AI Magazine Volume 24 Number 3 (2003) (© AAAI)
Articles Where Are the Semantics in the Semantic Web?
Michael Uschold
■ The most widely accepted defining feature of the ing primarily intended for human consump- semantic web is machine-usable content. By this tion to being intended for use both by humans definition, the semantic web is already manifest in and machines. shopping agents that automatically access and use Web tasks and services: The web is evolv- web content to find the lowest air fares or book ing from being primarily a place to find things prices. However, where are the semantics? Most to being a place to do things as well (Smith people regard the semantic web as a vision, not a 2001).1 reality—so shopping agents should not “count.” To use web content, machines need to know what All these new capabilities for the web de- to do when they encounter it, which, in turn, re- pend in a fundamental way on the idea of se- quires the machine to know what the content mantics, giving rise to another perspective means (that is, its semantics). The challenge of de- from which the web evolution can be viewed: veloping the semantic web is how to put this Semantics: The web is evolving from con- knowledge into the machine. The manner in taining information resources that have little which it is done is at the heart of the confusion or no explicit semantics to having a rich se- about the semantic web. The goal of this article is mantic infrastructure. to clear up some of this confusion. Despite the widespread use of the term se- I explain that shopping agents work in the com- mantic web, it does not yet exist except in iso- plete absence of any explicit account of the seman- lated environments, primarily in research labs. tics of web content because the meaning of the In the World Wide Web Consortium (W3C) Se- web content that the agents are expected to en- mantic Web Activity Statement, we are told counter can be determined by the human pro- that grammers who hardwire it into the web applica- tion software. I therefore regard shopping agents the Semantic Web is a vision: the idea of as a degenerate case of the semantic web. I note having data on the Web defined and various shortcomings of this approach. I conclude linked in a way that it can be used by ma- by presenting some ideas about how the semantic chines not just for display purposes, but web will likely evolve. for automation, integration and reuse of data across various applications (emphasis mine).2 he current evolution of the web can be As envisioned by Tim Berners-Lee: characterized from various perspectives (Jasper and Uschold 2003): the Semantic Web is an extension of the T current Web in which information is giv- Locating resources: The way people find things on the web is evolving from simple free en well-defined meaning, better enabling computers and people to work in cooper- text and keyword search to more sophisticated ation (Berners-Lee, Hendler, and Lassila semantic techniques both for search and navi- 2001, p. 35) (emphasis mine). gation. Users: Web resources are evolving from be- [S]omething has semantics when it can
Copyright © 2003, American Association for Artificial Intelligence. All rights reserved. 0738-4602-2003 / $2.00 FALL 2003 25 Articles
be ‘processed and understood by a com- The lack of an adequate definition of the se- puter,’ such as how a bill can be processed mantic web, however, is no reason to stop pur- by a package such as QUICKEN (Trippe suing its development any more than an inad- 2001, p. 1). equate definition of AI was a reason to cease AI There is no widespread agreement on exact- research. Quite the opposite, new ideas always ly what the semantic web is for or exactly what need an incubation period. it is. Some good ideas about what the semantic The research community, industrial partici- web will be used for have emerged from the pants, and software vendors are working with W3C effort to define a standard ontology lan- the W3C to make the semantic web vision a re- 6 guage.3 From the previous descriptions, there is ality (Berners-Lee et al 2001). It will be layered, clear emphasis on the information content of extensible, and composable. A major part will the web as machine usable and associated with entail representing and reasoning with seman- more meaning. tic metadata and providing semantic markup Note that machine refers to computers (or in the information resources. Fundamental to computer programs) that perform tasks on the the semantic infrastructure are ontologies, web. These programs are commonly referred to knowledge bases, and agents along with infer- as software agents, or softbots, and are found in ence, proof, and sophisticated semantic query- web applications. ing capability. Machine-usable content presumes that the The main intent of the semantic web is to machine knows what to do with information give machines much better access to informa- on the web. For this to happen, the machine tion resources so they can be information in- The web is reads and processes a machine-sensible specifi- termediaries in support of humans. According cation of the semantics of the information. to the vision described in Berners-Lee et al. evolving from This approach is robust and very challenging (2001), agents will be pervasive on the web, containing and largely beyond the current state of the art. carrying out a multitude of everyday tasks. information A much simpler alternative is for the human Hendler describes many of the important tech- web application developers to hardwire the nical issues that this approach entails, empha- resources that knowledge into the software so that when the sizing the interdependence of agent technolo- have little or machine runs the software, it does the correct gy and ontologies (Hendler 2001). To carry out thing with the information. In this second sit- their required tasks, intelligent agents must no explicit uation, machines already use information on communicate and understand meaning. They semantics to the web. There are electronic broker agents in must advertise their capabilities and recognize routine use that make use of the meaning asso- the capabilities of other agents. They must lo- having a rich ciated with web content words, such as price, cate meaningful information resources on the semantic weight, destination, and airport. Armed with a web and combine them in meaningful ways to infrastructure. built-in understanding of these terms, these so- perform tasks. They need to recognize, inter- called shopping agents automatically peruse pret, and respond to communication acts from the web to find sites with the lowest price for a other agents. book or the lowest airfare between two given In other words, when agents communicate cities. Thus, we still lack an adequate character- with each other, there needs to be some way to ization of what distinguishes the future seman- ensure that the meaning of what one agent tic web from what exists today. “says” is accurately conveyed to the other Because the RESOURCE DESCRIPTION FRAMEWORK) agent. There are two extremes, in principle, for (RDF) is hailed by