< Back to previous page

Project

Bitstring Semantics at the interface between psychology and computer science

This interdisciplinary PhD research, as the psychology and computer science part of the interdisciplinary BITSHARE project, will be carried out in the context of a tight collaboration between the Laboratory for Experimental Psychology of the Faculty of Psychology and Educational Sciences and the EAVISE research group of the Faculty of Engineering Technology at De Nayer Campus in Sint-Katelijne-Waver.

Bitstring semantics, which was recently developed in logic, is a formalism for exploring the meaning relations between different expressions. Because it operates in a thoroughly local fashion, this formalism is able to achieve a good balance between high expressivity and low mathematical/computational complexity. The overarching goal of the BITSHARE project is to explore new applications of bitstring semantics in other disciplines that are concerned with meaning and reasoning, viz., philosophy, linguistics, psychology and computer science. Our target applications include some of the most important research topics in these disciplines today, such as scalar implicatures, the semantics/pragmatics interface, collaborative knowledge extraction, and probabilistic causal models. Through these well-chosen target applications, the project will demonstrate that bitstring semantics offers a shared representation formalism for a broad, interdisciplinary community of reasoning researchers, in which the insights from different disciplines can be expressed and combined with each other.

Within computer science and cognitive science, the field of Knowledge Representation & Reasoning investigates how human experts can provide domain knowledge to a computer system in such a way that this knowledge can be used to solve different kinds of problems. This knowledge elicitation process requires the careful design of appropriate Knowledge Representation languages. The EAVISE research group brings knowledge representation technology to industrial applications. For instance, the group are currently collaborating with a company that wishes to automate some of its design activities. EAVISE knowledge engineers have therefore conducted knowledge extraction workshops with the company’s domain experts (i.e. product designers), which resulted in a formal representation of the relevant domain knowledge. By applying appropriate logical inference methods to this formal representation, products can be designed (semi-)automatically.

A key step in this entire process is the knowledge extraction phase: any errors in this phase will render the eventual system unreliable. Traditionally, knowledge extraction has been a one-way process, in which the domain experts provide information which the knowledge engineer then writes down in a logical formalism that is incomprehensible for the domain experts. However, this process often gives rise to misunderstandings, thus leading to errors in the system and hampering the domain experts’ confidence in the produced results. Therefore a collaborative process has recently been adopted, in which the knowledge engineer and the domain experts work together to construct the formal representation. Here, a fundamental trade-off exists between expressivity of the language and its ease-of-use for the domain expert. Therefore, this promising collaborative approach hinges on the availability of a suitable representation language: it must be expressive enough to allow all the relevant knowledge to be represented, yet also simple enough to be understood by domain experts with no background in ICT, logic or mathematics. The recent Decision Model and Notation (DMN) standard has been developed specifically in order to be easy for domain experts. So far, this standard has been adopted and extended further, e.g. cDMN. However, it still lacks in the expressivity that is required for some kinds of applications that we want to tackle.

In this research, we will investigate extensions to the DMN standard that increase the expressivity of the language without making it harder to use for domain experts. Bitstring semantics offers an appealing setting in which to analyze this trade-off between expressivity and usage-difficulty. Such an analysis will be concerned both with the formal properties of the language (a computer science perspective) and with how well it matches the way in which experts reason about their domain (a psychology perspective). We address three concrete research questions: 
             (1) what are the most suitable logical connectives to add to the language (e.g., in light of linguistic and psychological findings, the 'some but not all'-quantifier might be more intuitive for domain experts than the existential quantifier from classical logic)? 
             (2) can we classify different possible combinations of connectives into a complexity hierarchy, not according to computational complexity, but according to the complexity of understanding for domain experts? 
             (3) can we extend bitstring semantics from logical connectives to more knowledge-rich terms (e.g., suitable, intelligent, happy, good)? 
Bitstring semantics is ideally suited to address these questions, because it allows us to consider different ways of putting fundamental concepts together easily and it provides a semantically meaningful way of measuring their ‘understanding complexity’ through the size of the required bitstrings. The effectiveness of collaborative learning is well-supported by existing psychological research. For high-load tasks, such as knowledge extraction, collaboration increases effective working memory capacity because the task can be executed while relying on the working memories of more than one task performer. Finally, representing knowledge-rich terms is related to the topic of scalar diversity, and will be tackled experimentally.

To conclude, this research will consist in defining a number of orthogonal extensions to the DMN language — each aimed at preserving a specific aspect of its user-friendliness —, combining these extensions in various ways, and conducting experiments to assess how well domain experts are able to work with these combinations. This research will allow us to boil the DMN standard down to its essential aspects, and deliver a framework within bitstring semantics that explains how these aspects can be generalized and recombined in order to produce more expressive languages for domain experts, whose ‘understanding complexity’ we can formally measure using the bitstring apparatus. Experimental-psychological work will be essential to gain a realistic representation of human understanding complexity, which is crucial for the success of the envisaged enrichment of the DMN standard.

Date:1 Jul 2020  →  Today
Keywords:Artificial Intelligence, Knowledge Representation, Human Reasoning
Disciplines:Knowledge representation and reasoning, Cognitive science and intelligent systems not elsewhere classified, Artificial intelligence, Knowledge management, Cognitive processes, Logic, Computational logic and formal languages, Mathematical psychology
Project type:PhD project