Boston University Graduate School of Arts and Sciences

Web Mediators for Accessible Browsing

by

Benjamin N. Waber

B.A., Boston University, 2006

Submitted in partial fulfillment of the requirements for the degree of Masters of Arts

2006

Approved by

Margrit Betke, Associate Professor of Computer Science, Boston University

Acknowledgements

This thesis is the result of joint work with many people in the Image and Video Computing Group atBoston University. My advisor, Margrit Betke, has been so supportive and helpful during my time here.John Magee has also been extremely gracious with his time and advice, and has helped my work immensely.I would also like to thank James Gips for providing motivation for this work, and Michelle Paquette fordeveloping the IWebBrowser interface. Finally, I’d like to thank my fiancee Rebecca and my family forsupporting me and helping me refine my goals and ideas throughout my academic career. Without them,none of this would have been possible.

2

Abstract

We present a highly accurate method for classifyingweb pages based on link percentage, which is thepercentage of text characters that are parts of linksnormalized by the number of all text characters ona web page. K -means clustering is used to createunique thresholds to differentiate index pages and ar-ticle pages on individual web sites. Index pages con-tain mostly links to articles and other indices, whilearticle pages contain mostly text. We also presenta novel link grouping algorithm using agglomerativehierarchical clustering that groups links in the samespatial neighborhood together while preserving linkstructure. Grouping allows users with severe disabil-ities to use a scan-based mechanism to tab througha web page and select items. In experiments, we sawup to a 40-fold reduction in the number of commandsneeded to click on a link with a scan-based interface,which shows that we can vastly improve the rate ofcommuncation for users with disabilities. We usedweb page classification and link grouping to alter webpage display on an accessible web browser that we de-veloped to make a usable browsing interface for userswith disabilities. Our classification method consis-tently outperformed a baseline classifier even whenusing minimal data to generate article and index clus-ters, and achieved classification accuracy of 94.0%on web sites with well-formed or slightly malformedHTML, compared with 80.1% accuracy for the base-line classifier.

1 Introduction

People who cannot physically use a mouse, for exam-ple because of quadriplegia, often rely on an assistivedevice that moves the mouse pointer by tracking theuser’s head or eyes. Computer access with such de-vices is difficult because they typically do not pro-vide the same selection accuracy as a mouse pointer.Moreover, since the user cannot type with a physi-cal keyboard, text entry, for example of a web ad-dress, requires the use of an onscreen keyboard. Se-lection of a letter on this keyboard or a small textlink in a web page may be particularly difficult on

traditional browsers for users who experience tremorsor other unintentional movements that prevent themfrom holding the mouse pointer still. One possible so-lution is to change (1) the display of a web page and(2) how the interface navigates the information basedon its content. A page can be rendered and navi-gated differently depending on the “type” of page.Such classification allows us to create a variety ofcustomizations to occur in interaction mode and dis-play depending on the intended application and user.We could also allow the user to select not just a sin-gle link, but a group of links. This would allow theweb browser to either enlarge the single group of linkson the page or, for users who only have control of abinary interface, allow them to scroll through indi-vidual links within that group.

The principal technical contributions of this paperare a clustering method to accurately determine thetype of a web page based on a technique that exam-ines the text characters on a page and a link groupingmethod that respects the structure of the web pagewhile providing groupings that substantially increasethe effectiveness of browsing.

The clustering method computes the link percent-age, the percentage of text characters that are partsof links as compared to all text characters on a webpage. We posit that there are only two types of pages– articles and index pages – on web sites that dealwith news media: article pages contain mostly textand index pages contain mostly links to articles andother indices. This classification may be helpful toallow people with disabilities to browse the web inan effective and efficient fashion. After determiningthe content of a web page as an index or article, ourmethod can render the page to meet the needs ofusers with disabilities, for example, by increasing thesize of links on index pages. This makes the linkseasier to read, but more importantly, it makes themeasier to select. All text on article pages is enlargedto increase readability.

We also present a novel link grouping algorithmthat preserves link structure to enable disabled usersto browse web pages orders of magnitude faster thancurrent systems allow.1 Our link grouping method

1The link grouping algorithm is implemented in Javascript

1

proceeds in two stages: first it builds a link tree. Theleaves of the link tree are the links on a web page.The parents of these leaves are the first common par-ent between different links in the HTML DocumentModel (DOM) tree [1]. Our method then leveragesthis structure by moving up from the leaves of thelink tree, attempting to group links at their parentnode. If all links could not be merged subject to theconstraint that the sum-of-squared differences (SSD)error of the new grouping is less than a thresholdingfunction.

Our method for classifying web pages based on linkpercentage is highly accurate. We used k -means clus-tering [10] to automatically create unique thresholdsto differentiate index pages and article pages on in-dividual web sites. We also used web page classifica-tion to alter web page display on an accessible webbrowser that we developed to make a usable browsinginterface for users with disabilities. Our method con-sistently outperformed a baseline classifier even whenusing minimal data to generate article and index clus-ters, and achieved classification accuracy of 94.0% onweb sites with slightly malformed HTML, comparedwith 80.1% accuracy for the baseline classifier.

The rest of this paper is organized as follows: sec-tion 2 reviews related work in the area of web pageclassification and point clustering, section 3 describesour accessible web browser, section 4 describes ourmethodology for detecting web context, while section5 examines our link grouping algorithm. In section 6we describe our experiments, section 7 analyzes ourresults, and section 8 discusses possible avenues forfuture work. Finally, section 9 summarizes our find-ings.

2 Related Work

Previous work on classification of web pages into spe-cific types has been limited. It is hoped that this pa-per will spur interest in the use of web context andcomputer context in general and to improve accessi-bility in particular. In the research community, “con-text” has too often referred to the physical context

and can run on any web browser that supports the Document

Object Model.

of the user, such as location. Little research has beendone on the context of the computer environment,which is crucial to understand so that users with dis-abilities and mobile users can effectively utilize ap-plications. For example, references [25] and [33] dis-cussed this notion of computer context.

For mobile users, computer context may be as sim-ple as the currently open applications, while for a per-son with disabilities computer context includes theirdisability, the human-computer interface system theyare using, the applications they normally use, and nu-merous other factors. Harnessing the power of con-text in the web, when we have detailed informationon the current state of the application, is a good placeto start the investigation of this concept.

We define web context as the type of web page theuser is currently viewing. Extending this concept toinclude pages that the user previously browsed inthe current session is beyond the scope of this pa-per, but was examined by Milic-Frayling et al. [21].Methods for determining context or content of webpages vary widely. Cimiano et al. [7, 8] proposed asystem called PANKOW (pattern-based annotationthrough knowledge on the Web) and its derivativeC-PANKOW (context-driven PANKOW). Plessers etal. [26] examined web semantics by proposing a webdesign method combined with the Dante approach toaid visually impaired users through web annotation.The HearSay system of Ramakrishnan et al. [27],in contrast, uses semantic and structural analysis toprovide an audio-based web browser to users with vi-sual impairments. Larson and Gips [19] created a webbrowser for people with quadriplegia that reads webpage text and provides other accessibility options forpeople with disabilities. Gupta et al. [13] determinedweb site context (in their case genre, such as news orshopping) in order to facilitate content extraction.