the W3C as a semantic web handling this problem. The simplest (and per- language,4 some people seem to have the view haps the most common) approach is to ignore that if an application uses RDF, then it is a se- the problem altogether. That is, just assume mantic web application. This is reminiscent of that all agents are using the same terms to the “if it is programmed in Lisp or Prolog, then mean the same things. In practice, this as- it must be AI” sentiment that was sometimes sumption will usually be built into the appli- evident in the early days of AI. There is also cation. The assumption could be implicit and confusion about what constitutes a legitimate informal, or it could be an explicit agreement semantic web application. Some seem to have among all parties to commit to using the same the view that an RDF tool such as CWM is one.5 terms in a predefined manner. This approach This is true only in the same sense that KEE and only works, however, when one has full con- ART were AI applications. They were certainly trol over what agents exist and what they generating income for the vendors, which is might communicate. In reality, agents need to different from the companies using the tools to interact in a much wider world, where it can- develop applications that help their bottom not be assumed that other agents will use the line. same terms, or if they do, it cannot be as-
26 AI MAGAZINE Articles sumed that the terms will mean the same have semantics, where the semantics are, and thing. how they are used. We identify a kind of se- The moment one accepts the problem and mantic continuum ranging from the kind of grants that agents might not use the same semantics that exist on the web today to a rich terms to mean the same things, one needs a semantic infrastructure on the semantic web of way for an agent to discover what another the future. agent means when it communicates. Thus, Real-world semantics: Real-world seman- every agent needs to publicly declare exactly tics7 are concerned with the “mapping of ob- what terms it is using and what the terms jects in the model or computational world on- mean. This specification is commonly referred to the real world … [and] issues that involve to as the agent’s ontology (Gruber 1993). If it human interpretation, or meaning and use of were written only for people to understand, data or information” (Ouksel and Sheth 1999). this specification could just be a glossary. How- An object in the model might be a tag or a ever, meaning must be accessible to other soft- term or, possibly, a complex expression in ware agents, requiring the meaning to be en- some language. We might also speak of the se- coded in some kind of formal language. This mantics of a possibly large set of expressions, approach will enable a given agent to use auto- which collectively are intended to represent mated reasoning to accurately determine the some real-world domain. The real-world se- meaning of other agents’ terms. For example, mantics correspond to the concepts in the real suppose agent 1 sends a message to agent 2 and world that the objects in the model refer to. in this message is a pointer to agent 1’s ontol- Agent communication language perfor- ogy. Agent 2 can then look in agent 1’s ontol- matives: In the context of the semantic web, ogy to see what the terms mean, the message is performatives such as request or inform in agent successfully communicated, and the agent’s communication languages (Smith et al. 1998) task is successfully performed. At least this is require semantics to ensure that agents com- the goal. The holy grail is for this processing to municate effectively. happen consistently, reliably, and fully auto- Axiomatic semantics: An axiomatic seman- matically. In practice, there is a plethora of dif- tics for a language specifies “a mapping of a set ficulties, most arising from various sources of of descriptions in [that] language into a logical heterogeneity. For example, there are many theory expressed in first-order predicate calcu- different ontology representation languages, lus” (p. 4). The basic idea is that “the logical different modeling styles, and an inconsistent theory produced by the mapping … of a set of use of terminology. For further discussion, see such descriptions is logically equivalent to the the section entitled Why Do Web Shopping intended meaning of that set of descriptions” Agents Work? (p. 1) (Fikes and McGuinness 2001). Axiomatic semantics have been given for RDF, RDF SCHEMA Semantics: A (RDF-S), and DAML + OIL. The axiomatic seman- tics for a language helps to ascribe a real-world Many-Splendored Thing semantics to expressions in this language, in The core meaning of the word semantics is that it limits the possible models or interpreta- meaning itself. However, there is no agreement tions that the set of axioms might have. about how this definition applies to the term Model-theoretic semantics: “A model-theo- semantic web. In what follows, I characterize retic semantics for a language assumes that the the many things that one might mean when language refers to a ‘world,’ and describes the talking about semantics as it pertains to the se- minimal conditions that a world must satisfy mantic web. It is not my intention to define in order to assign an appropriate meaning for the term but, rather, to make some important every expression in the language.”8 Model-the- distinctions that people can use to communi- oretic semantics are used as a technical tool for cate more clearly when talking about the se- determining when proposed operations on the mantic web. language preserve meaning. In particular, they In the context of achieving successful com- characterize what conclusions can validly be munication among agents on the web, I am drawn from a given set of expressions indepen- talking about the need for agents to under- dent of what the symbols mean. stand the meaning of the information being Intended versus actual meaning: A key to exchanged between agents and the meaning of the successful operation of the semantic web is the content of various information sources that the intended meaning of web content be that agents require to perform their tasks. We accurately conveyed to potential users of this focus attention on the questions of what kinds content. In the case of shopping agents, the of semantics there are, what kinds of things meaning of terms such as price is conveyed
FALL 2003 27 Articles
Pump: "A device for moving a gas or liquid from one place or container to another." (pump has (superclasses (…))
Shared Human Semantics Hardwired and Semantics Processed Consensus Text Descriptions Used at Runtime and Used at Runtime
Implicit Informal Formal Formal (Explicit) (For Humans) (For Machines)
Figure 1. Semantic Continuum. Semantics can be implicit, existing only in the minds of the humans who communicate and build web applications. They can also be ex- plicit and informal, or they can be formal. The further I move along the continuum, the less ambiguity there is, and the more likely it is to have interoperable, robust, and correctly functioning web applications.
