Toward meaningful communication: Engineering semantics with the geometry of conceptual spaces

Abstract

As the demand for data transmission around the globe continues to rise, we face a challenge brought about by the fundamental limits of communication imposed by Claude Shannon's information theory. Modern communication systems are able to operate with a rate close to the theoretical channel capacity, meaning we must increase this capacity in order to transmit additional data. This is achieved by either increasing the power of the signal or bandwidth of the channel, neither of which are desirable options. Alternatively, what if we could communicate more efficiently given the resources we already have? This is the goal of semantic communication, which is a paradigm that aims to communicate a given meaning rather than exactly reproduce bits at the receiver, potentially resulting in efficient semantic representations that can communicate the same meaning using less physical bits. However, modern approaches to semantic communication suffer from a lack of consensus on how to quantify and optimize a system around the notion of "meaning" and a reliance on black-box machine learning models that obscure the true functionality of semantic communication modules. In this dissertation, we introduce an approach termed "semantic communication with conceptual spaces," which is based on the geometric conceptual spaces model of meaning. Conceptual spaces offer a general, expressive, and interpretable mechanism by which a system can utilize semantic knowledge and optimize its operation around this knowledge. We first lay the mathematical groundwork for such a system, and demonstrate its potential for efficient semantic communication via simulations. We then address a significant challenge of the conceptual space-based approach, which is obtaining the conceptual space model itself. To overcome this challenge, we develop a novel machine learning architecture capable of learning the complex domains of a conceptual space model using only high-level semantic property information. Next, we introduce a causal reasoning-based mechanism into the proposed system, which allows the semantic communication system to determine which semantic elements are most important for performing a given task. This reasoning mechanism enables the system to achieve even greater efficiency by transmitting only the semantics that are important for the task at hand. Finally, we connect our approach to traditional information theory by deriving an upper bound for the goal-oriented rate distortion function. This bound is based on a fixed threshold for the semantic distortion within the system and the novel notion of distortion discrepancy. In short, this dissertation proposes, for the first time, a method of semantic communication based on a general and expressive model of meaning that allows for automated learning of semantics, causal reasoning-based optimization, and theoretical analysis similar to classical information theory.