Autonomous The Zen of the Web Semantic Web Services The DARPA Agent Markup Language for Services (DAML-S) provides a mechanism that begins to bridge the gap between the Web services infrastructure and the Semantic Web. Massimo Paolucci he Web, as we know it, is a collec- between them to date. Semantic Web and Katia Sycara tion of human-readable pages that research focuses mostly on markup lan- Carnegie Mellon University Tare virtually unintelligible to com- guages for annotating Web pages and on puter programs. While the Web emerged the inferential power needed to derive as a global repository of digitized infor- consequences from the annotated pages mation, this very information is, by and — essentially transforming the Web into large, unavailable for automatic compu- a knowledge base. Web services efforts tation. Two parallel efforts have emerged concentrate on interoperability standards in recent years that could overcome this and protocols for performing business-to- paradox: the Semantic Web1 is providing business (B2B) transactions. tools for explicit markup of Web content, In our work at Carnegie Mellon, we and Web services could create a network focus on research that attempts to bridge in which programs act as independent the gap. We adopt the vision of Web ser- agents that produce and consume infor- vices as autonomous goal-directed agents mation, enabling automated business that select other agents to interact with transactions. and that flexibly negotiate their interac- Ideally, these two efforts should sup- tion models, acting variously in port each other. The Semantic Web will client–server and peer-to-peer modes. The help create a repository of computer- resulting Web services, which we call readable data, and Web services will pro- autonomous Semantic Web services, use vide the tools for automatically using that ontologies and semantically annotated data. Somewhat surprisingly, however, Web pages to automate the fulfillment of there have been few points of contact tasks and transactions. In particular, these 34 SEPTEMBER • OCTOBER 2003 Published by the IEEE Computer Society 1089-7801/03/$17.00©2003 IEEE IEEE INTERNET COMPUTING Autonomous Semantic Web Services Web services Discovery: services use the Semantic Web to support capabil- composition: UDDI ity-based discovery and interoperation at runtime. WSCI, BPEL4WS A first step toward this vision is to develop for- XML Interface specification layer: WSDL mal languages and inference mechanisms for rep- resenting and reasoning with core Web service Messaging layer: SOAP concepts. The DARPA Agent Markup Language for Transport layer: HTTP, TCP, UDP, and so on Services (DAML-S)2 is the first attempt to define such a language. This article describes DAML-S Figure 1. Web services infrastructure. The current and some example computational models that infrastructure uses XML to describe multiple show how it can be viewed as the first step in layers of abstraction from the transport bridging the gap between the Semantic Web and mechanism. These include message description proposed industry standards for Web services. (SOAP), a mapping from messages to operations performed by the Web service (WSDL), abstract The Semantic Web process representation (BPEL4WS and WSCI),and for Web Services discovery (UDDI). One objective behind the Semantic Web is to pro- vide languages for expressing the content of Web pages and making that information accessible to lems through collaboration and negotiation. With- agents and computer programs. More precisely, the in this scheme, ontologies not only define a shared Semantic Web is based on a set of languages such conceptualization for interpreting semantic markup as the Resource Description Framework (RDF), of Web sites, but also provide a shared vocabulary DAML+OIL, and the more recent Web Ontology that lets services across the Web use the same ter- Language (OWL), which can be used to annotate minology to interpret each other’s messages. Web content. These languages have well-defined Ultimately, the Semantic Web will provide the semantics and inferential procedures that let basic mechanisms for extracting information from agents draw inferences from the languages’ state- Web pages and the basic knowledge that Web ser- ments. Using the semantic markup for the US vices will use in all transactions. In addition to National Oceanic and Atmospheric Administra- knowledge, however, Web services need an infra- tion’s page reporting Pittsburgh’s weather condi- structure that facilitates reliable communication — tions, for example, an agent could learn that the registries to locate other services, reputation ser- current condition is heavy snow. The agent might vices, guarantees of secure and private transac- further learn from the Pittsburgh