Social tagging using marked strings in Web pages Yuki Matsuokaⅰⅱⅲ, Ryuuki Sakamotoⅲ, Sadanori Itoⅳⅲ, Hideaki Takedaⅰⅱⅲ, Kiyoshi Kogureⅲ ⅰThe Graduate University for Advanced Studies, Shonan Village, Hayama, Kanagawa 240-0193 Japan ⅱNational Institute of Informatics, 2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo 101-8430 Japan ⅲAdvanced Telecommunications Research Institute International, 2-2-2 Hikaridai, Seikacho, Soraku-gun, Kyoto 619-0288 Japan ⅳTokyo University of Agriculture and Technology, 3-8-1 Harumi-cho, Fuchu-shi, Tokyo 183-8538 Japan E-mail : m- yuki@grad.nii.ac.jp ABSTRACT words that differ from what users intended. One possible In this paper, we propose a system called “Aikuchi,” which solution to this is to employ user-generated tags from Social enables users to mark a string on a Web page and share it with Bookmark Services (SBS), enabling users to easily find other users. If users mark a string on their Web browser using appropriate pages. However, as users freely add tags, numerous their mouse cursor, Aikuchi recommends links to other Web other types of tags also accumulate on Web pages. pages based on recommendation algorithms. When a user clicks Because users view each page subjectively, they add tags to a recommended link, Aikuchi highlights a marked string. We express the contents, their feelings about the page, the purpose of offered the system as part of a conference support system. the page, and so on. When users find Web pages using such tags, According to the analysis of user logs, users preferred Web page though, such diversity becomes an obstacle to finding pages that recommendations based on strings marked by other users to those are actually appropriate. In this research, we employ strings based on page similarity using TFIDF and collaborative filtering. marked by users as tags to exclusively express a given page’s As a result, we think that marked strings can act as user- contents. The system we have implemented, called Aikuchi, generated metadata. enables users to mark strings in Web pages and share them with other others. If a user marks a string in the Web browser using Categories and Subject Descriptors his or her mouse cursor, Aikuchi recommends links to other Web H.3.3 [Information Search and Retrieval]: Search process, pages by applying it recommendation algorithms, and when a Selection process; H.3.5 [Online Information Services]: Data user clicks a recommended link, the system highlights the sharing marked string, which is called a footprint. The recommendation algorithms are based on page similarity using TFIDF, collaborative filtering, and word matching using words in the General Terms footprint or another Web page. We offered the system as a part of Experimentation, Human Factors a conference support system and analyzed whether users prefer the recommendation algorithm based on marked strings to search Keywords Web pages. Social tagging, marking, user-generated metadata 2. RELATED RESEARCH 1. INTRODUCTION Semantic annotation systems have previously been proposed It goes without saying that there is an enormous number of that highlight strings in Web browsers. One example is CHOSE pages on the Internet; therefore, we often need to use search [1], which stores metadata as hyperlinks in a Distributed Links engines to find ones that interest us. However, search results Service [2] and uses it to highlight ontology terms in the text frequently include “noise” pages. While search engines do return strings on the Web browser. Magpie [3] highlights strings related pages matching users’ query words, some pages may include to the ontology of the user’s choice on the browser. Both systems highlight strings to make the connection between a string and an ontology. Our system links from a string to offered pages and Permission to make digital or hard copies of all or part of this work for differs from them by the lack of connection between the link and personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies a meaning; it is instead related to the user’s interest. In this bear this notice and the full citation on the first page. To copy otherwise, or paper, we reuse a selected string as metadata to search Web republish, to post on servers or to redistribute to lists, requires prior specific pages. permission and/or a fee. SAAW’06, November 6, 2006, Athens, GA, USA. Copyright 2006 ACM 1-58113-000-0/00/0004…$5.00. 