The Tetherless World Mobile Wine Agent is a Semantic Web application for making wine and food recommendations to users. In order to make the interface easier for users, instances are generated through a common interface that uses data from the underlying ontology to drive user interaction. The agent is being integrated with social applications, such as Facebook and Twitter, to allow users to leverage and share generated content with other individuals on the World Wide Web. Anyone can contribute data by constructing an RDF graph and making it available over the web in XML, which the agent will read and incorporate into its internal graph. Thus, users of the wine agent can access menus published in RDF on the web. By bringing together semantics, data sharing, and extensibility, the Mobile Wine Agent demonstrates how semantics and user interactions can work in tandem to grow the Semantic Web.
The primary goal of the Semantic Web is to rede ne how we store, access, and
think about dat [
In this paper, we will describe the Mobile Wine Agent's ontology-sensitive interface for generating instances and classes, the underlying command architecture system for the agent, and how the agent integrates with social web applications, speci cally Twitter and Facebook. Additionally, we will highlight how to utilize existing web technologies to provide a seamless user experience with the Wine Agent.
Wine and food are a rich domain and the history of recommender systems
goes back almost twenty years [
The Mobile Wine Agent expands on the food and wine ontologies which were
released as part of the World Wide Web Consortium's OWL 1.0 speci cation [
The mobile client application currently runs on Apple's iPhone R and iPod touch R platforms (see Fig. 1) with an Android R version planned. The earlier iterations of these devices have a 400 MHz process with 80 MB available RAM. Due to these computational limitations, the system was designed such that basic reasoning could be performed locally on the device while any complex reasoning would take place on an external server. The device keeps a cached copy of the ontology stored locally in RDF to improve startup times and data access times. The current version of the ontology contains 5500 triples and the reasoner takes an average of 1572 ms ( =12 ms) to load the ontology from NAND and consumes 827 KB of memory. The primary reason for keeping the ontology on the device is to drive the user interface to assist users in generating food and wine descriptions (see Sec. 4). The resulting interface is ontology-aware and can use information present to simplify user interaction.
Performing the full breadth of OWL reasoning is not currently supported by the client's reasoner. Complicated processing can be costly in terms of battery life
and responsiveness of the device. Therefore, it is necessary to o oad complicated reasoning tasks (such as nding appropriate recommendations) to the remote server. When external reasoning is required, the client transmits a list of URLs for the server to load and then issues an appropriate command to tell the server what information the client needs to continue. Additionally, the user's food/wine preferences are pushed to the server if they have not been made available for sharing. 3
The Mobile Wine Agent relies on three core ontologies which describe the agent,
food, and wine. While the wine ontology is substantially unchanged from its
release as part of the OWL Speci cation (see [
http://wineagent.tw.rpi.edu/data/agent.rdf instances as well as added associations between foods and what they are made from. For example, pork products, such as bacon and ham, are declared to be made from pigs. This increased speci city a ords users who do not to eat porkbased products to lter out dishes containing these ingredients. Sauces have also been separated from foods so that restauranteurs can identify speci c sauces that are then realized as members of sauce classes. These classes can then be used to describe dishes, instead of being xed to the class hierarchy as was the case in the old ontology. Flavor pro les have also been added so that individuals can lter out bitter or spicy foods depending on their particular palette. The result is a more expressive ontology to help make it easier for individuals to describe foods to the Wine Agent.
Often services use WSDL [
Every command that the agent processes is a descendant of the agent:Command class. When the agent processes a command, it will return an agent:Response. A successful response will contain any necessary information for the client to use as a result of the command. In cases where processing fails, agent:Error, a subclass of agent:Response is returned with a human-readable string describing the error that occurred. Other classes include agent:RecommendationList, a subclass of rdf:Seq, which stores recommendations and agent:Recommendation which pairs dish classes with wine classes or dish instances with wine instances. 4
Due to their small size, mobile devices require cleaner, more structured interfaces than their desktop counterparts and given the complex domain the Wine Agent operates over, it was important to construct a highly dynamic interface which could be used to describe foods and wines. In order to accomplish this goal, an RDFS reasoner that understands all six OWL restrictions and intersectionOf was built into the Mobile Wine Agent. This reasoner is utilized whenever the user describes a dish or wine to the agent.