based on human consensus. However, mistakes A Semantic Continuum are always possible because of inconsistency of Three questions can be asked about how se- natural language usage. When formal lan- mantics can be specified: First, are the seman- guages are used, an author attempts to commu- tics explicit or implicit? Second, are the seman- nicate meaning by specifying axioms in a logi- tics expressed informally or formally? Third, cal theory. In this case, one can talk about are the semantics intended for human process- intended versus actual models of the theory. ing or machine processing? There is normally just one intended model. It These questions give rise to four kinds of se- corresponds to what the author wanted the ax- mantics: (1) implicit, (2) explicit and informal, ioms to represent. The actual models corre- (3) explicit and formal for human processing, spond to what the author actually has repre- sented. They consist of all the objects and and (4) explicit and formal for machine pro- relationships, and so on, in the real world that cessing. are consistent with the axioms. The goal is to For implicit and informal semantics, there is create a set of axioms such that the actual mod- no alternative to hard wiring the semantics in- els include only the intended model(s). to web application software. In the case of for- I believe that the idea of real-world seman- mal semantics, hard wiring remains an option; tics, as described earlier, captures the essence of in which case, the formal semantics serve the the main use of the term semantics in a seman- important role of reducing ambiguity in speci- tic web context. However, it is only loosely de- fying web application behavior compared to fined. The ideas of axiomatic and model-theo- implicit or informal semantics. There is also retic semantics are being used to make the idea the new possibility of using automated infer- of real-world semantics for the semantic web ence to process the semantics at run time, thus more concrete. allowing for much more robust web applica- From this discussion, it is clear that several tions in which agents automatically learn things have semantics: (1) terms or expres- something about the meaning of terms at run sions, referring to the real-world subject matter time. of web content (for example, semantic mark- I define these four kinds of semantics to be up); (2) terms or expressions in an agent com- four somewhat arbitrary points along a seman- munication language (for example, inform); tic continuum (figure 1). At one extreme, there and (3) a language for representing the previ- are no semantics at all, except what is in the ous information (for example, the semantics of minds of the people who use the terms. At the DAML = OIL or RDF).9 other extreme are formal and explicit seman-
28 AI MAGAZINE Articles tics that are fully automated. The further you mantics that is expressed in a formal move along the continuum, the less ambiguity language).14 there is, and the more likely the web applica- Typically, the semantics expressed in infor- tions are to be robust, correctly functioning, mal documents are hard wired by humans in easy to maintain, and interoperable. I consider working software. Compiler writers use lan- these four points on the semantic continuum, guage-definition specifications to write compil- in turn. Note that there are likely to be many ers. The specifications for RDF and UML are used cases that are not clear cut and, thus, arguably to develop modeling tools such as CWM and RA- might fall somewhere in between. TIONAL ROSE. Implicit Semantics In the simplest case, the I characterized this (somewhat arbitrary) semantics are implicit only. Meaning is con- point on the semantic continuum by having veyed based on a shared understanding de- the semantics expressed in an informal lan- rived from human consensus. A common ex- guage. However, we can further distinguish ample of this case is the typical use of XML tags, whether the semantics are informal or formal. such as price, address, or delivery date. No- The latter case reduces ambiguity and can be where in an XML document, or DOCUMENT-TYPE seen as further along the semantic continu- 15 DEFINITION (DTD) or SCHEMA, does it say what um, which helps to avoid inconsistent and in- these tags mean.10 However, if there is an im- compatible implementations. Users might no- plicit shared consensus about what the tags tice features and start depending on them, mean, then people can hardwire these implicit resulting in problems if interoperability is re- semantics into web application programs, us- quired or implementations change. For these ing screen scrapers and wrappers. This example and other reasons, informal semantics are illustrates how one implements shopping sometimes inadequate, which motivates efforts agents that search web sites for the best deals. to create formal semantics, for example, for UML From the perspective of mature commercial ap- (Evans et al. 1998),16 RDF, and DAML + OIL.17 plications that automatically use web content Formal semantics are not just playthings for as conceived by semantic web visionaries, this logicians and academics. Given the emerging approach is at or near the current state of the importance of RDF and the semantic web, ven- art. The disadvantage of implicit semantics is dors were demanding to know what the RDF that they are rife with ambiguity. People often specification actually meant.18 don’t agree about the meaning of a term. For Although the formal semantics are a big example, prices come in different currencies, help, if they are expressed informally, they are and they might or might not include various not amenable for machine processing. Next, I taxes or shipping costs. The removal of ambi- consider the case of semantics expressed in a guity is the major motivation for the use of formal language. I distinguish between whether specialized language in legal contracts. The they are intended for human processing only or costs of identifying and removing ambiguity for machine processing as well. are very high. Formally Expressed Semantics for Human Informally Expressed Semantics At a further Processing Further along the continuum, point along the continuum, the semantics are there are explicit semantics expressed in a for- explicit and are expressed in an informal nota- mal language. However, they are intended for tion or language, for example, a glossary or a human processing only. These semantics can text specification document. They are mainly be thought of as formal documentation or as for humans. Given the complexities of natural formal specifications of meaning. Some exam- language, machines have an extremely limited ples of this include the following: ability to make direct use of informally ex- Modal logic is utilized to define the semantics pressed semantics. There are many examples of of ontological categories such as rigidity and informally expressed semantics, usually found identity (Guarino et al. 1994). These descrip- in specification documents written in (often tions are for the benefit of humans, to reduce highly technical) natural language: (1) the or eliminate ambiguity in what is meant by meaning of terms in the Dublin core;11 (2) the these ideas. meaning of tags in HTML such as
, which Modal logic is used to define the semantics means second-level header; (3) the meaning of of performatives such as inform and request in expressions in modeling languages such as the agent communication languages (ACLs) UNIFIED MODELING LANGUAGE (UML);12 (4) the infor- (Smith et al. 1998). Humans use the formal de- mal semantics in the original specification of finitions to understand, evaluate, and compare RDF SCHEMA;13 (5) the model theory (formal se- alternative ACLs. They are also used to imple- mantics) of RDF SCHEMA that is developed subse- ment agent software systems that support quently (does not include the axiomatic se- these notions.
