Hausser R. Computational Linguistics and Talking Robots. Processing Content in Database Semantics

Springer, 2011. — 297 p.The practical task of building a talking robot requires a theory of how natural language communication works. Conversely, the best way to computationally verify a theory of free natural language communication is by demonstrating its functioning concretely in the form of a talking robot as the epitome of human-machine communication.
To build an actual robot requires hardware which provides recognition and action interfaces appropriate for this task – like those of C3PO, the bronzy protocol robot in the Star Wars movies. Because such a hardware is presently hard to come by, our method is theoretical: we present an artificial cognitive agent with language as a software system called Database Semantics (DBS).
DBS complements work in robotics which does include the interface hardware for reconstructing the sensorimotor loop. For example, Roy (2003, 2008) has been working on an actual robot with language, called Ripley.
Ripley and the DBS system have in common that they are grounded. This is because both approaches are agent-oriented, in contradistinction to Phrase Structure Grammar and Truth-Conditional Semantics, which are sign-oriented.
Unlike the fictional C3PO, the real Ripley consists of a robotic arm mounted on a table. Attached to the arm is a hand, a video camera, a microphone, and a loudspeaker. To manage the difficulties of building a robot as a hardware machine, Ripley is focused on simple language use, such as that of small children (Roy 2008, p. 3). The work concentrates on implementing examples, such as Human: Hand me the blue one, referring to a cup on Ripley’s table (op. cit., p. 10). This approach requires expending much time and labor on practical troubles like motor burnout from overexertion (op. cit., p. 19).
A theoretical approach, in contrast, does not have to deal with the technical difficulties of hardware engineering. Thus there is no reason to simplify our system to a toddler’s use of language. Instead the software components of DBS aim at completeness of function and data coverage in word form recognition, syntactic-semantic interpretation, inferencing, and so on, leaving the procedural implementation of elementary concepts, e.g., blue or cup, for later.
In the meantime, DBS uses placeholders for concepts assumed to be implemented as elementary recognition and action procedures of the agent. Embedded into flat feature structures called proplets, the concepts help to build the agent’s context component for non-language cognition, and are reused as meanings of the agent’s language component. In the hear and speak modes, the two components interact automatically on the basis of pattern matching.
Language proplets differ from context proplets in that language proplets have a language-dependent surface3 as value, while context proplets do not. As a consequence, the context level may be constructed from the language level simply by omitting the surfaces of the language proplets (4.3.3).
The export of language-level constructs to the context level allows us to make the structural corelation between the two levels simple and direct. For example, there is no processing in language interpretation and production except for the literal mapping from the context to the language component in the speak mode and from the language to the context component in the hear mode. All non-literal interpretation, for example, of a metaphor, is done by inferencing before the utterance in the speak mode and after in the hear mode.4
The compositional semantics of the language component, i.e., coordination and functor-argument, intra- and extrapropositional, and the basic parts of speech, are reused in the context component. By connecting the context component to the agent’s interfaces, recognition and action at the non-language level may be processed in the same way as at the language level.
The direct interaction between the language and the context5 components implies linguistic relativism, also known as the Humboldt-Sapir-Whorf hypothesis. It is counterbalanced by the hypothesis that the different natural languages all work essentially the same way. This is supported by the wide range of languages6 which have been analyzed using the DBS approach.
The similarity of language processing and context processing in DBS is motivated by function and simplicity of design. The similarity poses also an empirical question in psychology and neurology. It is not implausible, however, that language processing and context processing are cut from the same cloth.Introduction: How to Build a Talking Robot
Five Mysteries of Natural Language Communication
Mystery Number One: Using Unanalyzed External Surfaces
Mystery Number Two: Natural Language Communication Cycle
Mystery Number Three: Memory Structure
Mystery Number Four: Autonomous Control
Mystery Number Five: Learning
The Coding of Content
Compositional Semantics
Simultaneous Amalgamation
Graph Theory
Computing Perspective in Dialogue
Computing Perspective in Text
Final Chapter