"... In PAGE 3: ... Here, it is often desirable to extend the alphabet to incorporate n-grams that include the frequent characters; in the expanded alphabet, the highest probability will be reduced and Hu man coding could be e ective. That Hu man coding is e ective for databases of natural text is shown in the rst two columns of Table1 , which is discussed in detail below. There we nd the Hu man cost (de ned as... In PAGE 6: ...and 735 bigrams), the distribution has been computed using the database of The Re- sponsa Retrieval Project (RRP) [14] of about 40 million Hebrew and Aramaic words; the distribution for Italian, Portuguese and Spanish (26 letters each) can be found in Gaines [18], and for Russian (32 letters) in Herdan [23]. The results are summarized in Table1 , the two last lines corresponding to the bigrams. The rst two colums list the average codeword length of Hu man codes and arith- metic codes respectively, and the third column gives the increase of the former over the latter in percent.... In PAGE 10: ...rom 4.16 to 5.60! We thus see that arithmetic codes give worse compression in this case, even without considering the overhead caused by EOF. Table 3 summarizes an experiment in which we took the probability distributions of English, German, Finnish and French (as in Table1 ), adding to them the character distribution in Gadsby, and checked what happens if they are mutually interchanged. The rows correspond to the distributions which are used to generate the codewords (the assumed distribution), and the columns correspond to the distribution that actually occurs (the true distribution).... In PAGE 17: ... To store the alphabet with methods B, C and D one needs 26, 32 and 3 bytes respectively, to which 13 bytes have to be added for the lengths of the Hu man codewords, or 52 bytes for the probabilities for arithmetic codes. From Table1 we know that the average loss per character by using Hu man instead of arithmetic codes is 0.0251 bits, so the text has to be at least of length 12431 characters to justify the excess of these 39 bytes for methods B, C and D.... ..."

"... In PAGE 5: ... Why Multiple Modes ? The justification for using multiple output media to present information are given elsewhere in this book. The motivation for trying to use multiple input modes rather than relying on direct manipulation or command languages alone is that individual modes have different strengths and weaknesses as illustrated in Table1 (after Cohen, 1992). ------------------------ Table 1 about here ------------------------ The objective of using multiple input modes is to allow the user to utilise the strengths of each mode while providing mechanisms for overcoming the weaknesses of each.... In PAGE 5: ... The motivation for trying to use multiple input modes rather than relying on direct manipulation or command languages alone is that individual modes have different strengths and weaknesses as illustrated in Table 1 (after Cohen, 1992). ------------------------ Table1 about here ------------------------ The objective of using multiple input modes is to allow the user to utilise the strengths of each mode while providing mechanisms for overcoming the weaknesses of each. In conjunction with multiple output media the use of the corresponding input media will provide maximum engagement with the information.... ..."

"... In PAGE 5: ...Why Multiple Modes ? The justification for using multiple output media to present information are given elsewhere in this book. The motivation for trying to use multiple input modes rather than relying on direct manipulation or command languages alone is that individual modes have different strengths and weaknesses as illustrated in Table1 (after Cohen, 1992). ------------------------ Table 1 about here ------------------------ The objective of using multiple input modes is to allow the user to utilise the strengths of each mode while providing mechanisms for overcoming the weaknesses of each.... In PAGE 5: ... The motivation for trying to use multiple input modes rather than relying on direct manipulation or command languages alone is that individual modes have different strengths and weaknesses as illustrated in Table 1 (after Cohen, 1992). ------------------------ Table1 about here ------------------------ The objective of using multiple input modes is to allow the user to utilise the strengths of each mode while providing mechanisms for overcoming the weaknesses of each. In conjunction with multiple output media the use of the corresponding input media will provide maximum engagement with the information.... ..."

"... 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 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 10: ... We used the C++ programming language with the GNU g++ compiler under the Linux operating system. Table1 highlights the performance of our algorithm on these models. The execution times are on a 2 GHz Pentium IV PC with a GeForce 4 graphics card and 1 GB RAM.... In PAGE 10: ... The execution times are on a 2 GHz Pentium IV PC with a GeForce 4 graphics card and 1 GB RAM. Table1 provides a breakup of the total time in computing pairwiseconvexMinkowskisums,constructingaconnectivity roadmap and performing the graph search in the connectivity graph. It shows that most of the time is spent on roadmap con- struction and a very small fraction of the total time is spent on graph search and pairwise Minkowski sum computation.... ..."

"... In PAGE 4: ... Table1 : A list of selected text mining products and applications based on the text refining and knowledge distillation functions as well as the intermediate form... ..."

"... In PAGE 32: ... A controlled vocabulary and natural language are two tools that can be implemented for this search process. Examples of controlled vocabulary and accompanying natural language are in Table3 . The natural language area will need to be frequently updated based on input from the professionals using the system.... ..."