FALL 2003 29 Articles
Many axioms and definitions in the ENTER- hand, and it is intended for machine process- PRISE ontology (Uschold et al. 1998) were creat- ing. When the compiler encounters the sym- ed without the expectation that they would be bol, it places a call to the appropriate proce- used for automated inferencing (although this dure. The meaning of the symbol is what remained a possibility). The primary purpose happens when the procedure is executed. The was to help communicate the intended mean- agent determines the meaning of the symbol ing to people. by calling the appropriate procedure, so in Formal semantics for human processing can some sense, these symbols can be viewed as go a long way toward eliminating ambiguity, having machine-processible semantics. but because there is still a human in the loop, I am instead focusing on a declarative view. there is ample room for error. From this perspective, I ask how an agent can Note that in all practicality, there is little dif- learn the meaning of a new term from a formal, ference between formal semantics expressed declarative specification of its semantics. Ideal- informally and formal semantics expressed for- ly, this learning should occur without making mally but intended only for human process- any assumptions at all. In this case, all symbols ing. In both cases, the processing is intended might as well be in a never-before-seen script for humans, and the formality serves to reduce from a long-extinct intelligent species on Mars. ambiguity. There is no knowledge of the meaning of the Formally Expressed Semantics for Machine symbols and the rules of syntax for the lan- Processing Finally, there is the possibility of guage, nor is there any information on the se- explicit, formally specified semantics that are mantics of the language. This general case is the intended for machines to directly process using most challenging kind of cryptography. … how [can] automated inference. The idea is that when Issues and Assumptions an agent … new terms are encountered, it is possible to au- tomatically infer something about their mean- Cryptography is extremely difficult for hu- learn the ing and, thus, their use. Inference engines can mans, never mind machines, so we have to meaning of a be used to derive new information for a wide start making some assumptions. Here, I consid- variety of purposes. I explore this topic in er some key issues and assumptions. new term depth in the next section. Language heterogeneity: Different ontol- from a ogy languages are often based on different un- derlying paradigms (for example, description formal, Machine-Processible Semantics logic, first-order logic, frame-based representa- declarative The defining feature of the semantic web is ma- tion, taxonomy, semantic net, and thesaurus). specification chine-usable content, which implies that the Some ontology languages are very expressive, machine knows what to do with the web con- and some are not. Some ontology languages of its tent it encounters. This does not mean that have a formally defined semantics, and some semantics[?] there is any explicit account of the semantics. do not. Some ontology languages have infer- Instead, the semantics (whether implicit, infor- ence support, and some do not. If all these dif- mal, or formal) can be hardwired into the web ferent languages are to be allowed, then there applications. A more robust approach is to for- is the challenging problem of translating be- mally represent the semantics and allow the tween them. For simplicity then, I assume that machine to process it to dynamically discover the expressions encountered by our agent are what the content means and how to use it—I from a single language whose syntax and se- call this machine-processible semantics. This goal mantics are already known to the agent, for ex- might be impossible to achieve in its full gen- ample, RDF SCHEMA or OWL. erality, so I restrict this discussion to the fol- Incompatible conceptualizations: Even lowing specific question: How can a machine with a uniform language, there can still be in- (that is, software agent) learn something about compatible assumptions in the conceptualiza- the meaning of a term that it has never en- tion. For example, in Hayes (1996), it is shown countered before enough to accomplish its that two representations for time, one based on task? time intervals and another based on time One way to look at this questions is from a points, are fundamentally incompatible. That procedural perspective. For example, how does is, an agent whose time ontology is based on a compiler know how to interpret a symbol time points can never incorporate the axioms such as + in a computer language, or how does of another agent whose ontology for time is an agent system know what to do when it en- based on time intervals. From a logic perspec- counters the performative inform? The possibly tive, the two representations are like oil and informal semantics of these symbols are hard- water. Thus, I further assume that the concep- wired into a procedure by a human before- tualizations are compatible.