Kim et al. [18] described a method for segment-ing topics in discussion boards in order to help blindusers more effectively browse the web. Particularlyimportant is that the authors also identified naviga-tional context as an important cue for web browsing,especially for users with disabilities. Milic-Fraylinget al. [21] also explored the notion of context by uti-lizing page importance to implement a web browsingfeature called SmartBack. SmartBack functions sim-

2

Figure 1: Left: The rendering of an index page in our web browser. Notice that the link text is englargedrelative to the plain text to address the problem of mouse clicks generated by dwell time in mouse substitutiondevices. Right: The rendering of the same page in Internet Explorer.

Figure 2: Left: The rendering of an article page in our web browser. All text has been enlarged to enhancereadability. Right: The rendering of the same page in Internet Explorer.

3

ilarly to the common “Back” button on traditionalweb browsers except that it attempts to skip to pagesof high importance in the browsing history ratherthan simply the previously visited web page.

Mobile devices usually have small screens andtherefore have difficulty fully displaying traditionalweb pages. Classification of web pages could assistmethods that use text summarization to facilitateweb browsing on these devices ([4] and [24]) or aidtext summarization methods that drive user inter-faces for people with disabilities [24]. Index pages,for example, typically display text snippets that aresummaries of larger articles, and thus further sum-marizing these snippets is probably unnecessary. Inaddition, for text extraction and news delivery pur-poses knowing the type of page is necessary and anaccurate classification method would further enhancethe accuracy of these systems. Reis et al. [28] clus-tered pages by layout features to attempt to distin-guish between “section pages” and article pages tofacilitate news extraction. Henzinger et al. [16] de-scribed a system that suggests news articles on theweb based on broadcast news. It is evident that clas-sification of a web page as an article would aid thismethodology.

While the web page modification methods that wepresent here preserve all web page content, Guptaand Kaiser [12] extracted content from web pages inorder to render them more accessible. Clearly, ourmethods could be combined with this extraction toallow for even simpler browsing for users with dis-abilities.

Various previous work has been performed in reren-dering web pages for mobile devices. Buyukkoktenet al. [4] presented a number of text summariza-tion methods for display on personal data assistants(PDAs) or mobile phones. The web pages are brokeninto segments of text that can be displayed or hidden.Summaries are constructed from keyword extractionor a determination of significant sentences.

Chen et al. [5] detect and organize a web page intoa two-level hierarchy. Each section of a page is dis-played as a thumbnail that the user can zoom in toview more closely. For pages that are not able tobe split, an intelligent block scrolling method is usedto present the web page. In a similar fashion, Rob-

bins et al. [30] presented a method for navigating amap interface on Smartphone devices by segmentinga map into nine squares corresponding to keys on thephone. Hornbæk et al. [17] analyzed the effectivenessof these zoomable interfaces for the user’s navigationexperience. While this work shows results for nav-igating a map interface, similar conclusions may bedrawn for navigating a zoomable web page displayinterface. This result supports the technique that wechose to alter web page display. Our method essen-tially also “zooms-in” on a web page.

Many alternative user interfaces limit the numberof ways the user can interact with a computer. Var-ious mouse substitution devices, for example, havebeen developed both for people with disabilities andfor other purposes. The EagleEyes [9] project useselectrodes placed around a user’s eyes to detect eyemovements and translate them into mouse pointermovements on a screen. The Camera Mouse [2] tracksa user’s face or other body part to control the mousebased on the user’s movements. These interfaces haveproven very successful with many users with severedisabilities, however fine “pinpoint” control of themouse is difficult. Generating a mouse click requiresthe user to dwell the mouse pointer over the item tobe selected for a short period of time. Given the smallsize of a link in a regular web page, users may havedifficulty navigating web sites as they are normallypresented.

The accessibility problem for web pages comes withthe openness of the web. Web designers are generallyfree to present their information in whatever layoutthey find appealing. Drop down menu bars or click-able image maps may aid in the navigation of a website with the traditional user interface of a mouse, butmay hinder the usefulness of the web site for alterna-tive accessibility interfaces. Sullivan and Matson [32]surveyed accessibility on some of the web’s most pop-ular sites, while Chi et al. [6] presented a method thatautomatically generates a web site usability report.

Richards and Hanson [29] examined web accessibil-ity and how user interfaces should be built to facil-itate interaction for users with disabilities and theelderly. Harper et al. [14, 15] developed mobilityheuristics for visually impaired web surfers and in-corporate this into a web browser plug-in to aid nav-

4

igation. They also investigated the notion of webcontext as it applies to link anchor text, which is of-ten misleading. They posited that in order for webbrowsing to be more efficient, some notion of previewand context of the linked page is necessary.

Link analysis has been employed in many differentcontexts. Lu et al. [20] leverage the structure of theweb graph to remove web pages of low in-degree fromconsideration in the PageRank algorithm and achievehigh speed up with low decrease in accuracy. Foga-ras and Racz [11], in contrast, use links to facilitatesearches of the web graph using graph analysis meth-ods. Bharat et. al [3] similarly use the web graph tostudy the evolving structure of the web.

Duda et al. [10] summarize a number of point clus-tering algorithms and implementation techniques.Particularly relevant to this work is their descriptionof agglomerative hierarchical clustering and cluster-ing in the presence of unkown data structure. Ag-glomerative hierarchical clustering creates clusters bymerging the closest clusters together until the desirednumber of clusters is reached, thus giving the result aminimum variance flavor. They also describe meth-ods for evaluating the validity of cluster splitting byexamining the behavior of a fitness function as thenumber of clusters are increased, stopping splittingonly if the split results in a fitness function increasethat falls below that found by a thresholding func-tion.

3 Web Browser

In our IWeb Explorer web browser [23], we addressedseven problems that users with disabilities had withtraditional web browsers such as Microsoft InternetExplorer and on-screen keyboards:

1. The buttons in the toolbar and scrollbar are toosmall.

2. There is no way to type in the URL address with-out using the keyboard.

3. The links are all very close together in each ofthe pages.

Figure 3: Interface window to confirm that the userwanted to select a link.

4. It is easy to unintentionally end up following thewrong link since there is no check to make surethat the user has not mistakenly selected a link.

5. Each user has different abilities and preferencesand so the button sizes and colors in the key-board and interface windows should not be fixed.

6. The favorites menu is not easily accessible andthe links again are small and are difficult to selectfor people with disabilities.