When the user chooses to describe, for example, a wine for a meal, the client requests the class vin:Wine from the reasoner and passes it to the description generator interface. The generator then asks the reasoner for a list of properties that might apply to this class, which will include any property that has as its http://wineagent.tw.rpi.edu/data/food.rdf http://wineagent.tw.rpi.edu/data/wine.rdf rdfs:domain vin:Wine or any of its superclasses, as well as properties which lack a value on rdfs:domain. The user is presented with a picker view5 listing the properties. When the user selects a property, the interface is adjusted based on whether the rdfs:range is the class owl:Class or not. In the event the range of the property is owl:Class an additional entry in the display is added entitled Union Of which allows the user to construct an anonymous union class. Otherwise, two entries are added: One Of and Any which allow the user to specify that the value of a property must be a member of the speci ed class (either an anonymous enumerated class or a previously named class).
The behavior of the interface when selecting the value for a property is also conditional. The generator asks for the appropriate values for the property (or subclasses of the range in the case of the Any operation) to present them to the user. If the property is an owl:FunctionalProperty or the supplied class, vin:Wine in this case, is a subclass of an owl:Restriction on the property with an owl:maxCardinality of 1, then the generator displays another picker view to limit the property selection to a single element. If neither of these conditions applies, then a generic table view is displayed where the user can select one or more values. The resulting property/value pair(s) are rendered on an alternate screen which keeps a cumulative representation of the user's choices. Upon completion, the description generator returns either a string with the necessary RDF to describe the individual created or an owl:Class which will match any individual with the speci ed properties. 5
Since the underlying platform for the Mobile Wine Agent is di erent from a standard desktop application, it can take advantage of features unique to mobile devices. Users can activate location updates via the Global Positioning System (GPS) for the Wine Agent. Once the agent receives information about the user's location, it presents a map with known restaurants serving their menus in RDF6 (see Fig. 2). The user selects the appropriate restaurant, and the system will retrieve the menu and wine list for the reasoner to use. These lists are presented to the user so they can pick the appropriate meal or wine instead of manually inputting a description using the interface in Sec. 4. Restaurant selection is also used to increase reasoning performance by reducing the ABox to only contain dishes and wines available on the restaurants menu. 5 The UIPickerView is a component of the iPhone SDK and allows for single-item selection from a list. 6 We currently provide a number of RDF representations from HTML and PDF for some restaurants for demonstration purposes. Plans are underway to develop a Desktop Wine Agent which can be used by proprietors to generate RDF representations of menus and wine lists. The Wine Agent provides a variety of ways for users to interact with one another, both internally and through external applications. Interaction data is stored as RDF and as such can be queried by external applications. When users select a recommendation, for example, they can indicate whether they approve of the recommendation or not. The agent stores this information as an RDF triple and in the positive case prioritizes the recommendation and in the negative case removes it from future results for this individual. Recommendations and preferences can also be shared with other users via Facebook and Twitter. Facebook is the largest social networking community and its API is extremely exible. We have integrated access to Facebook into the Wine Agent to provide a number of features. For users to enable Facebook for the Wine Agent, they must rst add the Tetherless World Wine Agent application in Facebook. Once they grant permission to use their information in the application, they can log into Facebook through the Wine Agent. This will enable a number of features for obtaining and sharing content, such as sharing preferences and posting to their news feed.
