"... In PAGE 4: ...4 Model Summary We have identified mixture models that can be used at three stages in the retrieval process to infer parsimonious language models. Table2 shows the models used in these three stages and the relation between them [?]. In each case in this table the base model is externally defined (in- dependent of the object under consideration).... In PAGE 4: ... In initial query formulation, we fit the RM(GM) to the only information that we have about the request, namely its original text. In [?], it was assumed that the search process would involve trying out the level 3 model on every document (shown as 3a in Table2 ), against a null hypothesis which would be the level 2 model derived at indexing time. This comparison would again have to appeal to parsimony.... ..."

"... In PAGE 4: ...4 Model Summary We have identified mixture models that can be used at three stages in the retrieval process to infer parsimonious language models. Table2 shows the models used in these three stages and the relation between them [24]. In each case in this table the base model is externally defined (in- dependent of the object under consideration).... In PAGE 4: ... In initial query formulation, we fit the RM(GM) to the only information that we have about the request, namely its original text. In [24], it was assumed that the search process would involve trying out the level 3 model on every document (shown as 3a in Table2 ), against a null hypothesis which would be the level 2 model derived at indexing time. This comparison would again have to appeal to parsimony.... ..."

"... In PAGE 6: ... 6.2 Results Table2 displays the storage imbalance in terms of the number of postings and inverted lists for the two partitioning schemes with varying number of index servers, K =2, 8, 32. This table also shows the total number of disk accesses,... ..."

"... In PAGE 7: ...3.2 Comparison and Combination with Relevance Models In Table5 we compare the retrieval results of the LBDM with the relevance model (RM), which incorporates pseudo-feedback information and is known for excellent performance (Lavrenko and Croft, 2001). On some collections, the results of the two models are quite close.... In PAGE 7: ... Although we tune parameters on the AP collection, parameter adjustment for the other collections does not improve the performance much. Compared to the relevance model results in Table5 , we conjecture that it is due to the characteristics of the documents and queries that the improvement on the AP collection is larger than on the other collections. We can also combine the relevance model and LBDM to do retrieval.... In PAGE 7: ... More importantly, unlike the Relevance Model, LDA estimation is done offline and only needs to be done once. Therefore LDA-based retrieval can potentially be used in applications where pseudo-relevance feedback would not be 3 The QL amp;RM baseline in Table 6 is slightly different with Table 5 because in the experiments of Table5 , in order to compare with the results in Liu and Croft (2004), we directly load their index into our system and then run the experiments on their ... ..."

"... In PAGE 2: ...Table 1: MAP and precision at 10 for the runs, the improvement is relative to the document-based run The first three rows of Table1 show the results of three runs, each based on a different source of evidence for rele- vancy. The document-based run is quite competitive, but the element-based and the environment-based runs give sig- nificantly better results.... In PAGE 2: ... We simply rank elements by adding up the scores from each run. The result is shown in the last row of Table1 . The mixed run performs between 3.... ..."

"... In PAGE 14: ... This ability to retrieve relevant information based on context or meaning rather than literal term usage is the main motivation for using LSI. Table4 lists the LSI-ranked documents (medical topics) with dif- ferent numbers of factors (k). The documents returned in Table 4 satisfy a cosine threshold of :40, i.... In PAGE 15: ...orderingonly as Table4 clearly demonstratesthat its value associated with returned documents can significantly vary with changes in the number of factors k. Table 4: Returned documents based on different numbers of LSI factors.... ..."

"... In PAGE 18: ... We ran this experiment for several minimum page sizes, ranging from 250 bytes to 4,000 bytes. Results are shown in Table3 . The second column shows the number of indexed fragments; while the subsequent columns show the precision at the indicated volume of answers.... In PAGE 18: ... The second column shows the number of indexed fragments; while the subsequent columns show the precision at the indicated volume of answers. Because the number of known relevant documents for these queries is as little as one, it is impossible to achieve 100% precision even at k = 5; and the nal row of Table3 shows the maximum average precision value that could be achieved for each value of k. The entries in the table should be considered relative to this maximum.... ..."

"... In PAGE 5: ... Table2 : Performance of our 4 runs Retrieval performance was superior with Wikipedia-based query expansion than with- out. Adding Wikipedia-based expansion im- proved performance in 30/45 queries under the small document model and 34/45 queries under the large document model.... ..."