7. It is difficult to reach keys and links on the farcorners of the computer screen. Therefore ascrolling device is needed to scroll around theweb browser interface.

Screenshots of our browser for both an article andindex page compared to the rendering provided by In-ternet Explorer are shown in figures 1 and 2. The linktext clearly stands out on index pages much more inour browser since it is enlarged to address the prob-lem of mouse clicks generated by dwell time in mousesubstitution devices, while article text is enlarged toenhance readability. This is done automatically us-ing our web context recognition method, which is de-scribed below.

5

Perhaps the most glaring problem of conventionalweb browsers is the lack of an opportunity for theuser to confirm that a link was selected correctly. Ourconfirmation window, shown in figure 3, has yieldeda positive response in preliminary tests with userswith disabilities, and we hope to further improvethis browser by allowing users and their caretakersto change the way that different pages are displayedto suit their individual needs.

When people with disabilities used this webbrowser without our web context component, effec-tive web browsing was not attainable [23]. It there-fore became clear that to make web browsing appli-cations viable for all users, web context needed to beleveraged.

4 Web Context Recognition

4.1 Page Classification

The key observation of our technique is that by ex-amining link percentage we can accurately determinethe “type” of a web page. The link percentage is thepercentage of text characters that are parts of linksas compared to all text characters on a web page.We posit that on sites dealing with news media thatthere are only two types of pages: articles and in-dices, where articles contain mostly text and indicescontain mostly links to articles and other indices. Be-low we refer to pages as “dynamic” if their contentschange from day to day.

While the idea to classify web pages based on linkpercentage seems intuitive, the question is, is it actu-ally feasible to break down pages into categories bythis one-dimensional characteristic? Figure 4 givesan example where a dynamic index page and multi-ple article pages are clearly separable over time. Doesthis mean that a single threshold on link percentagewill be an accurate classifier for all web sites? Fromthe graphs of the link percentage of article and indexpages drawn from four popular web sites over a pe-riod of two weeks shown in figure 5, it is clear that asingle threshold does not suffice. Index pages have ahigher link percentage than articles in most cases, buta single threshold cannot separate these two types of

pages across the web.

Figure 4: The link percentage the BBC web site overa period of 8 days. The index page that was cho-sen was the same, but different articles were chosenat random every day. Observe that there is a cleardistinction in link percentage between the index pageand the article pages.

For the 50 web sites that we tested, the averageratio of link percentage in index pages to link per-centage in article pages was approximately 3, includ-ing malformed HTML characters (for the interestedreader, the web site that had the highest link percent-age ratio in our corpus was that of the Real Madridfootball team, with an average index page link per-centage of 95.3% and an average article page linkpercentage of 0.001%). Some examples from our website corpus are shown in figure 6.

To find these link percentages, we use an HTMLparsing mechanism that works on most web pageswith well-formed HTML. HTML is often not well-formed, however, so we perform some further pro-cessing after the link text and plain text has been ex-tracted from the HTML code by removing as manyextraneous tags as possible. Extra spaces are alsocollapsed to a single character to better represent theactual amount of text that is within a page.

An important issue for our algorithm is the notionof character codes. In particular, for characters suchas ’& ’ and ’<’, the character codes used will be longerthan the actual single character value. Thus we re-place these codes in our extracted text strings withtheir character value.

6

Figure 6: Two web pages included in our corpus as rendered in Microsoft Internet Explorer. Left: Theindex page for the Looksmart FindArticles site (link percentage: 60.2). Right: An article on the LooksmartFindArtilces site (link percentage: 11.5).

The substitution does not apply to pages in lan-guages with different character sets that use charac-ter codes to represent text. As long as the languageis consistent across a page the link percentage dis-tinction remains intact. This is demonstrated by ourresults from some Japanese news web sites, shown infigure 7. The index and article pages are still separa-ble, implying that our methodology works correctlyhere as well.

Once we have parsed a web page’s HTML code anddetermined the link percentage, the issue of deter-mining a proper threshold arises. If the user has notvisited other web pages from this site, then we canonly use generic thresholds to determine the type ofa page. An initial threshold of 0.4 was used in exper-iments and found to perform reasonably well. Afterthe user has visited at least one page of each type,however, we can begin to discover how the link per-centage values of index pages and article pages areclustered. Using the k -means clustering algorithmfor each web site, with k = 2, one cluster for articlepages and one cluster for index pages, we can accu-

rately classify future web pages from this site. Weuse as the initial mean points of the cluster the pageswith the lowest and highest link percentage for thearticle and index clusters, respectively. We then runthe k-means algorithm to determine the final clus-ters. Using these clusters, we choose as our decisionthreshold the value midway between the link percent-age of the page with the highest link percentage inthe article cluster and the link percentage of the pagewith the lowest link percentage in the index cluster.An example where clusters are separated by a thresh-old computed in this way is given in figure 8. Thek-means algorithm can be viewed as a method to ap-proximate the maximum-likelihood estimates for themeans of the clusters.

We observed that even dynamic web pages’ linkpercentages do not fluctuate very much over shortperiods of time. We studied four popular websitesover the course of seven months and found that thelink percentages for the same index pages had a to-tal range of less than eight percent. Therefore, oncea page is visited its link percentage is stored in a

7

Figure 5: The link percentage for various sites over aperiod of 8 days. The examined index pages are dy-namic and typically change every day. Different arti-cles at these sites were chosen at random every day.While it is apparent that index pages have higher linkpercentages than articles, there is not one thresholdthat can separate these two types of pages for all websites, as shown by the MSN article link percentagerising above the link percentage for the BBC indexpage. It is also important to notice that, except forthe Yahoo index page, the dynamic index pages’ linkpercentages do not change by more than 3%. Thebehavior of the Yahoo index page’s link percentage isdue to malformed HTML.

database and is retrieved if the page is visited againand no HTML parsing is performed. This saves com-putational effort and can easily be overridden by theuser if the classification results falter because of achange in the web site structure.

Figure 7: The classification accuracy of our methodon three Japanese web sites. Six pages, three indicesand three articles, were used for testing, while a singleindex page and a single article page were used tocreate the two clusters.

Figure 8: The optimal threshold for link percentageon the Yahoo Sports web site given the link percent-ages of three index pages and three article pages.

4.1.1 Database Storage

Each user’s database of visited web pages and websites is stored on the user’s computer in a hash tablewhich is processed by the web context program priorto browsing. We could, alternatively, have stored thelink percentages of pages and web site thresholds inan online database. In this scheme, users would firstquery an online database system with the URL of thepage that they will download. The database would

8

either respond with that page’s classification or statethat the page is not in the database. If the page wasnot in the database, the user could then upload theweb page’s link percentage after the page is loaded.

