### Table 3: Comparison of k-nn error rates using orig- inal distance measure, and using a 256-dimensional parameter-space query-sensitive embedding. We show the error rates for 1-nn, and for the value of k that achieved the lowest k-nn error rate (the best k equals 5 in both cases).

2004

"... In PAGE 15: ... These results agree with our expectations. Figure 11 and Table3 compare the classification error rates achieved using the original chamfer distance and using a 256-dimensional parameter-space query-sensitive embed- ding. It is interesting to note that for most values of k the embedding actually achieves a lower error rate than the original distance measure.... ..."

Cited by 2

### Table 3: Comparison of k-nn error rates using orig- inal distance measure, and using a 256-dimensional parameter-space query-sensitive embedding. We show the error rates for 1-nn, and for the value of k that achieved the lowest k-nn error rate (the best k equals 5 in both cases).

2004

"... In PAGE 15: ... These results agree with our expectations. Figure 11 and Table3 compare the classi cation error rates achieved using the original chamfer distance and using a 256-dimensional parameter-space query-sensitive embed- ding. It is interesting to note that for most values of k the embedding actually achieves a lower error rate than the original distance measure.... ..."

Cited by 2

### Table 1. Query term statistics for the six test collections This is verified by the behaviour of the test collections used in this experimental environment. Table 1 presents in the first row the percentage of relevant documents which contain at least one query term for each of the six collections2 used in this work, and in the second row the average number of query terms contained in a relevant document. The figures in the first row of this table all exceed 91%, an exceptionally high value that verifies the highly topical and algorithmic nature of relevance that is employed in standard IR evaluation (Schamber et al., 1990). The implication of this for the query-sensitive measures presented here, and especially for M1, is that the likelihood for pairs of co-relevant documents to contain at least one query term in common is high. It should be mentioned that if the pair of documents under comparison contains non-overlapping sets of query terms, this will not be taken into account as an indication of co-relevance by any of the similarity measures presented here. Although the presence of query terms in both documents can be seen as a source of evidence of their co-relevance, this is not incorporated by the query-sensitive similarity measures. The main reason for

"... In PAGE 13: ...7 terms on average per document. The length of the queries for this collection is large (almost 20 terms per query on average, Table1 ), almost half the average document size. Moreover, as it was mentioned in section 3, relevant documents in this collection contain on average 4.... In PAGE 13: ... Moreover, as it was mentioned in section 3, relevant documents in this collection contain on average 4.5 query terms ( Table1 ). Taking these characteristics into account, it can be appreciated why query influence in this database is strong: the combination of short documents, long queries and relatively large 1.... ..."

Cited by 2

### Table II. Stress Values for Query-Insensitive and Query-Sensitive Embeddings Dataset Query-insensitive stress Query-sensitive stress

### Table 7. CISI and Medline results Statistical tests of the results reveal significant improvements of M1 and M3 over the cosine (significance level lt;0.001 for the majority of cases) for all experimental conditions except for the CISI collection when n=100. Measure M2 is significantly more effective than the cosine for the CACM (except for n=100), LISA (all values of n), and Medline (except for n=100, 750, full) collections. It is also significantly more effective than the cosine when using the WSJ collection for n=750, 1000. Significance levels for M2 are not as low as the ones for M1 and M3, but they are still lower than 0.04 for all significant cases. The gains in effectiveness introduced by using QSSM are in most cases material , i.e. over 10%, which confirms the significance of the results (Keen, 1992). The largest differences occur when using the LISA collection, where all three query-sensitive measures are over 50% more effective than the cosine in all experimental conditions. Even M2, which relies only on common terms between documents that are query terms, introduces improvements of that magnitude. This behaviour for LISA can be explained on the basis of

Cited by 2

### Table 6.6: Characteristic of inter-document similarity matrix for query quot;jaguar quot;.

2003

### Table 3: The step by step calculation of similarity measure between pairs of objects

1998

"... In PAGE 7: ...To address these problems, ECOBWEB de nes two other methods for computing the interval values, a dynamic approach, where interval width is a function of the feature-value variance, and an adaptive approach, which uses the geometric mean of the static and dynamic approaches in computing interval size. Table3 lists the three schemes that have been implemented in ECOBWEB to generate the 2I i value, all of which rely on a user de ned parameter, the \expected number of distinct intervalsofproperty A i quot;, n. Since the choice of the 2I value is likely to haveavery signi cant e ect on the performance of the system, it is important to see howthetwo user controlled parameters: the method for calculating 2I i , and n, the expected distinct intervals of a property, a ect the structure of the clustering hierarchy.... ..."

