DATR is a declarative knowledge representation language designed for encoding default inheritance networks. It was introduced in 1989 by Roger Evans and Gerald Gazdar at the University of Sussex. The language is used in computational linguistics to represent lexical knowledge, such as morphological, syntactic, and phonological properties of words. DATR is a subfield of artificial intelligence and is notable for its formal semantics and support for default and multiple inheritance.
Historical context
DATR emerged during the late 1980s, when symbolic approaches dominated artificial intelligence research. Institutions such as MIT CSAIL, Stanford AI Lab, and Carnegie Mellon University were developing knowledge representation systems and logic-based formalisms. Research at Xerox PARC and Bell Labs contributed foundational work in programming languages and AI. Later, Berkeley AI Research and the University of Toronto became centers for connectionist and statistical approaches, which contrasted with the symbolic tradition that produced DATR.
Syntax and semantics
In DATR, a knowledge base consists of a set of nodes, each containing path equations. A path is a sequence of atoms, and a node can define values for paths. For example, a node for the noun "dog" might specify <root> = "dog" and <plural> = "dogs". Default inheritance is achieved by ordering equations, so that more specific nodes override more general ones. The language also supports multiple inheritance, where a node can inherit from several parents. The formal semantics of DATR is defined using a least fixed-point construction over a global inheritance network, which guarantees a unique model for any well-formed knowledge base.
The language has a simple syntax with no variables. Paths are written as sequences of atoms enclosed in angle brackets, and equations use the equals sign. Nodes can be defined by listing equations, and inheritance is expressed by referring to other nodes. DATR also supports orthogonal inheritance, where different paths can inherit from different parents independently.
Implementation
The original implementation of DATR was written in Prolog, and the language was designed to be easily embedded in larger natural language processing systems. A formal specification of DATR was published by Evans and Gazdar, providing a denotational semantics that maps knowledge bases to models. This specification allowed DATR to be implemented in other programming languages, and several later versions were developed for research purposes.
Applications
DATR was primarily used in computational linguistics for representing lexical information. It was a central component of the ACQUILEX project (1989-1993), which built multilingual lexical knowledge bases for English, Italian, Dutch, and Spanish. The language was also applied to research on morphology, syntax, and phonology, where default inheritance naturally captures regularities and exceptions. For example, a general noun node might define a default plural form, while a specific node for "child" overrides it with "children". DATR's explicit encoding of lexical knowledge contrasts with modern large language models, which learn statistical patterns from large corpora. It also differs from generative AI systems that produce text, as DATR is a declarative formalism rather than a generative model.
Influence and legacy
Although DATR is no longer widely used, it remains an important example of a symbolic knowledge representation formalism. Its principles of default inheritance and multiple inheritance have influenced later knowledge representation languages and ontologies. DATR is often contrasted with machine learning and deep learning approaches, which rely on neural networks and transformers. Modern AI systems such as those developed by OpenAI and Google DeepMind are based on deep learning rather than symbolic knowledge representation. DATR thus serves as a historical reference point for understanding the shift from symbolic AI to statistical and connectionist methods.