Clearly, there are problems with this approach.Privacy concerns rank chief among these, since theuser would be informing the database of their brows-ing behavior at all times. In addition, user-definedbehavior, such as changing the classification of a cer-tain web page for individual reasons, played a factorin our decision to keep the database on the local ma-chine. Finally, an online database would introducewasted computational effort employed in searchingsuch a large database since most users visit a nar-row band of web sites during the course of normalbrowsing [22].

Clusters for individual web sites are also stored lo-cally for similar reasons, and since clusters are com-puted extremely quickly and only when the user endstheir browsing session this imposes a minimal com-putational burden on the user’s system.

4.1.2 URL Analysis Feature

We initially attempted to use analysis of the web pageURL as a fast, effective indicator of page type. Thismethod evaluates a series of conditions to classify aweb page (a more sophisticated URL analysis methodis presented in [31]). If the page URL satisfies any ofthe conditions, then it is classified as an index. Oth-erwise, it is classified as an article. The conditionsare:

1. The web page has the same URL as theroot of the URL (for example, the root ofwww.cnn.com/WORLD/ is www.cnn.com).

2. The URL ends with a “/”, indicating that it isthe index of some subdirectory.

3. The URL ends with “index.*”, where “*” is avalid file name extension (such as .html, .jsp,.psp).

This approach yields fairly good results in certaincases, but it fails in a number of situations, since URLnaming conventions appear to differ widely across the

web. Particularly troublesome are articles where theURL satisfied the third constraint since the web pagerepresented a summary of all news pertaining to acertain event, such as the URL for an article on apolitical story on CNN.com: http://www.cnn.com/... /delay.indiantribes.ap/index.html. Even if we re-move step 3, however, numerous index pages werestill falsely classified. Thus it was determined thatthe far more accurate link percentage analysis waspreferable, since the pages that the URL analysismethod faltered on were the very pages that oftenpresented trouble for the k -means clustering method.

4.2 Customized Page Display

Our web browser enlarged link text on index pages tosupport mouse substitution interfaces that use dwelltime to generate mouse clicks. The browser also en-larged plain text on article pages to enhance read-ability. This is just one possible display modifica-tion, and for certain web pages this may or may notbe useful. The user can undo these display modifi-cations if they wish, putting the ultimate decision ofthe page’s display in their hands (see figures 1 and 2for a comparison of our browser’s rendering of a webpage using web context to that of Internet Explorer).

We could further enhance the user’s interaction ex-perience by changing the way that keyboard or mousecommands are issued depending on the type of webpage that the user is viewing in order to facilitatenavigation. Moreover, we could provide the user withtools to create their own rules for modifying web pagedisplay based on page type. This kind of control isextremely crucial for people with disabilities. Withmany of the currently available assistive interface sys-tems web browsing is still difficult. For example, us-ing an interface system such as the Camera Mouse [2]to select links or scroll down a page is hard even forusers without disabilities. With our method, theseusers could scroll down an article web page merelyby moving the mouse pointer to the left half of thescreen, or iteratively cycle through links by perform-ing the same action on an index page. It is our hopethat this method can alleviate some of the accessibil-ity problems that people with disabilities have withcurrent interface systems. We also hope that future

9

work will place more emphasis on the context of theiractions to enrich the interactive experience and makeit more effective and efficient.

Another aspect of customized page display is theuse of manual corrections if a user has a preference fora different display than provided by our method orif the web page was misclassified, which occurs whenpages were created with malformed HTML (see be-low). We decided that the best course of action is tosimply omit the page in question from the clusteringalgorithm altogether.

5 Link Grouping

Link grouping allows users with severe disabilities touse a scan-based mechanism to tab through a webpage and select item. This can substantially im-prove the user’s rate of communication, and couldbe applied to mouse substitution interfaces by allow-ing users to click on a group of links so they can moreeasily select the desired single link.

5.1 Link Tree Creation

The link tree creation step of our link grouping al-gorithm creates the framework under which we cancluster points according to their location on the webpage as well as their location in the HTML code. Es-sentially our method leverages the structure of theHTML code, which can be quite nicely representedin tree form by allowing an element to be a node ina tree and the elements that it encapsulates to beits children. An example link tree along with theHTML code that it was created from is picture infigure 9. The tree representation allows us to employa divide-and-conquer method from grouping as de-scribed below. Our method can also use this tree toconstrain link grouping so that groups do not spaninappropriately across the web page. Note that it isnot that such across-page grouping is incorrect per se,rather that it would create a very unnatural groupingconsisting of circles of links that did not respect thestructure of the web page. In addition, this wouldleave us with a very unconstrained clustering prob-lem, and we certainly prefer an approach that can ap-

ply the divide-and-conquer paradigm to the groupingproblem.

The pseudocode for link tree creation is shown inalgorithm 1. Our algorithm, CreateLinkTree, takesas input an HTML DOM tree and outputs the linktree. The algorithm starts at the links of the DOMtree and traverses it until it reaches the root node ofthe document, marking every node that it visits alongthe traversal (line 9). If, however, it comes acrossan already marked node, our algorithm converts thatnode into an internal node in the link tree, connectedto the link nodes that already marked it (lines 5 and6). If that link node has already been incorporatedinto a subtree of the link tree, that subtree becomesthe current node’s child. Traversal of the DOM treethen stops. The last case that can occur is that alink traversal will arrive at a node that is alreadyan internal node in the link tree. Here the link willsimply add itself as another child to that node andstop traversal.

In the worst case, the links do not intersect untilthe “<BODY>” node of the DOM tree (since thisnode must encapsulate all links). Then, if there aren links, and the height of the DOM tree from thelink nodes is log(q), where q is the number of ele-ments in the DOM tree, it is clear that the link treecreation algorithm is bounded by O(nlog(q)), sinceour method must make n traversals of length log(q).Note that typically n � q.

5.2 Link Group Creation

The previous step of our algorithm, that of link treecreation, has now given our method the machineryto perform link grouping. For all of the clusteringsteps below, our algorithm, GroupLinks, uses a setof points defined by the position of the links on therendered web page in cartesian coordinates. In gen-eral, our algorithm traverses the link tree from thetop down and attempts to merge the clusters of linkpoints of one of its child nodes with those of anotherchild node if each child has only one link cluster. Theoptimal number of clusters is chosen by the crite-rion function defined below. Pseudocode for the linkgrouping algorithm is given in algorithm 2.

More specifically, using the link tree as a struc-

10

Figure 9: The left figure represents the link tree for the HTML code on the right. At the leaves are the linknodes, and the internal nodes are the first common parent of its children in the DOM tree. The parents ofeach node in the DOM tree that are not represented as internal nodes of the link tree are shown in blue.This example can best be understood by examining the HTML code and viewing the structure of the DOMtree. Observe that the first common parent of the link to “bar.html” and the link to “boo.html” is the“<P>” node. Therefore this is their parent in the link tree. Next, note that the first common parent of thelink to “foo.html” and the “<P>” node is the “<BODY>” element, which is the root of the tree.

