"... In PAGE 3: ... In the keyword-based mode, the original query is used directly to calculate the relevance of query and document. The values of precision of ontology-based mode and key- word-based mode for the three IR models were calculated and shown in Table2 . In both retrieval modes, TMM has the best performance than the other IR models with the value of preci- sion in keyword-based mode and ontology-based mode 0.... ..."

"... In PAGE 9: ...able 3 Recall/Precision/F score for Two Search Techniques .......................................95 Table4... In PAGE 111: ...Generic 91 19 379 96 89 8 93 30 210 Specific 92 71 30 85 88 -3 87 78 12 Context 91 78 17 84 26 223 86 39 121 Overall 91 56 63 88 68 20 89 49 81 In Table4 , the data for average precision, recall, and F score for each type of query has been reported. We have also reported on the effectiveness of our ontology-based model over keyword based search by measuring the difference in scores between these two for recall, precision, and F score in each of the query types.... ..."

"... In PAGE 9: ...64. Table7 shows the performance of our approach.... ..."

"... In PAGE 25: ...x. Table3 lists the reference counts (recall rate) and the number of ... ..."

"... In PAGE 4: ... The weights of words are calculated by using a keyword weight- ing strategy of having specialized in partial XML document retrieval. Table2 shows a result of analyzing the XML docu- ment shown in Figure 1 using the content analyzer. If we use this analysis, we can retrieve partial XML documents related to a keyword-based query based on the vector space model because we can generate an inverted file for partial XML document retrieval.... ..."

"... In PAGE 2: ... Twenty percent of the training data is set aside for measuring training progress. The first two rows of Table4 provide word error rates for the baseline models and a set of models trained on train 12000 us- ing the above algorithm. The discriminative training produces relative error rate reductions of BIBMBJB1 for test 500 and BEBMBJB1 for test 2500.... In PAGE 4: ...e.g., function words). If our hypothesis is correct, we expect the overall word accuracies on the test sets to approach the levels achieved by the previous keyword-based training experiments, since the poor quality data is excluded from both training runs. Referring again to Table4 , the overall word accuracies using both train 12000 and train 18000 can be found on the last two rows, labeled as OFW (omitted function words). We see that they are indeed nearly as high as those for the keyword-based training experiments, providing support for our hypothesis.... ..."

"... In PAGE 3: ... Thus, compared with non-keyword training on train 12000, using train 12000 measures the effect of key- word training when the same amount of training data is avail- able, while using train 18000 measures the effect when similar amounts of training data are actually used. Improvements in keyword error rate on test 500 and test 2500 relative to the baseline models are shown in Table2 for three training runs. The keyword error rate barely decreases from the baseline when using standard MCE training; it appears most of the previously observed overall error rate reduction comes from other words.... In PAGE 3: ... Because our eventual goal is to improve understanding error rates more so than word error rates, we compute a measure of understanding error rate that is equal to the sum of substitutions, insertions, and deletions of semantic frame entries following the method described in [6]. Table 3 summarizes the results as mea- sured on test 500, which we see correlate with the keyword error rate results of Table2 . We find that the keyword-based discrimi- native training results in a larger decrease in understanding error rate as compared to full discriminative training.... ..."