In this paper we illustrate the use of Named Entities as a basis to share associative trails [1] between Semantic Desktops using a Firefox [6] extension called Trailblazer. Trailblazer can automatically detect the context of a Web page. It identifies the most relevant category for the Web page based on its context along with relevant links to additional information and notes. Named Entities belonging to that category are highlighted. Trailblazer facilitates the building and augmentation of categories with links from Named Entities to Web pages, documents, mail messages, personal notes, etc. It enables the import/export of categories and their related associative trails between users. It facilitates the distribution of discovered information through the sharing of associative trails in order to support community based knowledge. This allows faster initiation of new members to a group and information noted by one member as important can be highlighted to the rest of the group.
We describe Trailblazer in terms of how a user, Ted, might experience it for the task of collecting information on horse racing. Then we extend the scenario further to include how he shares his collected information with other enthusiasts. He also uses Trailblazer to create a separate category of trails about wine to share with his friends.
Ted enjoys horse racing and to further his interest joins a horse racing club. The club consists of a group of knowledgeable racing enthusiasts who share their knowledge between themselves. They maintain this knowledge using a Firefox extension called Trailblazer which allows them to identify Horses, Trainers, Jockeys, Racecourses, etc. as Named Entities. They add associative trails to these Named Entities in the form of links to relevant Web pages and notes that they write. When Ted joins the club he is given Trailblazer and access to the club’s associative trails. The Firefox browser has been altered to include the Trailblazer panel at the bottom of the screen as seen in Figure 1. He follows the club’s instructions to import their horse racing trail and visits his favorite online racing news Web site, ‘Irish Racing’. He can see from the Trailblazer panel that the page has been categorized as ‘Horse Racing’ as shown in Figure 1.
As he is studying the information Ted recalls how he overheard a Trainer at the Cork Racecourse talking about ‘Chardonnay’ and how the horse is recovering from a cough and will not be back to full fitness for the race on Thursday. He adds this information as a note to the Named Entity ‘Chardonnay’ as shown in Figure 3.
Trailblazer enabled Ted to import the horse racing associative trail. It also allows him to export a copy of the horse racing associative trail with his added information back to the club’s Web site. Ted’s information on Chardonnay’s cough is now available for sharing with the other club members. Other members of the club can now import this horse racing trail from the shared account and merge the additional information gathered by Ted with their horse racing trail. The local racing enthusiasts at the club can now make an informed decision with regards to the horse Chardonnay’s chances of winning his race.
Ted is also a keen wine connoisseur and has many friends who share this hobby. Having seen how powerful Trailblazer is for sharing horse racing information he decides to use Trailblazer to create associative trails for wine. He creates a wine category and goes to a Web page about wine to show Trailblazer the words and phrases which typically occur on a page about wine. This is used to train Trailblazer as shown in Figure 4.
Ted continues to train Trailblazer about wine using a number of typical wine pages. He then adds a number of Named Entity types to his wine category, for example ‘Grape Type’ and ‘Region’. When he selects a Named Entity on a wine page, for example ‘Chardonnay’, he adds it to the category as a ‘Grape Type’. He continues to add a number of Named Entities with bookmarks and notes from his own knowledge of wine as shown in Figure 5.
Using his new category, when Ted opens a Web page about wine and presses the ‘Find Entities’ button any of the Named Entities that are on the page will get highlighted. When he passes a mouse over the entity, a tool tip informs him of its type. When he clicks on an entity for example ‘Chardonnay’, the associated bookmarks and notes are displayed as shown in Figure 6.
Ted links the selected Named Entities with bookmarks and notes information. He adds his first note on the best foods to pair with Chardonnay. His second note is a reminder that his girlfriend’s favorite wine is Chablis which is made from Chardonnay grapes. However, Ted does not want to share this information with anyone so he marks it as a private note to himself and therefore only he can see the note about his girlfriend as shown in Figure 7.
His selected Named Entities will form the basis for his wine trail. He will then browse several other wine Web sites from which he will follow the same steps to collect more information about wine. When trained Trailblazer will then be able to automatically detect if a Web page applies to wine and will load that domain specific category with its additional information.
Although Ted has the Named Entity ‘Chardonnay’ as an instance of ‘Horse’ in his ‘Horse Racing’ category, he can also add ‘Chardonnay’ to his ‘Wine’ category as an instance of ‘Grape Type’. Trailblazer differentiates between the two as shown in Figures 8 and 9.
Trailblazer categorizes the context of the page in Figure 8 as a horse racing Web page so the Named Entity ‘Chardonnay’ is recognized as a horse. In Figure 9, the Web page has a context of wine so using the context of this page ‘Chardonnay’ is a type of grape. Ted can now share his associative trails with his wine connoisseur friends and together they can build and enhance their knowledge of wine. 3
Trailblazer is an extension for the Mozilla Firefox [
Interfaces for Mozilla Firefox extensions are built using the XML User-interface Language (XUL) [10]
and Cascading Style Sheets (CSS). Extensions are coded in JavaScript [12] and can link to native
components using the cross-platform Component Object Model (XPCOM) [
JavaScript is an open, platform-independent, event-driven, interpreted programming language. RDF is an open, general-purpose data format for representing information in the Web. It is a common framework therefore application designers can leverage the availability of common RDF parsers and processing tools. This ability to exchange information between different applications means that the information may be made available to applications other than those for which it was originally created. XUL (pronounced zool, rhymes with cool) is an open, cross-platform language designed specifically for building portable user interfaces. A XUL overlay is a XUL file containing elements to be inserted into another XUL file. This insertion occurs when the other XUL file is rendered into an application interface. The browser window of Firefox is a XUL window. Extensions can be made to the browser window using an overlay file. XUL is used to overlay the browser to add the Trailblazer features. The Trailblazer overlay is rendered into the Firefox browser, generating the main Trailblazer panel as part of the browser window.
