Academia.eduAcademia.edu

Figure 10 – uploaded by Sanna Kumpulainen

See full PDFdownloadDownload figure
Figure 1: Example of a set of candidate pages with probabilities and keywords and a hierarchical structure for this set. ‘Nota’ is short for ‘None of the above’.

Figure 1 Example of a set of candidate pages with probabilities and keywords and a hierarchical structure for this set. ‘Nota’ is short for ‘None of the above’.

Related Figures (55)

Contents
Contents
//Email[ subject = "Multimedia" and from = "Makici" and date="0ctober"] Figure 1: Examples of different ways of expressing an information need.
//Email[ subject = "Multimedia" and from = "Makici" and date="0ctober"] Figure 1: Examples of different ways of expressing an information need.
Table 2: Statistics on the main tags in the query set. were 91 instances where query terms were “about” a field in an email, which indicates that a considerable amount of noise is present within the queries. We refer to this noise as ambiguity as the query being expressed to the system contains uncertain information with respect to the target email. We classified queries according to how much ambigu- ity was present, using three grades: (0) not, (1) somewhat, or (3) very ambiguous. If all the features occurred in the known-item, then there is little or no ambiguity (i.e., we assumed that the query term was put there with specific reference to some field in the email). However, if more than half the query features are present in the known-item email, then there is some ambiguity in the query. If the majority of query terms do not occur in the email then the query is very ambiguous. To some extent this measure reflects the loss of recall that is experienced by a user when formulating the query; assuming that they are trying to select (remem- ber) the exact words and phrases from the email they have in mind. The more vague the user is about their missing email, presumably the less precise and more ambiguous the query will be as a result. Figure 2: Examples of different ways of expressing an information need.
Table 2: Statistics on the main tags in the query set. were 91 instances where query terms were “about” a field in an email, which indicates that a considerable amount of noise is present within the queries. We refer to this noise as ambiguity as the query being expressed to the system contains uncertain information with respect to the target email. We classified queries according to how much ambigu- ity was present, using three grades: (0) not, (1) somewhat, or (3) very ambiguous. If all the features occurred in the known-item, then there is little or no ambiguity (i.e., we assumed that the query term was put there with specific reference to some field in the email). However, if more than half the query features are present in the known-item email, then there is some ambiguity in the query. If the majority of query terms do not occur in the email then the query is very ambiguous. To some extent this measure reflects the loss of recall that is experienced by a user when formulating the query; assuming that they are trying to select (remem- ber) the exact words and phrases from the email they have in mind. The more vague the user is about their missing email, presumably the less precise and more ambiguous the query will be as a result. Figure 2: Examples of different ways of expressing an information need.
Table 3: Classification Accuracy shown as a percent- age (%) correct per class on the Independence and Dependence Classification models.
Table 3: Classification Accuracy shown as a percent- age (%) correct per class on the Independence and Dependence Classification models.
Table 1: Manual element name expansion lists based on INEX 2004 assessments.
Table 1: Manual element name expansion lists based on INEX 2004 assessments.
Table 3: Equivalence classes in INEX IEEE collection.
Table 3: Equivalence classes in INEX IEEE collection.
Table 2: Score region algebra operators.
Table 2: Score region algebra operators.
Table 4: INEX 2004 CAS experiments with different expansion classes evaluated using inex_eval and precision different recall points. Table 5: INEX 2005 CAS experiments with different expansion classes evaluated using nxCG at different recall point and ep/gr.
Table 4: INEX 2004 CAS experiments with different expansion classes evaluated using inex_eval and precision different recall points. Table 5: INEX 2005 CAS experiments with different expansion classes evaluated using nxCG at different recall point and ep/gr.
Table 6: INEX 2005 COS experiments with different expansion classes evaluated using nxCG at different recall point and ep/gr.
Table 6: INEX 2005 COS experiments with different expansion classes evaluated using nxCG at different recall point and ep/gr.
Table 7: INEX 2004 CAS experiments with different vague scenarios and rewriting techniques evaluated using inex_eval and precision at different recall points.
