Low-level representations have proven to be good at certain kinds of adaptive learning. High-level representations make effective use of existing knowledge and perform inference well. To promote using both forms of representation cooperatively rather than engaging in the perennial sectarian debate of supporting one paradigm at the expense of the other, this thesis presents a prototype system demonstrating knowledge retention using genetic algorithms and multiple levels of representation and learning. The prototype uses a mid-level of representation and transformations upward and downward for retaining domain-specific knowledge to bridge the gap between the high-level representation and learning and the genetic algorithm level. The thesis begins with an overview of the work, briefly introduces the principles of genetic algorithms, and states an illustrative domain. Then it reviews related work and two supportive systems. After that, it gives a general description of the prototype system's structure, three levels of representation, two transformations, and three levels of learning. Next, it describes methods of implementing the prototype system in some detail. Finally, it shows results with discussion, and points out conclusions and future work. / Master of Science

The study addresses the automated translation of architectural drawings from 2D Computer Aided Drafting (CAD) data into a Building Information Model (BIM), with emphasis on the nature, possible role, and limitations of a drafting language Knowledge Representation (KR) on the problem and process. The central idea is that CAD to BIM translation is a complex diagrammatic interpretation problem requiring a domain (drafting language) KR to render it tractable and that such a KR can take the form of an information model. Formal notions of drawing-as-language have been advanced and studied quite extensively for close to 25 years. The analogy implicitly encourages comparison between problem structures in both domains, revealing important similarities and offering guidance from the more mature field of Natural Language Understanding (NLU). The primary insight we derive from NLU involves the central role that a formal language description plays in guiding the process of interpretation (inferential reasoning), and the notable absence of a comparable specification for architectural drafting. We adopt a modified version of Engelhard's approach which expresses drawing structure in terms of a symbol set, a set of relationships, and a set of compositional frameworks in which they are composed. We further define an approach for establishing the features of this KR, drawing upon related work on conceptual frameworks for diagrammatic reasoning systems. We augment this with observation of human subjects performing a number of drafting interpretation exercises and derive some understanding of its inferential nature therefrom. We consider this indicative of the potential range of inferential processes a computational drafting model should ideally support. The KR is implemented as an information model using the EXPRESS language because it is in the public domain and is the implementation language of the target Industry Foundation Classes (IFC) model. We draw extensively from the IFC library to demonstrate that it can be applied in this manner, and apply the MVD methodology in defining the scope and interface of the DOM and IFC. This simplifies the IFC translation process significantly and minimizes the need for mapping. We conclude on the basis of selective implementations that a model reflecting the principles and features we define can indeed provide needed and otherwise unavailable support in drafting interpretation and other problems involving reasoning with this class of diagrammatic representations.

There is almost a universal agreement among scholars in information retrieval (IR) research that knowledge representation needs improvement. As core component of an IR system, improvement of the knowledge representation system has so far involved manipulation of this component based on principles such as vector space, probabilistic approach, inference network, and language modeling, yet the required improvement is still far from fruition. One promising approach that is highly touted to offer a potential solution exists in the cognitive paradigm, where knowledge representation practice should involve, or start from, modeling the human conceptual system. This study based on two related cognitive theories: the theory-based approach to concept representation and the psychological theory of semantic relations, ventured to explore the connection between the human conceptual model and the knowledge representation model (represented by samples of concepts and relations from the unified medical language system, UMLS). Guided by these cognitive theories and based on related and appropriate data-analytic tools, such as nonmetric multidimensional scaling, hierarchical clustering, and content analysis, this study aimed to conduct an exploratory investigation to answer four related questions. Divided into two groups, a total of 89 research participants took part in two sets of cognitive tasks. The first group (49 participants) sorted 60 food names into categories followed by simultaneous description of the derived categories to explain the rationale for category judgment. The second group (40 participants) performed sorting 47 semantic relations (the nonhierarchical associative types) into 5 categories known a priori. Three datasets resulted as a result of the cognitive tasks: food-sorting data, relation-sorting data, and free and unstructured text of category descriptions. Using the data analytic tools mentioned, data analysis was carried out and important results and findings were obtained that offer plausible explanations to the 4 research questions. Major results include the following: (a) through discriminant analysis category members were predicted consistently in 70% of the time; (b) the categorization bases are largely simplified rules, naïve explanations, and feature-based; (c) individuals theoretical explanation remains valid and stays stable across category members; (d) the human conceptual model can be fairly reconstructed in a low-dimensional space where 93% of the variance in the dimensional space is accounted for by the subjects performance; (e) participants consistently classify 29 of the 47 semantic relations; and, (f) individuals perform better in the functional and spatial dimensions of the semantic relations classification task and perform poorly in the conceptual dimension.

