Recently Sperberg-McQueen and others have argued that markup functions by licensing inferences about a text. They remark, however, that the information warranting such inferences may not be entirely explicit in the syntax of the markup language used to encode the text. (Sperberg-McQueen et al., 2001) For example, a language defined in SGML or XML may include an attribute (such as 'lang') that an encoder may apply to an element with the generic identifier 'QUOT.' One might then infer that the QUOT element marks an identifiable component of the document (called a quotation) and that the quotation has the property of being in a particular language (as indicated by the 'lang' attribute). It may also be valid to infer that children of the 'QUOT' element share the property of being in that language, unless overridden with a language attribute of their own. On the other hand, there may not be such a simple one-to-one mapping between components and elements: for example, a single quotation may be broken across two or more 'QUOT' elements. There are a number of other inferences that are typically assumed by tag set designers and application designers alike, but which cannot be formally expressed in the DTD, and may or may not be informally expressed in the tag set documentation. In order to adequately represent such inferences (the "meaning of markup") the Sperberg-McQueen group developed techniques for expressing in predicate logic, (i) the facts signalled by the encoding of a particular document instance and (ii) the logical relationships commonly understood to exist and license further inferences. A Prolog database was used to demonstrate the effectiveness of this approach. The present paper builds directly on this previous work, and reflects new results which provide more rigorous and explanatory layers of abstraction and progress in understanding problems with "deictic" expressions and domains of variables, etc. But the fundmental new result presented is the completion of a complete integrated working system with an entirely new and substantially redesigned Prolog database at its core. This Prolog database has been redesigned to improve functionally, better reflect the theoretical results, and increase functionality, flexibility, and performance. The system permits an analyst to specify facts about the markup syntax (e.g., generic identifiers and attribute values) separately from facts and rules of inference about semantic entities and properties. The system provides a level of abstraction at which the performative or interpretive meaning of the markup can be explicitly represented in machine-readable and executable form. Inferences can then be drawn regarding document components, including problematic structures, such as those participating in overlapping hierarchies. The new Prolog database is integrated with an SGML/XML parser so that SGML and XML instances can be input and output. Facts and rules of inference concerning the document are expressed in Prolog's standard declarative syntax. We have developed a collection of predicates that emulate a subset of the W3C's Document Object Model methods for navigating the hierarchical structure of nodes, and retrieving attribute values and information from the document type definition. These predicates afford a clear separation of the syntactic information captured by the parser and the document semantics expressed by the analyst. Another collection of predicates support deictic expression resolution. These allow rules of inference to include location-relative pointing from one part of the document to another. For example, we have predicates for resolving an element's closest ancestor having a particular generic identifier, attribute, or attribute value pair. Another set of predicates resolves the identity of an element of a particular type occurring most closely in terms of the linear structure of the document (rather than the closest in the hierarchy). A third set of predicates supports the tracing of connections across elements, such as those linked by ID and IDREF attribute pairs. Rules (axioms) represent the further logical relationships mentioned above, such as for defeasible inheritance, distribution of distributive properties, etc. In developing the architecture of this system, we have adopted an object-oriented strategy: each node identified by the parser and semantic entity instantiated via a rule of inference has a unique identifier assigned by the system. Predicates for retrieving or manipulating that information are written with the aim of hiding the underlying data structure. The system architecture can therefore be understood to have several distinct layers of representation:

• 1. A parser that handles the serialized document instance.
• 2. Predicates for processing the output of the parser.
• 3. Predicates for storing a representation of the parse tree in the Prolog database.
• 4. Predicates for emulating DOM methods, deictic expression resolution, object instantiation, and general characteristics of properties (such as their inheritance and distribution).
• 5. Facts and rules of inference expressing the document semantics.

The first two layers are implementation-dependent, and are designed to be modular, allowing experimentation to improve robustness and scalability. We intend the upper layers to be consistent across different implementations of the system. Currently the interface between the lower and upper layers in written entirely in Prolog, but we may employ other technologies (such as XSLT) in the future. This system is being developed with the aim of advancing several interrelated research goals. We will develop applications that provide a complete formal account of the semantics of particular document classes. The system will also provide an environment for experimenting with proposed or conjectured semantics with the goal of improving document retrieval and conversion applications. We also propose to build applications that draw inferences based on both document semantics and domain or world knowledge. At this time the new Prolog database has been completed and tested and the entire working system can be demonstrated on small fragments of TEI. By the conference we hope to have completed the representation of the markup semantics of two XML systems (XHTML and TEI-lite), which will allow us to be able to present a substantial demonstration of the practical advantages of representing markup semantics, as well as the theoretical soundness of this approach to the meaning of markup.