Human-centric Ontology Evaluation (HOnEst)

In recent years, great breakthroughs have been made in the area of novel intelligent systems such as question answering systems (e.g., IBMs Watson) or conversational agents, also known as chatbots. It is naturally assumed that these systems behave in a way that is both truthful (making factually correct statements) and unbiased (acknowledging multiple, possibly contradictor viewpoints). Yet, the behaviour of these systems highly depends on the quality of world knowledge that they possess. Such world knowledge is often stored in advanced data structures known as ontologies which are typically very large and are extracted automatically from massive text collections. Unfortunately, automatic extraction algorithms are prone to deliver ontologies that contain factual errors or subscribe to a particular viewpoint. Identifying most of such shortcomings is, currently, still beyond being automated and is best performed with human involvement. The aim of HOnEst is to improve the quality of ontologies through scalable human-centric evaluation, so that they enable a factually correct and unbiased (i.e., honest) system behaviour. To that end, HOnEst will firstly refer to current literature to identify those ontology shortcomings that have been reported as requiring humans for their identification. Secondly, it will explore the use of Human Computation (HC) techniques to effectively and efficiently organize human-centric evaluation, by splitting the evaluation task into multiple smaller tasks that can be solved in parallel by a large and distributed workforce. The project will rely on the Design Science method to create HC-based solutions to the evaluation problem and then assess these in order to better understand which ontology shortcomings can actually be identified with HC. Finally, to address scalability concerns, HOnEst proposes the combined use of human and algorithmic components in the context of two real-life, very large ontologies: (1) the CSO ontology containing 15K topics about computer science research and (2) WebIsALOD which aims to increase the diversity of knowledge available to information systems with its 400M automatically extracted relations between so called tail entities (i.e., not very well-known entities that are typically not covered by other state-of-the-art large ontologies, such as those that reflect Wikipedia). In both cases, algorithmic components will determine which parts of the ontology should be evaluated by humans and improve their functionality based on this input, with the possibility of actually learning enough to perform (nearly) as well as humans do. Besides its novel scientific contributions, HOnEst has important implications on ensuring truthful and unbiased behaviour of ontology-based question answering and conversational systems. Such benefits will be already enjoyed by the systems that make use of the two ontologies improved as part of the project.

In recent years, Artificial Intelligence (AI) has undergone tremendous technological developments that have significant impact on all aspects of society from economical changes to sustainability concerns. In particular, we have witnessed the major impact of AI techniques that derive intelligence from large-amounts of data (also known as neural AI). A key representative of this trend are foundational models (frequently leveraged by conversational agents such as ChatGPT) which demonstrated the power of these technologies while highlighting their limitations in terms of lack of transparency in terms of their internal operation and weak factual correctness. As such, the key question arises: how can we leverage the tremendous power of such neural AI techniques while ensuring that AI systems remain transparent and behave truthfully by delivering factually correct information? As a solution approach, "neurosymbolic" architectures augment neural systems (such as foundational models) with factually correct knowledge captured using symbolic AI techniques. In particular, advanced data structures known as ontologies are suitable to capture verified and factually correct knowledge. Ontologies are typically automatically extracted from massive text collections and undergo a quality check that is often performed with human involvement. The aim of the HOnEst project is to improve the quality of ontologies through scalable human-centric evaluation (HCE), so that they enable a factually correct and unbiased (i.e., honest) behaviour of AI systems that rely on them. To address this goal, the project achieved the following results. First, to fully understand this topic, we systematically surveyed 15 years of work on human-centric ontology evaluation. Second, we experimentally explored solutions to ontology evaluation by splitting the evaluation task into multiple smaller tasks that can be solved in parallel by a large and distributed workforce. We proposed reusable process models and technical architectures that others can employ for their own HCE campaigns. Thirdly, we focused on making HCE scalable by combining human and algorithmic components. To that end, we investigated suitable neurosymbolic architectures that combine ontologies and neural AI systems, leading to the identification of over 40 such architecture types. Also, we proposed hybrid human-AI workflows in which humans and algorithmic elements (namely, foundational models) team up for verifying ontologies. We experimentally studied these workflows for verifying the Computer Science Knowledge Graph (CS-KG) which is an ontology-based large scale resource containing 41 Million facts about the Computer Science domain. As such, HOnEst has delivered an essential piece of the overarching puzzle of ensuring powerful and truthful AI systems thus contributing to the safe and effective uptake of these powerful technologies even in mission critical domains where factual correctness is key.