On improving knowledge modeling in knowledge-based systems

Abstract

Knowledge bases (KBs), such as Freebase and Wikidata, have received a lot of attention in the research community. The development of KB has benefited a number of interesting applications, including knowledge-based question answering (KB-QA) systems, recommendation systems and machine reading comprehension systems. One type of KBs, named Open Knowledge Bases (OKBs), is constructed by applying an open information extraction system to extract assertions from a large corpus. OKBs are noisy as they contain ambiguous facts. In the first part of the thesis, we put forward algorithms that resolve ambiguities in OKBs. After the discussion of OKBs, we present building of a domain specific KB. In particular, we build a legal KB from court judgments and demonstrate that the legal KB benefits a number of legal applications. Firstly, we propose an efficient algorithm to resolve ambiguities in OKBs. Previously a canonicalization approach based on hierarchical agglomerative clustering was proposed for entity resolution in OKBs. Despite the high effectiveness of the canonicalization approach, it cannot be applied to large OKBs due to its high computation cost. We propose Fast Assertion Canonicalization (FAC) to address the efficiency issue. FAC employs several optimization strategies, including pruning techniques to avoid unnecessary similarity computations and bounding techniques to efficiently approximate small similarities. We demonstrate that FAC achieves order-of-magnitude speedups over other approaches. Secondly, with the development of language models, many NLP problems can be solved with better accuracies, including the OKB canonicalization problem. We propose Multi-Level Canonicalization with Embeddings (MULCE), which solves the canonicalization using a two-step framework. In the first step, assertions are coarsely grouped into clusters according to the GloVe vectors of the subjects. The second step refines the clusters by modeling relation and object information with BERT embeddings. We show that combining multiple embedding approaches to solve the canonicalization problem is better than employing a single embedding method, and that MULCE outperforms other state-of-the-art approaches. Finally, we describe how to build a legal KB from court judgments. Judgments are an important source of legal knowledge, but the large number of judgments make one difficult to master all the knowledge. We study the problems of machine-assisted extraction and modeling of legal knowledge from legal texts, leveraging the domain knowledge provided by human experts. Through examples, we show how our knowledge models can lead to the development of a number of essential legal applications that facilitate legal studies and research. In particular, we develop a prediction model for illegal drug trafficking sentencing. We demonstrate how the extracted legal knowledge and the sentencing prediction model help address a number of interesting issues in legal information processing, which include (1) extraction and cleaning of legal data, (2) discovery of model drifts in legal rules, (3) identification of critical features in legal judgments, (4) recommendation of similar judgments, (5) fairness in machine predictions, and (6) explainability of machine predictions.