Design is hard and needs to be supported by software. One of the ways software can support designers is by providing analogical reasoning. To make analogical reasoning work well, the software makers need to know how to create a knowledge representation that will facilitate the kind of analogies that the designers want. This thesis will inform software makers by experimenting with two kinds of knowledge representations, called device-centric (DC) and environment-centric (EC), and to try to determine the relative benefits of using either one of them for analogical matching. We performed computational experiments, using Structure Mapping Engine for matching, to determine the quantity and quality of analogical matches that are produced when the representation is varied. We conducted a limited human experiment, using questionnaires and repertory grids, to determine if any of the computational results were novel, and to determine if the human similarity ratings between devices correlated with the computer results. We show that design software should use DC representations to produce a few focused matches which have high average weight. It should use EC representations to produce many matches some of high weight and some of low weight. Based on our human experiment, design software can use either DC or EC representations to produce novel matches. Our experiments also show that human matches correlate most strongly with a combined DC and EC representation and that their similarity reasons are more EC than DC. This suggests that designers tend to think more in EC terms than in DC terms.

<div> <div> <div> <div> <p>A recent surge of research on commonsense knowledge has given the AI community new opportunities and challenges. Many studies focus on constructing commonsense knowledge representations from natural language data. However, how to learn such representations from large-scale text data is still an open question. This thesis addresses the problem through statistical script learning, which learns event representations from stereotypical event relationships using weak supervision. These event representations serve as an abundant source of commonsense knowledge to be applied in downstream language tasks. We propose three script learning models that generalize previous works with new insight. A feature-enriched model characterizes fine-grained and entity-based event properties to address specific semantics. A multi-relational model generalizes traditional script learning models which rely on one type of event relationship—co-occurrence—to a multi-relational model that considers typed event relationships, going beyond simple event similarities. A narrative graph model leverages a narrative graph to inform an event with a grounded situation to maintain a global consistency of event states. Also, pretrained language models such as BERT are used to further improve event semantics.</p><p>Our three script learning models do not rely on annotated datasets, as the cost of creating these at large scales is unreasonable. Based on weak supervision, we extract events from large collections of textual data. Although noisy, the learned event representations carry profound commonsense information, enhancing performance in downstream language tasks.</p> <p>We evaluate their performance with various intrinsic and extrinsic evaluations. In the intrinsic evaluations, although the three models are evaluated in terms of various aspects, the shared core task is Multiple Choice Narrative Cloze (MCNC), which measures the model’s ability to predict what happens next, out of five candidate events, in a given situation. This task facilitates fair comparisons between script learning models for commonsense inference. The three models were proposed in three consecutive years, from 2018 to 2020, each outperforming the previous year’s model as well as the competitors’ baselines. Our best model outperforms EventComp, a widely recognized baseline, by a large margin in MCNC: i.e., absolute accuracy improvements of 9.73% (53.86% → 63.59%). In the extrinsic evaluations, we use our models for implicit discourse sense classification (IDSC), a challenging task in which two argument spans are annotated with an implicit discourse sense; the task is to predict the sense type, which requires a deep understanding of common sense between discourse arguments. Moreover, in an additional work we touch on a more interesting group of tasks about psychological commonsense reasoning. Solving these requires reasoning about and understanding human mental states such as motivation, emotion, and desire. Our best model, an enhancement of the narrative graph model, combines the advantages of the above three works to address entity-based features, typed event relationships, and grounded context in one model. The model successfully captures the context in which events appear and interactions between characters’ mental states, outperforming previous works.</p> <div> <div> <div> <p>The main contributions of this thesis are as follows: (1) We identify the importance of entity-based features for representing commonsense knowledge with script learning. (2) We create one of the first, if not the first, script learning models that addresses the multi-relational nature between events. (3) We publicly release contextualized event representations (models) trained on large-scale newswire data. (4) We develop a script learning model that combines entity-based features, typed event relationships, and grounded context in one model, and show that it is a good fit for modeling psychological common sense.</p><p>To conclude, this thesis presents an in-depth exploration of statistical script learning, enhancing existing models with new insight. Our experimental results show that models informed with the new knowledge aspects significantly outperform previous works in both intrinsic and extrinsic evaluations. </p> </div> </div> </div> </div> </div> </div> </div>