30 AI MAGAZINE Articles
Term heterogeneity and different model- An Example ing styles: Even if I assume a shared language I now consider a simple example of how ma- and compatible conceptualizations, it is still chine processing of formal semantics can be possible, indeed likely, that different people utilized to do something practical with today’s will build different ontologies for the same do- technology. As you can see, automatic ma- main. Two different terms can have the same chine processing of formal semantics is fraught meaning, and the same term can have two dif- with difficulties. I have made the following ferent meanings. The same concept can be simplifying assumptions: (1) all parties agree to modeled at different levels of detail. A given use the same representation language, (2) the idea can be modeled using different primitives conceptualizations are logically compatible, in the language. For example, is the idea of be- and (3) there are publicly declared concepts ing red modeled by having the attribute color that different agents can use to agree on mean- with value red, or is the idea modeled as a class ing. called something like RedThings? Is it both, Suppose that an agent is tasked with discov- where either (1) they are independent or (2) ering information about a variety of mechani- RedThings is a derived class defined in terms of cal devices. It encounters a web page with the the attribute color and the value red? text FUEL PUMP (figure 2). Lacking natural In the section entitled Machine-Processible language–understanding capability, the term is Semantics, I spoke of the intended versus the completely ambiguous. You can reduce the am- actual models of a logical theory. The former biguity by associating the text FUEL PUMP with correspond to what the author of the theory a formally defined term fuel-pump (this is called wanted to represent. The actual models corre- semantic markup). The agent might never have spond to what the author actually did repre- encountered this concept before. In this case, For a sent. The actual models consist of all the ob- the definition for the new term is defined in jects, relationships, and so on, in the real terms of the term pump, which, in turn, is de- computer to world that are consistent with the axioms. Be- fined in an externally shared hydraulics ontol- automatically cause the machine has access to the axioms, it ogy. The agent can learn that fuel-pump is a might, in principle, be possible for a computer subclass of pump, which, in turn, is a subclass determine the to determine whether two logical theories are of mechanical-device. The agent now knows that intended equivalent and, thus, whether the semantics of fuel-pump is not a typewriter or a spaceship be- meaning of a two terms are identical. This determination cause they are not kinds of pumps. The agent would be true, for example, if the two theories has no knowledge of what kind of pump it is, given term in had the same actual models. However, even if only that it is some kind of pump. However, this information is sufficient to allow the agent an ontology is the exact same language is used, and there is to return this document as relevant to substantial similarity in the underlying con- an impossible mechanical devices, even though it has never ceptualizations and assumptions, the inference task…. before heard of the term fuel-pump. It is possi- required to determine whether two terms actu- ble to do this processing with today’s technol- ally mean the same thing is intractable. ogy using research tools that have been devel- For a computer to automatically determine oped (Decker et al. 1999).20 This technology is the intended meaning of a given term in an also being commercialized (Staab and Maedche ontology is an impossible task, in principle; it 2001).21 Scale remains a significant barrier to would require seeing into the mind of the au- commercial success. thor. Therefore, a computer cannot determine This example illustrates the importance of whether the intended meaning of two terms is semantic markup and the sharing of ontolo- the same. This situation is analogous to formal gies. It also demonstrates the importance of specifications for software. The specification is formal ontologies and automated inference. what the author actually said he/she wanted Inference engines can be used to derive new in- the program to do. It might be possible to ver- formation for a wide variety of purposes; in ify that a computer program conforms to this particular, a formally specified ontology allows specification, but it will never be possible to agents to utilize theorem-proving and consis- verify that a program does what the author ac- tency-checking techniques to determine tually wanted it to do.19 whether they have agreement on the seman- To reduce the problems of term heterogene- tics of their terminology. ity and different modeling styles, I further as- The ability of the agent to infer something sume that the agent encounters a term that ex- about the meaning of fuel-pump depends on plicitly corresponds to a publicly declared the existence of a formal semantics for ontol- concept that it already knows about (for exam- ogy languages such as OWL. The language se- ple, using markup). mantics also allow the agent to infer the mean-
FALL 2003 31 Articles
Hey, I know this ontology Shared Hydraulics Ontology (SHO) (pump has (superclasses (mechanical-device)) (text-def (“A device for …”))) (every pump has (physical-parts (piston, valve, cylinder)) (device-purpose (Pumping-A-Fluid)))
The purpose of this review is to remind operators of the existence of the Operations Manual Bulletin 80 -1, which provides information regarding flight
Figure 2. Formal Semantics for Machine Processing. An agent is searching for information about mechanical devices, as defined in a shared hydraulics ontology (SHO). A document contains the term FUEL PUMP, which the agent has never encountered. Semantic markup reveals that it refers to the concept FUEL PUMP, which is a kind of pump, which is, in turn, defined in SHO as a kind of mechanical device. The agent infers that the document is relevant.
ing of complex expressions built up using lan- cent book on the semantic web (Daconta et. al. guage primitives. The semantics of the lan- 2003). For further discussion of inference on guage are not machine processible; they are the semantic web, see Horrocks (2002) and written for humans only. People use them to Jasper and Tyler (2001). write inference engines or other software to correctly interpret and manipulate expressions in the language. Why Do Web Note that today’s spectacularly impressive Shopping Agents Work? search engines by and large do not use formal semantics at all. Overall it remains an un- I have taken some time to consider what peo- proven conjecture that semantic approaches ple might mean when they talk about the se- will have significant impact anywhere on the mantic web. There appears to be consensus web. For example, there appear to be insuffi- that the key defining feature is machine-usable cient business drivers to motivate venture cap- web content. However, I argue that by this de- italists to heavily invest in semantic web com- finition there is an important sense in which panies. Fortunately, the W3C is moving the semantic web already exists. For example, forward on this issue by identifying a wide va- travel and bookseller shopping agents auto- riety of use cases (going well beyond search) to matically access web pages looking for good drive the requirements for a standard web on- deals. I don’t quibble about whether this tology language (OWL).22 Also, a strong business should “count” or how the definition of se- case for the semantic web is put forward in a re- mantic web might need to be modified accord-