11

Algorithm 1 CreateLinkTree

Input: HTML DOM Tree TOutput: Root of the Link tree

1: for each link a in T2: traverser = a3: while traverser.parent != root4: traverser.parent.child = a5: if marked(traverser.parent)6: make node(traverser.parent)7: break

8: end if

9: mark(traverser.parent)10: traverser = traverser.parent11: end while

12: end for

13: return root

tural guide, the algorithm recursively breaks downthe clustering problem into that of merging the linkpoint clusters of children nodes together, starting atthe root node, by a criterion function. If our methodcannot merge all of the children of some internal nodetogether, then, intuitively, these link groups shouldbe excluded from merging with other groups at higherlevels in the tree. As stated earlier, doing so wouldviolate our constraint of respecting the structure ofthe web page.

At the current node, call it node i, the method firstruns the link grouping algorithm on all of its children(line 4). If node i has no children, it is a link andthus simply returns its position to its parent as a sin-gle cluster (line 2). Otherwise, the algorithm checkswhich of node i’s children can be merged (line 7),which is the case if each child returned only a sin-gle cluster of points. For all of node i’s children thatcan be merged, our method runs the agglomerativehierarchical clustering algorithm on the mean pointsof each of its children’s clusters (line 8). This algo-rithm essentially merges the closest clusters at everystep. As stated above it is for this reason that wecan expect the clusters to have low variance. Assum-ing that there are c children of node i, this gives ouralgorithm a total of c clusters before i is processed.

The next step is to determine the optimal num-

ber of clusters. We do this using the equations (dueto [10]):

J(k + 1)

J(k)< 1− 2

dπ− α

√

2(1 − 8/π2d)

nd(1)

where J(k) is the SSD error for k clusters, d is thedistance between the means of the clusters that weresplit to create b + 1 clusters from b clusters, and n isthe number of points in all clusters. The parameterα is determined by solving the equation:

α =√

2 × erf−1(1 − p

2) (2)

where erf is the Gauss error function, defined by theequation:

erf(x) =2√π

x∫

0

e−t2

dt (3)

Here, p is the significance level at which we believethat we have at least k + 1 clusters. Our algorithmstarts with k = 1, stopping once the inequality inequation 1 is violated. It then returns the resultingclusters.

Naturally, having k + 1 clusters will yield a lowerSSD error than k clusters. Essentially, what equa-tion 1 does is model the clusters as k different normaldistributions and check that the error reduction thatwe see is not due to chance at the p significance level,since if there actually were only k clusters we wouldexpect any other clusters that formed to be formedby chance.

Once the final clusters have been returned to theroot node, we can modify the web page to make itmore accessible to users with disabilities. While thismodification mechanism can be accomplished by var-ious parameterized functions, in our implementationwe choose to use color to identify links in the samegroup, using different colors for different groups. Thisis shown in figures 10 and 11. Other options includelink enlargement when the mouse cursor hovers overa link group, making all links in a group lead to apage where the links are made very large for easynavigation, and many other possitibilities which wewill explore further in the Future Work section.

12

Figure 10: A Wikipedia (http://www.wikipedia.org) web page that has been processed by our link groupingalgorithm. Note that the different link groups are in different colors so that the user can easily pick outdifferent groups. In this page, shown only partially here, 66 clusters were found.

Figure 11: Left: The link tree for the web page on the right. The links in the tree have the color that theywere given in the web page after they were processed by our link grouping algorithm. Right: An exampleweb page processed by the link grouping algorithm.

13

Algorithm 2 GroupLinks

Input: Link Tree Node R,Significance pOutput: Link Groupings for TreeRooted at R

1: if isLeaf(R)2: return R.position3: end if

4: for each child c of R5: groups(c) = GroupLinks(c)6: end for

7: mergeGroups = { group(c) } s.t.8: num clusters(group(c)) = 19: create hierarchical clusters(mergeGroups)

10: current clusters = 111: while equation 1 is not satisfied12: current clusters++13: end while

14: return clusters(current clusters) +Σ (num clusters(group(c)) |num clusters(group(c)) 6= 1) ]

The link grouping creation algorithm is boundedby O(n3) in the worst case, where n is the number oflinks, since our method must find the closest group atevery step of hierarchical clustering, for n steps. Notethat this worst case is only realized if every link hasonly one common parent at the root of the link tree,which does not happen in practice since that wouldrequire essentially no other content on the web page.Since the link tree exhibits the structure of a treewith branching factor b in the average, the algorithmis bounded by O(b2logb(n)), since the height of thelink tree is logb(n) and the algorithm needs to cre-ate b clusters at each level of the tree. If no linkscan be merged, the algorithm stops at the leaves ofthe link tree and is bounded by O(b2). In practice,we measured a runtime that falls somewhere in be-tween the last two bounds that we derived, since typ-ically we can merge only some of the link point clus-ters at each node. Note that the runtime dependsheavily upon the p used in equation 2, since p essen-tially bounds the size of a cluster in 2D space. Thecomplete algorithm, including the creation of the link

tree, then, for the bound on the link grouping algo-rithm of O(b2logb(n)), is simply O(nlog(q)), as longas we have b � n, which is again normally the case.

Suppose that our user is using a simple tab inter-face which requires the user to press the tab buttononce to move to the next link and enter to click onthe link. Using our grouping algorithm, this usercould select a specific group and then an individuallink. Assuming that there are c final clusters andthat each group has an average of s links, the aver-age number of tabs required to click on a link wouldbe c+s

2. In comparison, the tab-based interface that

is currently employed on commercial web browsershas an average number of tabs of n

2. Therefore, our

method improves the communication rate of users bya factor of:

n

c + s(4)

Note that if s is too large, we could simply split eachcluster into subgroups in order to minimize the num-ber of tabs required and maximize equation 4, butthat is left to future work. As a precaution, if onlyone cluster is created then the grouping is not used,since this would result in requiring the user to pressthe tab button once just to be able to perform theoriginal selection task. Clearly, in the worst case,where the number of clusters equals the number oflinks, our method performs as well as the current tab-based implementations on web browsers, and due tothe trivial computational cost of our grouping algo-rithm (the highest number of links in our web corpusof 300 pages was 125, and the link grouping programran nearly instantaneously), our method could be anintegral component for an accessible web browser, orany web browser in general.

6 Experiments

To test our web context method we used it to clas-sify web pages from a corpus of the top 25 news andtop 25 sports web sites as rated by Alexa Web Search(URL: www.alexa.com). For each web site, three in-dex pages and three article pages, as categorized bya human observer, were used for testing for a total of

14

300 web pages. One index page and one article pagefrom each web site were randomly chosen to createan index cluster and an article cluster in the web con-text program, and these pages were not included inthe test set.