When Trailblazer is installed and the user launches Firefox, the standard browser window opens with the addition of a Trailblazer panel. In order to facilitate the addition of Named Entities and their bookmarks and notes, extra windows created using XUL dialog files are popped up by Trailblazer to request information from the user.
Firefox allows an installed extension to have additional privileges, such as reading local files and
modifying user preferences. The Input/Output component of Trailblazer uses another Firefox extension
called JSLib [
The Trailblazer Vocabulary store contains the information to calculate the most likely match for the context of the current Web page. Each Category stored within Trailblazer contains the Named Entity types, Named Entities, and their bookmarks and notes related to that category.
When a user chooses the ‘Train’ option on the currently viewed page, the Trainer component extracts the text from the page to supplement the information held against the current category within the Vocabulary. Trailblazer uses a Vector Space Model (VSM) to model the collected training information.
The VSM is an algebraic model used for information filtering and information retrieval [14]. It represents natural language documents in a formal manner by the use of vectors in a multi-dimensional space. Each dimension represents a term or concept found in the documents. The model creates a space in which documents are represented by vectors and allows decisions to be made about which documents are similar to each other. Spatial proximity is the physical distance between two items in space, for example the distance between two vectors or lines. Semantic proximity is the difference between two items in meaning, for example the context of the ‘Irish Racing’ Web page and the context of the ‘Wine Connoisseur’ Web page. The VSM uses spatial proximity to calculate semantic proximity. The Trainer uses the words and their frequency of occurrence from the current page to augment the existing training data.
When a page is loaded into the browser a comparison is made between the text on the page and the training data contained in the Trailblazer Vocabulary. A similarity measure is used to compare a vector representation of each category Trailblazer has had training on, and a vector representation of the currently viewed page. This calculates the most likely match, therefore deriving the context of the page in relation to the information obtained from training. Currently, Trailblazer is using the Cosine Similarity Measure [7] to compare the vectors of information. This is derived from the Law of Cosines, which is an extension of the Pythagorean Theorem.
The Cosine Similarity Measure is the cosine (in radians) of the angle between two vectors. For two vectors a and b the cosine similarity between a and b is given by: a ⋅ b a × b i.e. the dot product of a and b divided by the modulus of a multiplied by the modulus of b.
When a user presses the ‘Find Entities’ button, the NER functionality processes the currently viewed
Web page and highlights any discovered Named Entities from the previously derived category. NER [
Information held in Trailblazer can be exported on a category by category basis. Export files are named using the category name and an extension of EXP. These are generated in an RDF format. Each export file stores the words and their frequency of occurrence in relation to the export category. It also stores Named Entity types and Named Entities along with their related bookmarks and notes. Private notes are not exported.
Using Named Entities as a basis to share associative trails between Semantic Desktops 9 </RDF:RDF>
The import facility allows the user to import a file. Named Entities and their associated information are extracted from the file and appended into the users existing category file. Named Entity duplicates are not added, but their additional bookmarks and notes are merged. If a category file doesn’t exist a new one is created. The words and their weights for that category are then extracted from the export file and integrated into the user’s vocabulary file. This ensures that not only the annotation information is transferred between users, but also the training information to allow Trailblazer to detect the context of Web pages.
Annotea [
Magpie [19] is a Web browser extension which uses NER based on a supplied ontology to add links to Named Entities on a Web page. Each ontology contains metadata associated with the Named Entities which, when selected by the user, are resolved by a Web server and an appropriate Web page is returned. Based on the user’s interpretation of what the page is about, they choose an appropriate ontology to use to identify the Named Entities on that page and their associated links. Trailblazer differs from Magpie here in that it uses its cosine similarity logic to automatically identify the correct category to use for Named Entity identification. Magpie’s links are supplied in the ontology and cannot be added to by the user. Additionally it does not have the facility to allow users to add annotations.
XP Smart Tags [
In this paper we have described the Firefox extension Trailblazer and how it can be used in a collaborative environment. We have shown how Named Entities can be used as a basis for sharing associative trails between Semantic Desktops. Future work will extend these capabilities across the whole desktop infrastructure. Leveraging the XUL overlay technology we will apply the Trailblazer extension to Thunderbird [21], which is the Mozilla mail client. This will allow the making and sharing of associative trails across Web pages, RSS [22] and Mail. By building the capability in behind Smart Tags we should get the same coverage across Windows XP. The Annotea schema appears to have good potential to be extended to meet Trailblazer’s requirements. We also intend to build Annotea XPointer based bookmark and annotation capability into Trailblazer as this will allow the support of document specific associative trails. An additional area of exploration is the potential for using the Annotea server for sharing Trailblazer Named Entity based associative trails. Security needs to be addressed to ensure that links only come from trusted sources and that links are not to dangerous or undesirable Web resources.