32 AI MAGAZINE Articles ingly. It is more useful to regard these examples tions, and shared ontologies, one can get the collectively as a degenerate case of the seman- machine to process formal semantics specifica- tic web. In this section, I examine why web tions directly and do useful things (figure 2). shopping agents work, what their limitations Agreements and Public Declarations In are, and what you can expect in the future. general, the more agreement there is, the bet- Requirements for ter. Making agreements public is critical for the semantic web to take off. For example, the Machine-Usable Content news and magazine publishing industries have The following requirements are fundamental developed NEWSML and PRISM (publishing re- for enabling machines to make use of web con- quirements for industry standard metada- tent. ta).23,24 Agreements can also lessen the amount Requirement 1: The machine needs to of change, alleviating some maintenance is- know what to do with the content that it en- sues. counters. Although I brought up the issue of public For example, it needs to recognize that it has agreements in the context of machine-proces- found the content it is looking for and to exe- sible semantics, it is equally important when cute the appropriate procedures when it has the semantics are hardwired by the human. For been found. Ultimately, it is humans that write example, consider the Dublin core metadata el- By making the programs that enable the machines to do ement aet (DCMES), a set of 15 terms for describ- various the right thing. ing resources.25 The elements include such Requirement 2: Humans must know what things as title, subject, and date and are de- assumptions to do with the content that the program is ex- signed to facilitate search across different sub- regarding pected to encounter. ject areas. The meaning for these elements is Requirement 3: Humans know the meaning defined in English, not a formal language. Nev- languages, of the expected content or are able to encode a ertheless, if this meaning is hardwired into a conceptuali- procedure that can learn the meaning. web application, this application can make use In determining what makes the web shop- of web content that is marked up and points to zations, and ping agent examples work, consider the follow- the Dublin core elements. shared ing questions: (1) Hardwiring: What is hard- Semantics Specification Agreements about ontologies, wired and what isn’t? (2) Agreements and semantics should clearly be specified so that public declarations: How much agreement is humans can build reliable and correctly func- one can get there among different web sites in their use of tioning web applications. In the absence of the machine terminology and in the similarity of the con- agreements, effort is required to make sure that to process cepts being referred to? Are agreements pub- your application will do the right thing at each licly declared? (3) Specification of semantics: web site that uses terms in different ways (for formal To what extent are the semantics of the con- example, only some web sites include taxes in semantics tent clearly specified? Is it implicit, explicit and the price information), thus creating more informal, or formal? work in programming because a different ver- specifications Hardwiring The general case of automatical- sion is needed for every web site. directly and ly determining the meaning of web content is Even in the absence of agreements, it is very somewhere between intractable and impossi- important for the semantics of web site con- do useful ble. Thus, a human will always be hardwiring tent to clearly be specified (possibly informal- things…. some of the semantics into web applications. ly). Otherwise, there will be a lot of guesswork, The question is what is hardwired and what is thus undermining the reliability of applica- not? Shopping agent applications essentially tions. When information from different web hardwire the meaning of all the terms and pro- sites needs to be integrated, there must be cedures. The hardwiring enables the machine some way to map the different meanings to to know how to use the content. The hard- each other. Ontologies in conjunction with se- wiring approach is not robust to changes in mantic mapping and translation techniques web content. play a key role in semantic integration (Brad- The alternative is allowing the machine to shaw et al. 2003). process the semantics specifications directly. Thus, the semantics of the representation lan- Web Shopping Agents Work because… guages are made public and hardwired into the I regard the web shopping agents as degenerate inference engines used by the applications. examples of the semantic web. It is important This approach gives an additional degree of to understand why they are able to make use of flexibility because you do not hardwire the today’s web content in the apparent absence of meaning of every term. By making various as- semantics. sumptions regarding languages, conceptualiza- Here I show that they are able to make use of
FALL 2003 33 Articles
today’s web content because all three require- how to recognize the content they are looking ments from the section Requirements for Ma- for, and they must know what to do when they chine-Usable Content are met. I first provide find it. This knowledge requires access to the answers to the three questions in that section: meaning (that is, semantics) of the content, Question 1: Everything is hardwired. one way or the other. The manner in which the Question 2: There is no agreement among machine can access the semantics of web con- different web sites in their use of terminology, tent is at the heart of much confusion about although there is very strong overlap in the un- the semantic web. The main objective of this derlying concepts that are relevant. We are article was to shed light on this issue. aware of no public standards, although there I argued that paradoxically, today’s web could well be standard XML SCHEMA for the trav- shopping agents demonstrate the defining fea- el and bookseller industries, as there are for ture of a semantic web application without any other industries. explicit representation of semantics. Further- Question 3: I assume that the semantics of more, no one would seriously regard them as the terms and concepts are not specified at all, examples of the semantic web. I resolved the or if so, they are specified informally. paradox by regarding shopping agents as de- I now consider the three requirements for generate cases of the semantic web. I hope that machine-usable content in reverse order. this work will inspire people to generate “gen- Requirement 3: Humans know the meaning uine” examples of the semantic web. I antici- of the expected content, which seems surpris- pate that progress in development of the se- ing, given the lack of specification of terms and mantic web will take place by (1) moving along any public standards. The meaning is available the semantic continuum from less clearly spec- instead because there is sufficient human con- ified (implicit) semantics to more clearly spec- sensus on the use of terms such as price and ified (formal) semantics; (2) reducing the destination. One can think of this consensus as amount of hardwiring that is necessary or an implicit shared semantic repository, which en- changing which parts are hardwired and which Making ables web application developers to make edu- are not (This approach will entail a