Cited by 4

### Table 2: Parameters for Similarity and Confidence Calculation.

"... In PAGE 2: ...4. Table2 shows an example of an article pair. As shown in Table 2, an English article is not a literal translation of a Japanese article, although their contents are almost parallel.... In PAGE 2: ...s 7.4. Table 2 shows an example of an article pair. As shown in Table2 , an English article is not a literal translation of a Japanese article, although their contents are almost parallel. 2.... In PAGE 5: ...1 Usually, the equality score between I and S is equal to the number of phrases in I (the number of phrase corre- spondences in EQ), but sometimes these are slightly dif- ferent, depending on the conjugation type and function words. 1All constant values in Table2 and formulas were decided based on preliminary experiments. On the other hand, the similarity between the surround- ings of I and those of S is a sum of the similarity score of each phrase correspondence in CONTEXT, which is cal- culated as follows: SIMB4CXB5BP AQ CG CB CRD3D2D8 A2 BE AZ CRD3D2D8 B7BCBMBEA2 C8 CB CUD9D2CR A2 BE AZ CUD9D2CR AR A2CB CRD3D2D2CTCRD8 BM (4) Basically the calculation of SIM and EQUAL is the same, except that SIM considers the relation type between the phrase in I and its outer phrase by CB CRD3D2D2CTCRD8 .... In PAGE 5: ... The monolingual similarity between Japanese expres- sions I and S is calculated as follows: CG CXBEBXC9 EQUALB4CXB5B7 CG CXBEBVC7C6CCBXCGCC SIMB4CXB5BM (5) 3.3 Translation Confidence of Japanese-to-English Alignment The translation confidence of phrase alignment between S and T is the sum of the confidence score of each phrase correspondence in ALIGN, CONF(CX) in Table2 , and it is weighted by the WCR of the parallel sentences. As a final measure, the score of I-S-T is calculated as follows: D2 CG CXBEBXC9 EQUALB4CXB5B7 CG CXBEBVC7C6CCBXCGCC SIMB4CXB5 D3 A2 D2 CG CXBEBTC4C1BZC6 CONFB4CXB5 D3 A2 WCRBM (6) 3.... ..."

### Table 2: Parameters for Similarity and Confidence Calculation.

"... In PAGE 2: ...4. Table2 shows an example of an article pair. As shown in Table 2, an English article is not a literal translation of a Japanese article, although their contents are almost parallel.... In PAGE 2: ...s 7.4. Table 2 shows an example of an article pair. As shown in Table2 , an English article is not a literal translation of a Japanese article, although their contents are almost parallel. 2.... In PAGE 5: ...1 Usually, the equality score between I and S is equal to the number of phrases in I (the number of phrase corre- spondences in EQ), but sometimes these are slightly dif- ferent, depending on the conjugation type and function words. 1All constant values in Table2 and formulas were decided based on preliminary experiments. On the other hand, the similarity between the surround- ings of I and those of S is a sum of the similarity score of each phrase correspondence in CONTEXT, which is cal- culated as follows: SIMB4CXB5BP AQCG CBCRD3D2D8 A2 BE AZCRD3D2D8 B7BCBMBEA2 C8 CBCUD9D2CR A2 BE AZCUD9D2CR AR A2CBCRD3D2D2CTCRD8BM (4) Basically the calculation of SIM and EQUAL is the same, except that SIM considers the relation type between the phrase in I and its outer phrase by CBCRD3D2D2CTCRD8.... In PAGE 5: ... The monolingual similarity between Japanese expres- sions I and S is calculated as follows: CG CXBEBXC9EQUALB4CXB5B7 CG CXBEBVC7C6CCBXCGCC SIMB4CXB5BM (5) 3.3 Translation Confidence of Japanese-to-English Alignment The translation confidence of phrase alignment between S and T is the sum of the confidence score of each phrase correspondence in ALIGN, CONF(CX) in Table2 , and it is weighted by the WCR of the parallel sentences. As a final measure, the score of I-S-T is calculated as follows: D2 CG CXBEBXC9EQUALB4CXB5B7 CG CXBEBVC7C6CCBXCGCC SIMB4CXB5D3 A2D2 CG CXBEBTC4C1BZC6 CONFB4CXB5D3 A2 WCRBM (6) 3.... ..."

### Table 2: Parameters for Similarity and Confidence Calculation.

"... In PAGE 2: ...4. Table2 shows an example of an article pair. As shown in Table 2, an English article is not a literal translation of a Japanese article, although their contents are almost parallel.... In PAGE 2: ...s 7.4. Table 2 shows an example of an article pair. As shown in Table2 , an English article is not a literal translation of a Japanese article, although their contents are almost parallel. 2.... In PAGE 5: ...1 Usually, the equality score between I and S is equal to the number of phrases in I (the number of phrase corre- spondences in EQ), but sometimes these are slightly dif- ferent, depending on the conjugation type and function words. 1All constant values in Table2 and formulas were decided based on preliminary experiments. On the other hand, the similarity between the surround- ings of I and those of S is a sum of the similarity score of each phrase correspondence in CONTEXT, which is cal- culated as follows: SIMB4CXB5BP AQ CG CB CRD3D2D8 A2 BE AZ CRD3D2D8 B7BCBMBEA2 C8 CB CUD9D2CR A2 BE AZ CUD9D2CR AR A2CB CRD3D2D2CTCRD8 BM (4) Basically the calculation of SIM and EQUAL is the same, except that SIM considers the relation type between the phrase in I and its outer phrase by CB CRD3D2D2CTCRD8 .... In PAGE 5: ... The monolingual similarity between Japanese expres- sions I and S is calculated as follows: CG CXBEBXC9 EQUALB4CXB5B7 CG CXBEBVC7C6CCBXCGCC SIMB4CXB5BM (5) 3.3 Translation Confidence of Japanese-to-English Alignment The translation confidence of phrase alignment between S and T is the sum of the confidence score of each phrase correspondence in ALIGN, CONF(CX) in Table2 , and it is weighted by the WCR of the parallel sentences. As a final measure, the score of I-S-T is calculated as follows: D2 CG CXBEBXC9 EQUALB4CXB5B7 CG CXBEBVC7C6CCBXCGCC SIMB4CXB5 D3 A2 D2 CG CXBEBTC4C1BZC6 CONFB4CXB5 D3 A2 WCRBM (6) 3.... ..."