Table 7: INEX 2004 CAS experiments with different vague scenarios and rewriting techniques evaluated using inex_eval and precision at different recall points.
Table 8: INEX 2005 CAS experiments with different vague scenarios and rewriting techniques evaluated using nxCG at different recall points and ep/gr.
Table 8: INEX 2005 CAS experiments with different vague scenarios and rewriting techniques evaluated using nxCG at different recall points and ep/gr.
Table 9: INEX 2005 COS experiments with different vague scenarios and rewriting techniques evaluated using nxC' at different recall points and ep/gr.
Table 9: INEX 2005 COS experiments with different vague scenarios and rewriting techniques evaluated using nxC' at different recall points and ep/gr.
Table 10: INEX 2004 CAS experiments on combining vague search and rewriting techniques evaluated using inex_eva and precision at different recall points.
Table 10: INEX 2004 CAS experiments on combining vague search and rewriting techniques evaluated using inex_eva and precision at different recall points.
Table 11: INEX 2005 CAS experiments on combining vague search and rewriting techniques evaluated using nxCC at different recall points and ep/gr.
Table 11: INEX 2005 CAS experiments on combining vague search and rewriting techniques evaluated using nxCC at different recall points and ep/gr.
Table 12: INEX 2005 COS experiments on combining vague search and rewriting techniques evaluated using nxCG at different recall points and ep/gr.
Table 12: INEX 2005 COS experiments on combining vague search and rewriting techniques evaluated using nxCG at different recall points and ep/gr.
‘able 1. The MAP values (% ) for the test queries and their difference to the baselines (% ) (* statistically significant difference, ** statistically highly significant difference)
‘able 1. The MAP values (% ) for the test queries and their difference to the baselines (% ) (* statistically significant difference, ** statistically highly significant difference)
Figure 1. Recall-precision curves for all queries.
Figure 1. Recall-precision curves for all queries.
‘able 2. The interpolated recall precision averages (% ) at standard recall level 10 for the test queries, and their difference to the baselines. (* statistically significant difference, ** statistically highly significant difference) Table 3. The interpolated recall precision averages (% ) at standard recall level 50 for the test queries, and their difference to th baselines. (* statistically significant difference, ** statistically highly significant difference)
‘able 2. The interpolated recall precision averages (% ) at standard recall level 10 for the test queries, and their difference to the baselines. (* statistically significant difference, ** statistically highly significant difference) Table 3. The interpolated recall precision averages (% ) at standard recall level 50 for the test queries, and their difference to th baselines. (* statistically significant difference, ** statistically highly significant difference)
The extracted information is stored in an XML file which is then accessible by the retrieval component of the system. This component, which is currently under development, dynamically forms an index of the processed conferences based on the information found in the XML repository. When projected to the client’s browser conferences are classified as open or past and they are categorized based on their date. This tool will also allow multiriteria retrieval of conferences, such as “show me conferences in Athens or near Athens which are about Web mining and will take place this summer”. Supporting these queries will be based on the location knowledge base and on the month dictionary. Figure 1. Flowchart of the extraction procedure.
The extracted information is stored in an XML file which is then accessible by the retrieval component of the system. This component, which is currently under development, dynamically forms an index of the processed conferences based on the information found in the XML repository. When projected to the client’s browser conferences are classified as open or past and they are categorized based on their date. This tool will also allow multiriteria retrieval of conferences, such as “show me conferences in Athens or near Athens which are about Web mining and will take place this summer”. Supporting these queries will be based on the location knowledge base and on the month dictionary. Figure 1. Flowchart of the extraction procedure.