We compared our method to a static thresholdtechnique, which used a predetermined global thresh-old (a link percentage of 0.4) to differentiate betweenindex pages and articles. We treat this classifier asthe baseline in our discussion of results, since it isthe simplest solution to the classification problem.We also compared our method to an “optimal” clas-sifier, which chooses the best possible classificationthreshold for each web site. This is merely a theoret-ical classifier; given complete knowledge of what thecorrect classifications are, it finds the optimal classi-fication threshold.

There were many web pages that containedseverely malformed HTML, as evidenced by the factthat even the optimal classifier did not generate 100%accuracy on every web site. The results are shown infigures 12, 13, and 14, broken down by site category(news or sports) and combined. The average classifi-cation accuracy for each method is shown in table 1,and a graph comparing the static threshold methodand the clustering method is shown in figure 15. Infigures 12, 13, 14, and 16 accuracy of 100% impliesthat all pages in the test set for a particular web sitewere classified correctly, while 83% implies that fiveout of the six pages in the test set were classifiedcorrectly, and so on.

To evaluate our link grouping method, we testedit on a number of web pages from our original cor-pus using a signifigance level p = 0.001, which keptthe link groups to a reasonable size. We discuss theresults of these experiments below.

Figure 12: The classification accuracies for the threedifferent classifiers on 25 news web sites, containinga total of 150 web pages.

Figure 13: The classification accuracies for the threedifferent classifiers on 25 sports web sites, containinga total of 150 web pages.

15

Table 1: Classification accuracy of three classifiers as the fraction of the number of correctly labeled webpages out of 300 test web pages.

Static Threshold Clustering Optimal ClassifierNews Pages 0.734 0.840 0.946Sports Pages 0.700 0.760 0.866All Pages 0.717 0.800 0.906

Figure 14: The classification accuracies for the threedifferent classifiers on all 50 web sites containing atotal of 300 web pages.

Figure 15: A comparison of the static threshold clas-sifier and the clustering method. As the HTML pars-ing improves, evidenced by the higher accuracy of theoptimal classifier for the news web site data set, sodoes the performance of our clustering method.

7 Results and Discussion

7.1 Web Context

In our experiments, the clustering method gavehigher performance than the static threshold method,correctly classifying 80.0% of the test pages comparedto the static threshold’s 71.7%. Our method’s accu-racy, however, is clearly not equivalent to the opti-mal classifier, which classified 90.6% of the pages cor-rectly. The other 9.4% of pages had highly malformedHTML code, since when the experimenter hand la-beled the link and plain text, these pages were foundto cluster in the expected fashion and often the handlabeled link percentages differed from those given bythe parsing program by over 40%.

It is also interesting that our method achieved bet-ter results than the static threshold 38% of the time,while only having poorer performance in 14% of ourexperiments. This result is encouraging because itshows that we are consistently outperforming thebaseline classifier and reaching a level of performancethat is close to that of the optimal classifier.

It also makes sense to ask how the static thresh-old and clustering methods compare when all websites where the optimal classifier did not achieve100% accuracy are removed. Then we see the clus-tering method achieved 94.0% accuracy, while thestatic threshold technique rose to only 80.1% accu-racy. These results are shown in figure 16. Theclustering method performed better than the staticthreshold on 47% of the web sites, and the staticthreshold method performed better on only 9.4% ofsites. This indicates that as HTML parsing accuracyincreases, our clustering method gets closer to 100%accuracy. Indeed, there are still some parsing errorsleft over in this group of sites, only less of them and

16

of smaller magnitude. It seems clear that combin-ing our method with an “HTML cleaning technique”such as that introduced in [34] would yield extremelyhigh levels of accuracy.

Figure 16: The classification accuracy for the staticthreshold and clustering methods on sites for whichthe optimal classifier achieved 100% accuracy. Ourclustering method correctly classified 94% of webpages, while the static threshold method only clas-sified 80% correctly.

The strength of our method is that it finds theproper threshold between index and article pages ina web site given little data. The fact that we do notknow a priori which web pages are indices and whichare articles poses a problem if the user visits onlypages of a single type. The algorithm will then as-sume that one of the pages is in fact an index and thuserroneous clusters will emerge until the user visits apage of the other type. This is not a problem if theuser can turn off our web page classification methodfor certain sites that do not have different kinds ofpages. It is important to make the user aware of thiscaveat, lest they prematurely turn off the algorithmfor sites where it would work appropriately if it hadmore information.

With regards to our clustering algorithm, it maybe desirable for manual corrections to force the ini-tial mean points to be the mean of the web pagesthat were previously classified as articles. This comeswith the danger of an initial poor classification thatwas not corrected corrupting future classification at-

tempts, so we decided against this modification.

Note that, in general, users will collect more datafrom each site over the course of normal browsing,so we would expect even better results than thosereported here. Our experiments are meant to showmerely the bare minimum of what our method is ca-pable of.

We acknowledge that we could have used otherfeatures such as periods, punctuation character per-centage, or word rather than character link percent-age. But these do not capture the page type dataas accurately as link percentage. Period and punc-tuation percentages perform poorly because they areoften used as delimiters within a page, and occasion-ally index pages have many small text snippets, suchas when small summaries for different articles arepresent, giving these pages more punctuation char-acters than an article page. Word link percentage,or the percentage of words that were part of linksdivided by the total number of words, on the otherhand, treats words such as “a” and “interestingly” asequally weighted words, when clearly “interestingly”takes up more space than “a” on the rendered pageand so contributes more to its type. The two pagetypes appeared less separable using this feature, sowe opted in favor of the (character) link percentagecue, although admittedly using word link percentagemay blunt the influence of malformed HTML.

It also may be unsettling that we ignore images,display markup, and position information. While werealize that these are important parts of a page’s con-tent, it is difficult to develop rules that would gener-alize to the entire web, since some articles have manyimages while index pages have very few (see, for ex-ample, figure 17). Some images, however, are usedas links in place of text. It is unclear exactly howinfluence should be computed for these images, butthese images may prove useful as an additional cue ina future extension of our system, although currentlythey are ignored during processing.

In addition, text markup is used in many differentways with rather loose rules governing their use, andgiven that our classification performance is extremelyhigh, it does not seem that the extra processing re-quired would generate large gains in accuracy.

17

Figure 17: Left: The index page for the BBC News site. Right: An article on the BBC News site. Both webpages have three images, so a classification algorithm based solely on number of images would not work inthis case.

7.2 Link Grouping

The results for link grouping varied widely, with thereduction in tabbing according to equation 4 span-ning from a factor of 40 to an improvement of only1.5, with a mean of 12. In addition, our algorithmon average took 0.2 seconds to run, with the longesttime at 0.8 seconds. Thus our method clearly runsin real time and provides real performance gains forthe user.