correspond- agreements cated guesses and develop useful software. ing increase in the amount of automated infer- public is Requirement 2: Humans know what to do ence to determine the meaning of web con- with the content, which follows from under- tent, thus enabling agents on the semantic web critical for the standing what content means and knowing to correctly perform their tasks. The impor- semantic web the specifications of the functions of the web tance of compelling use cases to drive the de- agents. mand for this approach cannot be underesti- to take off. Requirement 1: The machine knows what mated.); (3) increasing the amount of public to do with the content because the human pro- standards and agreements, thus reducing the grammers hardwired the semantics of the con- negative impact of today’s pervasive hetero- tent and created appropriate procedures to be geneities; and (4) developing technologies for executed. semantic mapping and translation for the Shopping agents can work even if there is no many cases where integration is necessary but automatic processing of semantics; they can where it is not possible to reach agreements. work without any formal representation of se- Finally, I want to stress that there is no need mantics; they can even work with no explicit for semantics envy. The needs of the applica- representation of semantics at all. The key to tion dictate the appropriate place to be along enabling shopping agents to automatically use the semantic continuum. What is right is what web content is that the meaning of the web works. For many applications, there is no need content that the agents are expected to en- for rigorous formal approaches because the counter can be determined by the human pro- cost is too high. There might not be a need for grammers who hardwire it into the web appli- machines to automatically determine the cation software. meaning of terms; the human can simply hard- wire this meaning into the software. Web shop- Summary and Conclusions ping agents know how to find the fare for a giv- en trip or the price of a book. Every browser There are many different views of what the se- knows that
means it is a second-level mantic web is and how it can or should evolve. header. There is no need to do inference; it is I attempted to show a variety of perspectives sufficient to hardwire the meaning of in- and possibilities. The most frequently quoted to the browser. I believe that in the short and, defining feature of the semantic web is ma- possibly, the medium term, approaches that do chine-usable web content. Fundamentally, this not make use of machine-processible seman- definition requires that machines must know tics are likely to have the most impact on the
34 AI MAGAZINE Articles development of the semantic web. Based on Canada, in June 2001. This article is a signifi- this analysis, I conjecture that the following is cantly revised and extended version of a short a law of the semantic web: article that appeared in a special issue of the The more agreement there is, the less it is Knowledge Engineering Review that contained necessary to have machine-processible se- papers from this workshop. mantics. Notes Eventually, there will be a need for automat- 1. Semantic Web Services Initiative. www.swsi.org. ed semantics processing. Relying too heavily 2. Also see World Wide Web Consortium. 2001. Se- on hardwiring semantics can result in different mantic Web Activity Statement. www.w3.org/2001/ implementations having different functions, sw/Activity. which, at best, means interoperation is diffi- 3. www.w3.org/TR/webont-req/. cult; at worst, there might be incorrect func- 2. World Wide Web Consortium. 2003. RDF Seman- tions. Another disadvantage of the hardwiring tics, ed. P. Hayes. www.w3.org/TR/rdf-mt/. approach is brittleness and consequent main- 4. World Wide Web Consortium. 1999. RESOURCE DE- tenance difficulties. SCRIPTION FRAMEWORK (RDF) Schema Specification, eds. In closing, there are many answers to the D. Brickly, R. Guha. www.w3.org/TR/1998/WD-rdf- question, Where are the semantics in the se- schema/. mantic web? First, they are often just in the hu- 5. Closed-World Machine. infomesh.net/2001/cwm/. man-as-unstated assumptions derived from 6. DAML. 2001. See www.daml.org/. implicit consensus (for example, price on a 7. This term is commonly used in the literature on travel or bookseller web site). Second, they are semantic integration of databases. in informal specification documents, for exam- 8. World Wide Web Consortium. 2003. LBASE: Se- ple, the semantics of UML or RDF SCHEMA. Third, mantics for Languages of the Semantic Web, eds. P. they are hardwired in implemented code (for Hayes and R. V. Guha. www.w3.org/TR/lbase/. example, in UML and RDF tools and in web shop- 9. Recently, DAML + OIL has evolved into a W3C stan- ping agents). Fourth, they are in formal speci- dard called the WEB ONTOLOGY LANGUAGE (OWL). fications to help humans understand or write www.w3.org/TR/OWL-REF/. code (for example, a modal logic specification 10. R. Cover. 1998. XML and Semantic Transparency, of the meaning of inform in an agent commu- The XML Cover Pages. www.oasis-open.org/cover/- nication language). Fifth, they are formally en- xmlAndSemantics.html. coded for machine processing, for example, 11. S. Weibel and E. MIller. 2000. An Introduction to (fuel-pump has (superclasses SHO: pump)). Dublin Core. www.xml.com/pub/a/2000/10/25/- Sixth, they are in the axiomatic and model- dublincore/. theoretic semantics of representation lan- 12. Object Management Group. 2000. OMG Unified guages (for example, the formal semantics of Modeling Language Specification, version 1.3. RDF). www.omg.org/technology/documents/formal/uni- Finally, I want to note that there are many fied modeling language.htm. other important issues for the semantic web 13. World Wide Web Consortium. 1999. RESOURCE DE- that I merely touched on or failed to address in SCRIPTION FRAMEWORK (RDF) Schema Specification, eds. D. Brickly, R. Guha. www.w3.org/TR/1998/WD-rdf- this article. These issues include web services, schema/. semantic markup, semantic integration, and 14. World Wide Web Consortium. 2003. LBASE: Se- use of natural language–processing techniques mantics for Languages of the Semantic Web, eds. P. to glean the semantics of natural language doc- Hayes and R. V. Guha. www.w3.org/TR/lbase/. uments. 15. For simplicity, this distinction is not shown as Acknowledgments two separate points on the continuum in figure 1. 16. UML Working Group. 2001. The Precise UML This article was inspired by a stimulating e- Group home page. www.puml.org. mail debate with Dieter Fensel on how and 17. F. van Harmelen, I. Horrocks, and P. Patel-Schnei- whether XML has the ability to carry semantic der. 2001. Model-Theoretic Semantics for DAML + OIL. information. This article benefited from many World Wide Web Consortium Note 18. www.w3.org/ discussions with Peter Clark, Rob Jasper, TR/daml+oil-model. Michael Gruninger, John Thompson, and 18. Pat Hayes, personal communication, 2002. Frank van Harmelen. I am grateful for valuable 19. A much more detailed discussion of these formal feedback from Peter Clark, John Thompson, issues can be found in Complete Semantic Integration, Pat Hayes, and an anonymous referee on a pri- 2003, M. Gruninger and M. Uschold (forthcoming). or draft. The content of this article was first 20. ONTOPRISE. 2001. ONTOPRISE: Semantics for the presented as an invited talk at the Ontologies Web. www.ontoprise.de/com/. in Agent Systems Workshop held at the Au- 21. ONTOPRISE. 2001. ONTOPRISE: Semantics for the tonomous Agents Conference in Montreal, Web. www.ontoprise.de/com/.