First, users can allow others to read their personal preferences by granting permissions to friends or friend lists. The client application uploads the user's preferences to be stored on the server7 and issues a command to the server to allow permission to the indicated users to access these new triples. When a friend wants to reason about food and wine using these shared preferences, they can choose to import them into their instance of the agent through the appropriate menu items. For privacy reasons, users can also revoke permissions to view personal preferences at any time. 7 Unlike recommendations and the core ontologies, these triples are not publicly available for privacy reasons.
Users can join events via Facebook and share their preferences with these events (see Fig. 3). Any attendees to the event using the Wine Agent can register a set of triples so that the event organizer can load all of the preferences of those attending into an instance of the Wine Agent for reasoning. These permissions are only granted for the particular event and thus are more restricted than the general sharing mechanism. In the future, this option will also allow for creating group recommendations via iPhone discovery (see Future Work for details).
To facilitate sharing of recommendations, dishes, and wines, the Wine Agent also ties into the Facebook feed, allowing users to post a message to their friends regarding any entry in the agent they nd notable. These posts are accompanied with the appropriate URI so that users can launch the Wine Agent directly from Facebook.com. 6.2
Twitter In addition to the integration with Facebook, the Wine Agent is begin integrated with Twitter so that users with Twitter accounts can post updates to their Twitter feeds about the dishes, wines, and recommendations they discover through the agent. The Wine Agent presents a text box for the user to enter a message and uses URL shortening tools to provide links to any resource the user identi es for the message (see Fig. 4). The text is then sent to the Twitter API using the OAuth security protocol. This allows users to share Semantic Web data with friends without having to understand the complexities of RDF and OWL. 7
In order to facilitate interaction on the web and the sharing of data generated
by the Mobile Wine Agent, the agent responds to a custom URI scheme and
all recommendations and dish/wine classes generated by users are given unique
URIs starting with this scheme. These URIs can be used as part of an anchor
HTML tag:
<p>Check out this <a href="twwa://rec/dlm-5/">
recommendation</a> from the Wine Agent.
</p>
or with RDFa [
Many approaches have been taken to providing and describing web services.
The Web Services Description Language [
Such a request would generate an agent:Response which would contain a number of recommendations which match the instance LinguiniConGamberetti. Using the XML Stylesheet Language Transformations (XSLT), this output could be translated into new input which might match the recommendation against LoPorto's wine list. This system is discussed in more detail on the Wine Agent's website8.
The Mobile Wine Agent is a complex Semantic Web application that integrates with a number of social platforms to enable sharing of user-generated semantic content. It provides a common user interface to generate descriptions of wines, dishes, and recommendations that is sensitive to the underlying ontology driving the application. The ontology and instance data can be extended by anyone, with a particular focus on restauranteurs and winery proprietors, to re ect the large variety of wines and dishes. Custom URIs can be linked to from anywhere on the web that are recognized by Safari as being associated with the Wine Agent, thus allowing links on blogs, Facebook, and Twitter to directly launch the agent and let users start working with the data immediately. The Wine Agent server can also act as a web service, accepting commands and returning results all described in RDF. While many more features could be added to the Wine Agent to extend its capabilities, the core structure, both the ontologies and the application itself, provide a rich Social Semantic Web experience. 10
To further expand on the social aspects of the Wine Agent, a full OWL-DL capable reasoner is being developed to run on the iPhone. This will allow the phone to perform all of the necessary reasoning locally but will also facilitate the investigation of parallel reasoning across multiple devices using the iPhone's Bluetooth R interface. With this new capability, along with the Facebook Events integration as mentioned in Sec. 6.1, a group of friends could go out to dinner, import their dinner event, and have each of the devices talk to the others, sharing preferences and meal choices to pick the most appropriate wines for dinner.
Supporting provenance in the Wine Agent is another future goal. For example, if a wine critic such as Robert Parker began making recommendations using the Wine Agent, many individuals might want to know which recommendations are his and prefer those recommendations over others. Similarly, friends who are less wine literate may want to rely on recommendations from friends who are more wine literate when going out to dinner. Using data from Facebook, for example, could be one such way of building provenance into the Wine Agent.