3. A SOCIAL TAGGING SYSTEM 3.1 Overview We propose a system called Aikuchi that enables users to mark strings in Web pages and jump to other pages from the marked string. The system lets users search pages using marked strings from other users. Details of the system are as follows: Users can mark a string in their Web browser using their mouse cursor when they find something of interest on a page. (Fig. 1). Then the recommendation window pops out to display recommendations from other pages called recommendation links (Fig. 2). If the user selects a recommendation link, the system shows the specified page, and once a recommendation link is selected, the Figure 3. The footprint system highlights the used string. We call this highlighted string The right figure is translation of the left it. a footprint (Fig. 3). It is shared; i.e., users can see the footprints added by all other users. If the user places her or his mouse cursor over the footprint, the recommendation window pops out again, with Fig. 4 Showing examples of recommendation links and footprint links. The latter includes Web pages to which some users have jumped from the marked string. Figure 4. Recommendation window after the cursor is placed over the footprint The right figure shows translation of the footprint link and the recommendation links Figure 1. Marking on the Web browser The left figure indicates title, authors and abstract for a paper. 3.2 Implementation The right figure is translation of the left it. Figure 5 shows the system’s structure. The system runs as a script on the user client and on the Web server, and is implemented with JavaScript and PHP. When a user accesses a Web page, the JavaScript engine asks the Web server if a footprint exists, and shows it if that is the case. When a user marks a string in the Web browser, the JavaScript engine obtains the marked string and sends it to the Web server, which then takes the marked string and calculates the recommendation links according to certain algorithms. Following calculation, the Web server sends recommendation links and the JavaScript engine shows the recommendation window to the user. If the user selects a recommended link, the JavaScript engine sends it and the Web server saves the following information related to the selected link as a text file in the metadata storage. Date User ID Figure 2. Popped-out recommendation window The right figure shows translation of the recommendation links Marked string Position of the marked string on the Web page URL of the selected recommended link find useful pages based on the string or footprints. To make recommendations, we use the system in a closed environment Selected recommendation algorithm where the available Web pages are fixed1. The algorithms are as follows: A) Page similarity using TFIDF B) Collaborative filtering using the number of footprints on a Web page C) Word matching between a marked string and footprint’s strings D) Word matching between a marked string and the Web page’s strings. In algorithm A, the system employs calculated page similarity based on TFIDF [4]. We calculate the TFIDF value of words on Web pages and cosine similarity using them. If a user marks a certain string on a Web page, the system recommends high- similarity pages to it regardless of the marked string and footprints. In algorithm B, the system recommends pages with collaborative filtering [5], using the number of footprints on a Web page as the users’ evaluation of that page. We adopt not incoming links but outgoing links, because selected recommendation links are not valuable for users. Therefore, the system recommends Web pages with a high predictive value using the number of footprints. In algorithm C, if a word in the marked string matches one in the footprint’s string on other pages, the system recommends the matched page. In algorithm D, if a word in the marked string matches one in another page’s string, the system recommends the matched page. Table 1 shows a comparison of recommendation algorithms with respect to whether they use a marked string and footprints. Table 1. Comparison of recommendation algorithms Recommendation A marked string Footprints algorithm A × × B × ○ C ○ ○ D ○ × Algorithm A makes recommendations based on page similarity between Web pages without using a marked string or footprints. Figure 5. System structure This is assumed to be ordinary recommendation. Algorithm B, When a user places his or her mouse cursor over a footprint, the meanwhile, uses the number of footprints on a Web page to look JavaScript engine takes the string of the footprint and sends it to for neighborhood users, and algorithms C and D use a marked the Web server. The Web server gets the footprint’s string, string as a query to search Web pages with matching words. The calculates recommendation links, and obtains footprint links difference is the search target: a footprint’s strings or a Web from the metadata storage. The Web server then sends both page’s strings. footprint links and recommendation links. Next, the JavaScript We investigated which algorithms are preferred by users. The engine again displays the recommendation window to the user system recommends up to two Web pages for each algorithm and and if the user selects a footprint link or a recommended link, the shows them in random order. Algorithm priority was applied so JavaScript engine sends it and the Web server saves information that the same link was calculated in order of C, D, B, and A. The about it. system did not inform users about these algorithms. When users clicked a recommended page link, we determined that users 3.3 Recommendation algorithms preferred its recommendation algorithm. When a user marks a string or places the mouse cursor over a footprint on a Web page, the system shows recommendation 1 Aikuchi is not limited to closed environments, but we restricted links using four types of algorithms that explore Web pages to the system to test recommendations in the following test cases. 