FALL 2003 35 Articles
22. World Wide Web Consortium. 2003. Web 2002), 24–28 March, Prague, The Czech Republic. Ontology Language (OWL) Use Cases and Require- Jasper, R., and Tyler, A. 2001. The Role of Semantics ments. www.w3.org/TR/webont-req/. and Inference in the Semantic Web: A Commercial 23. International Press Telecommunications Council Challenge. Paper presented at the International Se- (IPTC). 2000. NewsML, Version 1.0 Functional Spec- mantic Web Working Symposium, 30 July–1 August, ification. citeseer.nj.nec.com/452173.html. Stanford, California. 24. PRISM Working Group. 2001. PRISM: Publishing Jasper, R., and Uschold, M. 2003. Enabling Task-Cen- Requirements for Industry Standard Metadata Ver- tered Knowledge Support through Semantic Metada- sion 1.0. www.prismstandard.org/techdev/prism- ta. In Spinning the Semantic Web, eds. D. Fensel, J. spec1.asp. Hendler, H. Lieberman, and W. Wahlster, 223–251. 25. S. Weibel and E. Miller. 2000. An Introduction to Cambridge, Mass.: MIT Press. Dublin Core. www.xml.com/pub/a/2000/10/25/- Ouksel, A., and Sheth, A. 1999. A Brief Introduction dublincore/. to the Research Area and the Special Section. SIG- MOD Record (Special Section on Semantic Interoper- References ability in Global Information Systems) 28(1): 5–12. Also see www.acm.org/sigmod/record/issues/9903/. Berners-Lee, T.; Hendler, J.; and Lassila, O. 2001. The Semantic Web. Scientific American 284(5): 34–43. Smith, R. 2001. What’s Required in Knowledge Tech- nologies—A Practical View. Paper presented at Bradshaw, J. M.; Boy, G.; Durfee, E.; Gruninger, M.; Knowledge Technologies Conference, 4–7 March, Hexmoor, H.; Suri, N.; Tambe, M.; Uschold, M.; and Austin, Texas. Also see www.gca.org/attend/2001_- Vitek, J., eds. 2003. Software Agents for the Warfight- conferences/kt_2001/default.htm. er. Menlo Park, Calif.: AAAI Press. Forthcoming. Smith, I.; Cohen, P.; Bradshaw, J.; Greaves, M.; and Daconta, M. C.; Obrst, L. J.; and Smith, K. T. 2003. Holmback, H. 1998. Designing Conversation Policies The Semantic Web—A Guide to the Future of XML, Web using Joint Intention Theory. Paper presented at the Services, and Knowledge Management. New York: Wi- Third International Conference on Multiagent Sys- ley. tems (ICMAS-98), 3–7 July, Paris, France. Decker, S.; Erdmann, M.; Fensel, D.; and Studer; R. Staab, S., and Maedche, A. 2001. Knowledge Por- 1999. ONTOBROKER: Ontology-Based Access to Distrib- tals—Ontologies at Work. AI Magazine 21(2): 63–75. uted and Semistructured Information. In Semantic Is- Trippe, B. 2001. Taxonomies and Topic Maps: Cate- sues in Multimedia Systems, Proceedings of DS-8, ed. R. gorization Steps Forward. Econtent Magazine, August. Meersman, Z. Tari, and S. Stevens, 351–369. Boston: www.econtentmag.com/Articles/ArticleReader.aspx. Kluwer Academic. See also ontobroker.aifb.uni-karl- sruhe.de/index_ob.html. Uschold, M.; King, M.; Moralee, S.; and Zorgios, Y. 1998. The Enterprise Ontology. The Knowledge Engi- Evans, A. S.; France, R. B.; Lano, K. C.; and Rump, B. neering Review (Special Issue on Putting Ontologies to 1998. The UML as a Formal Modeling Notation. In Use) 13(1): 31–89. UML’98—Beyond the Notation, 297–307. Lecture Notes in Computer Science 1723. New York: Springer. Fikes, R., and McGuinness, D. 2001. An Axiomatic Semantics for RDF, RDF SCHEMA, and DAML + OIL, KSL Technical Report KSL-01-01, Knowledge Systems Laboratory, Stanford University. See www.ksl.Stan- Michael Uschold is a research sci- ford.EDU/people/dlm/daml-semantics/abstract-ax- entist at Boeing, Phantom Works. iomatic-semantics.html. His interests center on the core area Gruninger, M., and Uschold, M. 2003. Complete Se- of developing and applying ontolo- mantic Integration. Forthcoming. gies, including the semantic web, Gruber, T. R. 1993. A Translation Approach to semantic integration, and world Portable Ontology Specifications. Knowledge Acquisi- modeling for autonomous systems. tion 6(2): 199–221. Uschold is on the industrial adviso- Guarino, N.; Carrara, M.; and Giaretta, P. 1994. An ry boards of various international projects and initia- Ontology of Meta-Level Categories.” In Principles of tives related to these areas. He receieved his B.S. in Knowledge Representation and Reasoning: Proceedings of mathematics and physics, a masters in computer sci- the Fourth International Conference (KR94), eds. E. ence, and a Ph.D. in artificial intelligence. From 1983 to 1987, he worked at the University of Edinburgh. Sandewall and P. Torasso, 270–280. San Francisco, Calif.: Morgan Kaufmann. Hayes, P. 1996. A Catalog of Temporal Theories. Technical Report, UIUC-BI-AI-96-01, University of Illinois. Hendler, J. 2001. Agents on the Semantic Web. IEEE Intelligent Systems 16(2): 30–37. Horrocks, I. 2002. DAML + OIL: A Reasonable Web On- tology Language. Paper presented at the Eighth Con- ference on Extending Database Technology (EDBT
36 AI MAGAZINE