090921 is considered RELEVANT as it has at least 1 RELEVANT passage. Baselines run results:
2007:
- Okapi significantly performed worse than tfidf in all metrics (no queries failed to run)
- Query Likelihood baselines did not show significant improvements compared to TFIDF (few metrics were even worse)
- RM3 did yield better results but the data did not provide a significant improvement compared to TFIDF baselines.
- MAP results are above mean and median of official results of TREC 2007 Genomics Track (whose MAP with mean 0.1862, median 0.1897)
2006:
- Okapi significantly performed worse than tfidf in most of the metrics (but not such huge difference like 2007).
- Query Likelihood baselines did show significant improvements compared to TFIDF for NDCG@20 and NDCG@100.
- RM3 provided a significant improvement compared to TFIDF baselines (MAP, P@100, NDCG@20, NDCG@100)
- MAP results are lower than mean and median of official results of TREC 2006 Genomics Track (whose MAP with mean 0.2887, median 0.3083)
Verify the methods with high MAP metric from official results.
UICGenRun1,2,3(~0.52-0.54): two dimensional ranking, query expansion.
NLMinter (~0.47) refined queries using combination of topic terms, query expansion (using Entrez Gene, GeneCard, MeSH)
THU1,2,3 (~0.43): best result, shorter passages return, longer passage return
iitx1,2,3 (~0.41-0.42): customized function
UIUCinter (~0.41): interactive run
PCPsgRescore,Clean,Aspect (~0.42): Combine multiple types of resources for constructing queries; Hierarchical language model smoothing; Post result filter
uchsc2,1,3 (~0.40 - 0.41): Expanded queries are sent to the search engine Lemur. Results undergo zone filtering, and top remaining Lemur results are sent to a singular value decomposition algorithm to expand the results pool by selecting similar paragraphs based on a latent semantic Dirichlet similarity score. Results of the SVD are filtered using Naive Bayes with lexical and conceptual features with training data dervied from manual evaluation of Lemer output
NLMfusion (~0.37): equally-weight fusion of the results of 4 automatic methods.
UniNE1,2,3 (~0.34-0.36): Data fusion of two IR systems (based on normalized RSV values (Z-score), baserun for comparisons) IR system
UofG0,1 (~0.35-0.36): Retrieval based on the language modeling approach. The results are further filtered based on document coverage.
Compare our results with similar methods used in official runs.Method 2006 Official Run Our run tfidf weighting 0.2755 (UniGe - a vector-space with tf.idf weightings and a modified version of pivoted normalization)0.129 (kyoto2 - probabilistic model of term occurrence)0.2714 Language Modelling 0.3459 (EMCUT1,2 - Document retrieval is performed using a language-modelling approach. Passage selection is based on identification of concepts from the UMLS metathesaurus and a gene thesaurus in both the query and the documents)
0.3517 (UofG0 - Retrieval based on the language modeling approach)0.2774 (JM)0.2939 (Dir)0.2884 (TS)Riochio relevance feedback 0.3634 (DUTgen1, DUTgen2 - using Rochio relevance feedback based on 2005's gold standard, Two levels of indexes, BM25, reranking)0.2964 (UMassCIIR1 - Query-biased pseudo relevance feedback. 250 word passages with overlap removed)0.3415 (RM3) Okapi 0.3365 (york06ga1 - 1. Use Okapi BM25 for concept-based structured query 2. Use the blind feedback with term selection technique 3. Use a dual index model for passage retrieval 4. No aspect-level retrieval) 0.2372