4. ANALYSIS previously but also jumped pages. We investigated which We offered Aikuchi as a part of a support system at the algorithms are preferred by users in such cases (Fig. 7). The Japanese Society for Artificial Intelligence 2006, which was held algorithm Z denotes the selection of jumped pages by users, and from June 7th to 9th, 2006. The support system was operated as a the figure shows that Z was selected in an overwhelming number Web system, and every conference participant could access it of cases. Therefore, we found that users prefer to select links by using a user ID and password. The target Web pages comprised using footprints. 276 pages that included authors and abstracts for papers. We prepared recommendation pages using algorithm A. When users 80 marked a string, the system took a few seconds to display 70 recommendation links. After the completing the experiment, we B e fo r e a n d D u r in g t h e 60 c o n fe re n c e obtained 324 footprints and 172 links from them. Analysis of the 50 results revealed that there were 45 users who marked strings one 40 or more times, 28 users who jumped from marked strings to 30 other pages, 88 users who placed their mouse cursors over 20 footprints one or several times, and 33 users who jumped from 10 footprints to other pages. In this section, we describe which 0 A B C D Z algorithms were preferred by users. Figure 7. Number of selected recommendation algorithms, 4.1 Preference of the algorithms when users placed their mouse cursor over a footprint. Figure 6 shows a comparison of the number of selected recommendation algorithms before and during the conference. 5. CONCLUSION Before the conference, algorithm A was most commonly selected We proposed to use strings marked by users on Web pages as and the number of selections of B and C were few since the tags, and developed a system called Aikuchi with which users footprints were few at the beginning the conference. The number can mark strings and share them as footprints. Aikuchi of selections for D was fewer than that of A before the recommends links based on a variety of algorithms when a user conference. This means users selected a recommended link based marks a string on a Web page. on similarity without relation to marked strings or footprints. During the conference, the number of selections of C and D were Based on the analysis of user logs, we found that users more than before it and the number of selections of B was fewer preferred recommendations based on words in marked strings than before it, since if Aikuchi recommended the same page links rather than page similarity. We consider that footprints are useful over time, then users would be tired of selecting the links for users to select links because links based on footprints are recommended by B. The number of selections for algorithm A frequently selected by users. Consequently, we believe that was also lower during the conference. This means users selected strings marked by users on Web pages can be used as user- the recommendation algorithms based on the word in the marked generated metadata for searching pages. string. Therefore, we found that users prefer recommendations based on words rather than similarity over time. As the number 6. ACKNOWLEDGMENTS of selections for algorithm C increased, we also found that This research was supported by the National Institute of footprints are effective for recommending links. Information and Communications Technology. 30 7. REFERENCES [1] L. Carr, S. Bechhofer, C. Goble, and W. Hall. Conceptual B e fo r e t h e c o n f e r e n c e 25 (5 / 2 2 - 6 / 6 ) Linking: Ontology-based Open Hypermedia. In WWW10, 20 D u r in g t h e c o n f e r e n c e Tenth World Wide Web Conference, May 2001. (6 / 7 - 6 / 9 ) 15 [2] L. Carr and D. de Roure, W. Hall, G. Hill, The distributed link service: a tool for publishers, authors and readers, 10 World Wide Web J. 1 (1) (1995) 647–656. 5 [3] M. Dzbor, J. B. Domingue, and E. Motta.: Magpie - towards 0 a Semantic Web browser, In Proc. of the 2nd Intl. Semantic A B C D Web Conf., October 2003, Florida US [4] G. Salton: Developments in automatic text retrieval, Science, Figure 6. Number of selected recommendation algorithms, Vol. 253, pp. 974-980 (1991). when users marked a string on a Web page. [5] P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, and J. Riedl, “GroupLens: An Open Architecture for Collaborative 4.2 Footprints Filtering of Netnews,” In Proceedings of the 1994 Computer When users place their mouse cursor over a footprint, they can Supported Collaborative Work Conference, pp. 175–186, 1994. obtain not only recommendations by the algorithms described