Our link grouping algorithm has shown itself tobe quite versatile and effective, working across webpages in multiple languages with a myriad of lay-outs. While is difficult to state what the “correct”link grouping would be in an objective manner, per-sonal experience with the algorithm has shown thatit does indeed respect the layout of the page andprovides very intuitive groupings on most pages. Asa bonus, our method is easy to plug in to any webbrowser, since it was built entirely in code that canbe inserted directly into a web page by a browser.

Actual modification of web page display, however,did not receive as much attention as the algorithmitself. There are innumerable possibilities for modifi-cations, and these vary drastically with the intended

audience. We will explore some of these possibilitesbelow.

8 Future Work

Extracted HTML text characters can clearly be usedto form a very accurate classification algorithm, butin order to push accuracy higher we may need to useother cues. Using rendering data to weight text ac-cording to its centrality in the displayed page (i.e.weighting text that is closer to the top and middlehigher than text that is more towards the sides andbottom) appears to be an attractive extension, al-though it is not clear if a general rule can be devel-oped that works for a broad segment of web sites.

We may also wish to handle pages that contain ahigh volume of images used as links rather than textlinks. While not encountered in our test corpus, han-dling of such pages is crucial, and perhaps assigninga default weight to each image-link would further im-prove results.

Detecting web pages and sites that are merely Javaapplets or Flash programs is also important, since wecan no longer determine the optimal mode of interac-

18

tion with the web page. If, however, we allow the userto specify what mode of interaction to use when theyvisit this page, then we can contact the interface sys-tem each time this page is visited again and instructit to output commands according to the user’s speci-fication. This is a very useful feature that will likelybe implemented in our web browser in future work.

It may also be useful to segment the rendered pageinto different regions using a decomposition methodsuch as that introduced by Chen et al. [5] and thenclassify each of the regions using our method. Theclassification of the entire page could come from aweighted sum of the classifications of each region.This decomposition could also aid in interaction mod-ification, since we can imagine displaying regions ofdifferent types in different ways and changing the userinteraction method if they select a particular region.This would incur a higher computational cost, but itmay be a necessary extension to further utilize webcontext on PDA-class devices.

We could extend our approach to web page classi-fication into the image processing realm by using abitmap image of the rendered page to classify text re-gions and other regions using pixel information only.This would be insulated from many of the problems ofparsing malformed HTML, but this type of algorithmwould be computationally expensive and require thepage to be rendered before it is altered, placing fur-ther burden on the user. It may be useful to combinethis image-based method with our current classifica-tion scheme, however, to yield a more robust estimateof page type.

One issue that we touched upon earlier was sitesthat have only one page type or web sites where onlypages of one type are visited. To handle this case au-tomatically it may be necessary to first attempt to fita single cluster to all web pages on a site and then seeif the fit is acceptable. If not, then the regular algo-rithm can resume. Characterizing what constitutesa “good” fit may prove troublesome and complicatesthis technique.

There are also web sites which do not fall withinthe domain of sports or news web sites that may havemultiple types of pages. A major component of fu-ture work is to identify these page types and examineif they generalize across a wide range of web sites as

the article and index types do. If it appears thatnew page types provide a nice fit for a wide rangeof web sites, incorporating these types into a futurealgorithm would be a definite possibility. It may be,however, that beyond the basic article-index distinc-tion different interaction modes and display modifi-cations are not useful. This issue clearly demandsfurther research.

We also wish to extend our web browser to givemore control to the user in displaying different typesof web pages. We are experimenting with variousways to offer this functionality, and it will surely cre-ate a greatly enhanced interaction experience for theuser. Detailed experiments on how page display mod-ification and changes in the interaction mode posi-tively impact the user interface experience will alsobe performed to further validate our results. This isparticularly relevant to our link grouping algorithm,which could allow users to click on a link group to en-large that group of links, or to highlight a link groupin a different color as it’s selected. What options aremost user-friendly and intuitive would make for in-teresting future research.

Our work is a preliminary step into the larger in-vestigation of computer context. In later work wewould also like examine other types of computer con-text and investigate whether or not extending webcontext to include previously browsed pages is feasi-ble. Work in this area has only just begun.

9 Conclusion

We have presented a highly accurate method for clas-sifying web pages based on link percentage. Our k -means clustering method created unique thresholdsto differentiate index pages and article pages on in-dividual web sites. Accuracy increased when we re-moved web sites from the corpus that had extremelymalformed HTML, and it is expected that more ro-bust HTML parsing will yield even more accurate re-sults. Our method consistently outperformed a base-line classifier even when using minimal data to gen-erate initial article and index clusters.

Our link tree creation algorithm and link group-ing method have been shown to be quite effective

19

and guaranteed to outperform or at least stay at thesame level as the functionality offered by current webbrowsers, leading to a possible improved communi-cation rate for users with disabilities. This methodis fast and is portable to nearly all available webbrowsers, giving it promise to become an integral toolfor web accessibility.

We also used web page classification and linkgrouping to alter web page display on a web browser,and future work will center around giving the usermore control in determining how different types ofweb pages are displayed and choosing intuitive waysto change interaction modes of an interface systembased on web page classification and link grouping.

References

[1] Document object model (dom) level 1 specifici-ation version 1.0. W3C Recommendation, 1998.

[2] M. Betke, J. Gips, and P. Fleming. The Cam-era Mouse: Visual tracking of body features toprovide computer access for people with severedisabilities. IEEE Transactions on Neural Sys-tems and Rehabilitation Engineering, 10(1):1–10, Mar. 2002.

[3] K. Bharat, B. Chang, M. Henzinger, andM. Ruhl. Who links to whom: Mining linkagebetween web sites. In International Conferenceon Data Mining (ICDM), pages 51–58, 2001.

[4] O. Buyukkokten, H. Garcia-Molina, andA. Paepcke. Seeing the whole in parts: text sum-marization for web browsing on handheld de-vices. In Proceedings of the 10th InternationalWorld Wide Web Conference (WWW), pages652–662, Hong Kong, 2001.

[5] Y. Chen, W.-Y. Ma, and H.-J. Zhang. Detect-ing web page structure for adaptive viewing onsmall form factor devices. In Proceedings of the12th International World Wide Web Conference(WWW), pages 225–233, Budapest, Hungary,2003.

[6] E. H.-H. Chi, A. Rosien, G. Supattanasiri,A. Williams, C. Royer, C. Chow, E. Robles,B. Dalal, J. Chen, and S. Cousins. The blood-hound project: automating discovery of webusability issues using the infoscent simulator.In Computer-Human Interaction 2003 Confer-ence on Human Factors in Computing Systems(CHI), pages 505–512, 2003.