DL-based ontologies have been widely used as knowledge infrastructures in knowledge management systems and on the Semantic Web. The development of efficient, sound and complete reasoning technologies has been a central topic in DL research. Recently, the paradigm shift from professional to novice users, and from standalone and static to inter-linked and dynamic applications raises new challenges: Can users build and evolve ontologies, both static and dynamic, with features provided by expressive DLs, while still enjoying e cient reasoning as in tractable DLs, without worrying too much about the quality (soundness and completeness) of results? To answer these challenges, this thesis investigates the problem of tractable and quality-guaranteed reasoning for ontologies in expressive DLs. The thesis develops syntactic approximation, a consequence-based reasoning procedure with worst-case PTime complexity, theoretically sound and empirically high-recall results, for ontologies constructed in DLs more expressive than any tractable DL. The thesis shows that a set of semantic completeness-guarantee conditions can be identifed to efficiently check if such a procedure is complete. Many ontologies tested in the thesis, including difficult ones for an off-the-shelf reasoner, satisfy such conditions. Furthermore, the thesis presents a stream reasoning mechanism to update reasoning results on dynamic ontologies without complete re-computation. Such a mechanism implements the Delete-and-Re-derive strategy with a truth maintenance system, and can help to reduce unnecessary over-deletion and re-derivation in stream reasoning and to improve its efficiency. As a whole, the thesis develops a worst-case tractable, guaranteed sound, conditionally complete and empirically high-recall reasoning solution for both static and dynamic ontologies in expressive DLs. Some techniques presented in the thesis can also be used to improve the performance and/or completeness of other existing reasoning solutions. The results can further be generalised and extended to support a wider range of knowledge representation formalisms, especially when a consequence-based algorithm is available.

The aim of this thesis is to develop theories and formal methods to endow a computing machinery with capabilities to identify, represent, reason and evaluate complex activities that are directed by an individual’s needs, goals, motives, preferences and environment, information which can be inconsistent and incomplete. Current methods for formalising and reasoning about human activity are typically limited to basic actions, e.g., walking, sitting, sleeping, etc., excluding elements of an activity. This research proposes a new formal activity-centric model that captures complex human activity based on a systemic activity structure that is understood as a purposeful, social, mediated, hierarchically organized and continuously developing interaction between people and word. This research has also resulted in a common-sense reasoning method based on argumentation, in order to provide defeasible explanations of the activity that an individual performs based on the activity-centric model of human activity. Reasoning about an activity is based on the novel notion of an argument under semantics-based inferences that is developed in this research, which allows the building of structured arguments and inferring consistent conclusions. Structured arguments are used for explaining complex activities in a bottom-up manner, by introducing the notion of fragments of activity. Based on these fragments, consistent argumentation based interpretations of activity can be generated, which adhere to the activity-centric model of complex human activity. For resembling the kind of deductive analysis that a clinician performs in the assessment of activities, two quantitative measurements for evaluating performance and capacity are introduced and formalized. By analysing these qualifiers using different argumentation semantics, information useful for different purposes can be generated. e.g., such as detecting risk in older adults for falling down, or more specific information about activity performance and activity completion. Both types of information can form the base for an intelligent machinery to provide tailored recommendation to an individual. The contributions were implemented in different proof-of-concept systems, designed for evaluating complex activities and improving individual’s health in daily life. These systems were empirically evaluated with the purpose of evaluating theories and methodologies with potential users. The results have the potential to be utilized in domains such as ambient assisted living, assistive technology, activity assessment and self-management systems for improving health.

Advances in Internet technologies have made it seemingly possible and easy to create digital collections, repositories and libraries. However, supporting diverse information uses that facilitate interaction beyond searching and browsing is in the early stages. Interactive digital libraries, or interactional digital libraries as we prefer to call them, are still evolving. This special issue tries to bring together work that is being done to incorporate interactivity in digital libraries.

published_or_final_version / abstract / Computer Science and Information Systems / Master / Master of Philosophy

This dissertation is about the design of a modular language for describing actions. The modular action description language, MAD, is based on the action language C+. In this new language, the possibility of "importing" a module allows us to describe actions by referring to descriptions of related actions introduced earlier, rather than by listing all effects and preconditions of every action explicitly. The use of modular action descriptions eliminates the need to reinvent theories of similar domains over and over again. Another advantage of this representation style is that it is similar to the way humans describe actions in terms of other actions. We first define the syntax of a fragment of MAD, called mini-MAD, and then extend it to the full version of MAD. The semantics of mini-MAD is defined by grounding action descriptions and translating them into C+. However, for the full version of MAD, it would be difficult to define grounding. Instead, we use a new approach to the semantics of variables in action descriptions, which is based on more complex logical machinery---first-order causal logic. Grounding is important as an implementation method, but we argue that it should be best avoided in the definition of the semantics of expressive action languages. We show that, in application to mini-MAD, the two semantics are equivalent. Furthermore, we prove that MAD action descriptions have some desirable, intuitively expected mathematical properties. We hope that MAD will make it possible to create a useful general-purpose library of standard action descriptions and will contribute in this way to solving the problem of generality in Artificial Intelligence. / text