school board tions, and so on. Such an infrastructure falls out- site’s semantic markup that all schools are closed side the current view of the Semantic Web’s scope. on days of heavy snow. Combining the two pieces of information, the agent could infer that Pitts- Current Web burgh schools are closed today. Services Infrastructure The Semantic Web’s second element is a set of Researchers and developers have already proposed ontologies that provide conceptual models for many definitions for Web services. Simply put, interpreting the information provided. An ontol- they are programs that interoperate on the Web. ogy of weather might contain concepts such as The success of e-commerce in the late ’90s and the temperature, snowy, cloudy, and sunny, for recent plethora of interoperability standards pro- example, and relationships between the terms. posed for Web-based business transactions have The Semantic Web vision is about transforming generated significant interest in automating pro- the Web into an Internet-wide knowledge-repre- gram interactions for B2B e-commerce. These sentation system in which ontologies provide the efforts have created a vision of a Web of dynami- conceptual framework for interpreting the infor- cally interoperating nodes rather than static pages. mation provided by Web pages. To produce the This is the vision that Web services try to realize. types of inferences we’ve described, the Semantic Web requires computational processes and agents Basic Standards that can interpret semantic content and derive The Web services infrastructure will provide the consequences from the information they collect. basic standards that let Web services interact. The The Semantic Web also supports a more distrib- diagram in Figure 1 shows how some popular uted computational model in which a requester proposed standards could fit together. The unify- transacts with multiple Web services, solving prob- ing factor is XML, as shown by the left column, IEEE INTERNET COMPUTING http://computer.org/internet/ SEPTEMBER • OCTOBER 2003 35 The Zen of the Web <types> ... </types> UDDI descriptions include a host of useful <message name=”GetTradePriceInput”> information, such as the Web service provider and <part name=”request” the binding (including the transport protocol port) type=”xsd:string”/> that lets a requester invoke the service. In addi- </message> tion, Web service descriptions in UDDI can refer <message name=”GetTradePriceOutput”> to TModels — unbounded attribute sets that can <part name=”response” be associated with Web services — which can rep- type=”xsd:float”/> resent any type of information about a given ser- </message> vice. TModels can specify a Web service WSDL description’s location, for example, or a Web ser- <portType name=”StockQuotePortType”> vice’s classification within a taxonomy, such as <operation name=”GetTradePrice”> the United Nations Standard Products and Ser- <input message=”GetTradePriceInput”/> vices Code. <outputmessage=”GetTradePriceOutput”/> </operation> Need for Autonomy </portType> One overarching characteristic of the Web services infrastructure is its lack of semantic information. <binding> ... </binding> It relies exclusively on XML for interoperation, but that guarantees only syntactic interoperability. <service> ... </service> Expressing message content in XML lets Web ser- vices parse each other’s messages, but it does not Figure 2. Example WSDL description. This service performs the facilitate semantic “understanding” of the mes- operation GetTradePrice, described by the input message sages’ contents. GetTradePriceInput (string) and the output message Consider the WSDL fragment in Figure 2, for GetTradePriceOutput (float). Much remains unknown, example, which describes a stock-reporting Web however, regarding the information to be encoded in the string and service. The service performs only one operation, how to interpret the output float. GetTradePrice, which requires one input, GetTradePriceInput (of type string), and pro- duces an output GetTradePriceOutput (of type which cuts across all layers. The two most popular float). This description specifies the syntactic type proposed standards are SOAP3 and the Web Ser- for the data transferred, but not what data it vices Description Language (WSDL).4 SOAP expects. We do not know, for example, whether the defines a format for passing messages between Web service expects a company name or a ticker Web services, and WSDL describes each service’s symbol as input; similarly, we do not know interface — how to contact it (through RPC or whether the output will be the latest quote, the asynchronous messaging, for example) and how stock’s beta value, or the 12-week average. to serialize the information exchanged. SOAP and Industry’s current