Table 1. Precision and Recall of the extraction procedure
Table 1. Precision and Recall of the extraction procedure
Table 2. Precision and Recall of location queries in Google 7. REFERENCES
Table 2. Precision and Recall of location queries in Google 7. REFERENCES
Table 1: Recall of Academy Award Winners Recall. The number of entries in IMDb exceeds our ontology by far. Although our algorithm performs especially well on re- cent productions, we are interested how well it performs on classic movies, actors and directors. First, we made lists of all Academy Award winners (1927-2005) in a number of relevant categories, and checked the recall (Table 1). IMDb has a top 250 of best movies ever. The algorithm found 85% of them. We observe that results are strongly oriented towards Hollywood productions. We also made a list of all winners of the Cannes Film Festival, the ‘Palme d’Or’. Alas, our algorithm only found 26 of the 58 winning movies in this category.
Table 1: Recall of Academy Award Winners Recall. The number of entries in IMDb exceeds our ontology by far. Although our algorithm performs especially well on re- cent productions, we are interested how well it performs on classic movies, actors and directors. First, we made lists of all Academy Award winners (1927-2005) in a number of relevant categories, and checked the recall (Table 1). IMDb has a top 250 of best movies ever. The algorithm found 85% of them. We observe that results are strongly oriented towards Hollywood productions. We also made a list of all winners of the Cannes Film Festival, the ‘Palme d’Or’. Alas, our algorithm only found 26 of the 58 winning movies in this category.
Table 2: Some examples to illustrate the difficulties in discrim- inating between persons names and other text fragments. Not all text fragments we have found in the extraction phase will be person names. Typically, historic periods, art styles, geographic names, etc. can also directly precede a time interval. Table 2 illus- trates the difficulties in discriminating between person names and other text fragments. We note that West Mae is an inversion of the person name Mae West and that Napoleon Hill refers to a person as well as to a geographic location in the state Idaho (USA).
Table 2: Some examples to illustrate the difficulties in discrim- inating between persons names and other text fragments. Not all text fragments we have found in the extraction phase will be person names. Typically, historic periods, art styles, geographic names, etc. can also directly precede a time interval. Table 2 illus- trates the difficulties in discriminating between person names and other text fragments. We note that West Mae is an inversion of the person name Mae West and that Napoleon Hill refers to a person as well as to a geographic location in the state Idaho (USA).
nist. Table 3 give the top-40 of the professions found, ranked by the number of times that these professions were found in the excerpts. Table 3: The professions that were found most often.
nist. Table 3 give the top-40 of the professions found, ranked by the number of times that these professions were found in the excerpts. Table 3: The professions that were found most often.
Table 1: Characteristics of the three main test col- lections used in the experiments. The total au- thor count (‘# total authors’) is the sum of the au- thor count over all documents; the total number of unique authors (‘# unique authors’) is the sum of the author count over all documents with each au- thor counted only once.
Table 1: Characteristics of the three main test col- lections used in the experiments. The total au- thor count (‘# total authors’) is the sum of the au- thor count over all documents; the total number of unique authors (‘# unique authors’) is the sum of the author count over all documents with each au- thor counted only once.
Table 2: Author-related characteristics of the six special test collections.
Table 2: Author-related characteristics of the six special test collections.
Figure 1: Precision vs. Recall for CACM.
Figure 1: Precision vs. Recall for CACM.
Table 3: Comparison of the re-ranking approaches on R-precision scores. The underlined scores are statisti- cally significant improvements over the baseline.
Table 3: Comparison of the re-ranking approaches on R-precision scores. The underlined scores are statisti- cally significant improvements over the baseline.
Figure 3: Precision vs. Recall for ILK. results of, for instance, a probabilistic IR model. However, it would also be interesting to investigate whether using other IR models such as probabilistic retrieval or a language mod- elling approach indeed show this increase to be universal over the entire range of IR approaches.
Figure 3: Precision vs. Recall for ILK. results of, for instance, a probabilistic IR model. However, it would also be interesting to investigate whether using other IR models such as probabilistic retrieval or a language mod- elling approach indeed show this increase to be universal over the entire range of IR approaches.
Figure 2: Precision vs. Recall for CISI.
Figure 2: Precision vs. Recall for CISI.