[7] P. Cimiano, S. Handschuh, and S. Staab. To-wards the self-annotating web. In Proceedingsof the 13th International World Wide Web Con-ference (WWW), pages 462–471, New York City,2004.

[8] P. Cimiano, G. Ladwig, and S. Staab. Gimme’the context: Context-driven automatic semanticannotation with C-PANKOW. In Proceedings ofthe 14th International World Wide Web Con-ference (WWW), pages 332–341, Chiba, Japan,2005.

[9] P. DiMattia, F. X. Curran, and J. Gips. AnEye Control Teaching Device for Students with-out Language Expressive Capacity – EagleEyes.The Edwin Mellen Press, 2001. See alsohttp://www.bc.edu/eagleeyes.

[10] R. . Duda, P. E. Hart, and D. G. Story. PatternClassification. Wiley-Interscience, 2nd edition,2001.

[11] D. Fogaras and B. Racz. Scaling link-based sim-ilarity search. In Proceedings of the 14th Inter-national World Wide Web Conference (WWW),pages 641–650, Chiba, Japan, 2005.

[12] S. Gupta and G. Kaiser. Extracting contentfrom accessible web pages. In Proceedings of the14th International World Wide Web Conference(WWW), pages 26–30, Chiba, Japan, 2005.

[13] S. Gupta, G. Kaiser, and S. Stolfo. Extract-ing context to improve accuracy for html con-tent extraction. In Proceedings of the 14th Inter-national World Wide Web Conference (WWW),pages 1114–1115, Chiba, Japan, 2005.

20

[14] S. Harper, C. Goble, and R. Stevens. Travers-ing the web: Mobility heuristics for visuallyimpaired surfers. In Proceedings of the FourthInternational Conference on Web InformationSystems Engineering (WISE 03), pages 200–209,Rome, Italy, 2003.

[15] S. Harper, Y. Yesilada, C. Goble, andR. Stevens. How much is too much in a hy-pertext link: Investigating context and preview– a formative evaluation. In Proceedings of thefifteenth ACM conference on Hypertext and hy-permedia (HT’04), pages 116–125, Santa Cruz,CA, USA, 2004.

[16] M. Henzinger, B.-W. Chang, B. Milch, andS. Brin. Query-free news search. In Proceedingsof the 12th International World Wide Web Con-ference (WWW), pages 1–10, Budapest, Hun-gary, 2003.

[17] K. Hornbæk, B. B. Bederson, and C. Plaisant.Navigation patterns and usability of zoomableuser interfaces with and without an overview.ACM Transactions on Human-Computer Inter-action, 9(4):362–389, Dec. 2002.

[18] J. W. Kim, K. S. Candan, and M. E. Donderler.Topic segmentation of message hierarchies for in-dexing and navigation support. In Proceedingsof the 14th International World Wide Web Con-ference (WWW), pages 322–331, Chiba, Japan,2005.

[19] H. Larson and J. Gips. A web browser for peoplewith quadriplegia. In 10th International Con-ference on Human-Computer Interaction, Crete,Greece, 2003.

[20] Y. Lu, B. Zhang, W. Xi, Z. Chen, Y. Liu,M. Lyu, and W.-Y. Ma. The powerrank weblink analysis algorithm. In Proceedings of the13th International World Wide Web Conference(WWW), pages 254–255, New York City, 2004.

[21] N. Milic-Frayling, R. Jones, K. Rodden,G. Smyth, A. Blackwell, and R. Sommerer.SmartBack: Supporting users in back naviga-tion. In Proceedings of the 13th International

World Wide Web Conference (WWW), pages63–71, New York City, 2004.

[22] A. Montogmery and C. Faloutsos. Indentifyingweb browsing trends and patterns. Computer,pages 94–95, 2001.

[23] M. Paquette, M. Betke, and J. Magee. IWeb Ex-plorer: A web browser designed for use with aneye controlled mouse device. In Boston Univer-sity Computer Science Department MA ThesisReport, 2005.

[24] B. Parmanto, R. Ferrydiansyah, A. Saptono,L. Song, I. W. Sugiantara, and S. Hackett.AcceSS: Accessibility through simplification &summarization. In Proceedings of the Sec-ond International Cross-Disciplinary Workshopon Web Accessibility (W4A2005), pages 18–25,Chiba, Japan, 2005.

[25] J. Pitkow, H. Schutze, T. Cass, R. Coo-ley, D. Turnbull, A. Edmonds, E. Adar, andT. Breuel. Personalized search. Commun. ACM,45(9):50–55, 2002.

[26] P. Plessers, S. Casteleyn, Y. Yesilada, O. D.Troyer, R. Stevens, S. Harper, and C. Goble.Accessibility: A web engineering approach. InProceedings of the 14th International WorldWide Web Conference (WWW), pages 353–362,Chiba, Japan, 2005.

[27] I. V. Ramakrishnan, A. Stent, and G. Yang.HearSay: Enabling audio browsing on hyper-text content. In Proceedings of the 13th Inter-national World Wide Web Conference (WWW),pages 80–89, New York City, 2004.

[28] D. Reis, P. B. Golgher, A. S. da Silva, andA. H. F. Laender. Automatic web news extrac-tion using tree edit distance. In Proceedings ofthe 13th International World Wide Web Confer-ence (WWW), pages 502–511, New York City,2004.

[29] J. T. Richards and V. L. Hanson. Web acces-sibility: A broader view. In Proceedings of the

21

13th International World Wide Web Conference(WWW), pages 72–79, New York City, 2004.

[30] D. C. Robbins, E. Cutrell, R. Sarin, andE. Horvitz. ZoneZoom: Map navigation forsmartphones with recursive view segmentation.In Proceedings of the working conference on Ad-vanced visual interfaces, pages 231–234, Gal-lipoli, Italy, 2004.

[31] L. Shih and D. Karger. Using URLs and tablelayout for web classification tasks. In Proceed-ings of the 13th International World Wide WebConference (WWW), pages 193–202, New YorkCity, 2004.

[32] T. Sullivan and R. Matson. Barriers to use:Usability and content accessibility on the web’smost popular sites. In Proceedings of the 2000Conference on Universal Usability, pages 139–144, Arlington, Virginia, USA, 2000.

[33] T. Winograd. Architectures for context. Human-Computer Interaction, 10(24):401–419, 2001.

[34] L. Yi, B. Liu, and X. Li. Eliminating noisy infor-mation in web pages for data mining. In KDD’03: Proceedings of the ninth ACM SIGKDD in-ternational conference on Knowledge discoveryand data mining, pages 296–305, Washington,D.C., USA, 2003.

[35] X. Yin and W. S. Lee. Using link analysis toimprove layout on mobile devices. In Proceed-ings of the 13th International World Wide WebConference (WWW), pages 338–344, New YorkCity, 2004.

22