Figure 1: ViPF interface after a query was submitted: result panel (1), graph panel (2), info panel (3), subject panel (4), fitness color bar (5)
Figure 1: ViPF interface after a query was submitted: result panel (1), graph panel (2), info panel (3), subject panel (4), fitness color bar (5)
Figure 2: In-degree distribution of the Citebase data set with yn = 2.1564
Figure 2: In-degree distribution of the Citebase data set with yn = 2.1564
Figure 3: Preferential attachment measurements of the Citebase data set
Figure 3: Preferential attachment measurements of the Citebase data set
Table 1: With an increase in network size, the aver- age in-degree k;, increases.
Table 1: With an increase in network size, the aver- age in-degree k;, increases.
Table 1: The average number of steps of simulated users with perfect choices.
Table 1: The average number of steps of simulated users with perfect choices.
Figure 4: The average number of steps of simulated users with various noise levels.
Figure 4: The average number of steps of simulated users with various noise levels.
Figure 5: The average number of steps that human users needed to reach their target pages. Table 2: The average number of steps of human users and the average number of targets that were found by human users.
Figure 5: The average number of steps that human users needed to reach their target pages. Table 2: The average number of steps of human users and the average number of targets that were found by human users.
Table 5: Responses on system suitability for answer- ing general questions: Mean rating and standard de- viation (in brackets). Answers we on a 5-point scale, ranging from 1 (“very unsuitable”) to 5 (“very suit- able” )
Table 5: Responses on system suitability for answer- ing general questions: Mean rating and standard de- viation (in brackets). Answers we on a 5-point scale, ranging from 1 (“very unsuitable”) to 5 (“very suit- able” )
Table 4: Responses on system suitability for answer- ing specific questions: Mean rating and standard de- viation (in brackets). Answers were on a 5-point scale, ranging from 1 (“very unsuitable”) to 5 (“very suitable” ).
Table 4: Responses on system suitability for answer- ing specific questions: Mean rating and standard de- viation (in brackets). Answers were on a 5-point scale, ranging from 1 (“very unsuitable”) to 5 (“very suitable” ).
Table 9: Page-link clicks vs. focused-link clicks in the focused interface: mean number of clicks and standard deviation (in brackets). Each search task contained three distinct search assignments.
Table 9: Page-link clicks vs. focused-link clicks in the focused interface: mean number of clicks and standard deviation (in brackets). Each search task contained three distinct search assignments.
Table 8: Time spent per search task (minutes): mean time and standard deviation (in brackets). Each search task was divided into three distinct search assignments.
Table 8: Time spent per search task (minutes): mean time and standard deviation (in brackets). Each search task was divided into three distinct search assignments.
Table 7: Page views per search task: Mean num- ber of page views and standard deviation (in brack- ets). Each search task was divided into three dis- tinct search assignments.
Table 7: Page views per search task: Mean num- ber of page views and standard deviation (in brack- ets). Each search task was divided into three dis- tinct search assignments.
Figure 4: Number of result clicks per user in the focused interface. Dark: Clicks on page-links. Light: Clicks on focused-links.
Figure 4: Number of result clicks per user in the focused interface. Dark: Clicks on page-links. Light: Clicks on focused-links.
Figure 5: Linear depth of Wikipedia pages which have one or more sections. The distribution of pages over the number of sections is plotted on a log-log scale.
Figure 5: Linear depth of Wikipedia pages which have one or more sections. The distribution of pages over the number of sections is plotted on a log-log scale.
Figure 6: Number of page visits per user in the focused interface. Dark: Pages visited via the result list. Light: Pages visited via internal links.
Figure 6: Number of page visits per user in the focused interface. Dark: Pages visited via the result list. Light: Pages visited via internal links.
Table 10: Analysis of focused-clicks in the focused interface. Left: Type of element clicked (hierarchi- cal depth). Right: Section number (in the Wiki- pedia source) of the of the sections clicked (linear depth).
Table 10: Analysis of focused-clicks in the focused interface. Left: Type of element clicked (hierarchi- cal depth). Right: Section number (in the Wiki- pedia source) of the of the sections clicked (linear depth).
Related topics:
Cross Language Information RetrievalString MatchingQuery Translation
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved