Towards Unified Representations of Knowledge Graph and Expert Rules for Machine Learning and Reasoning

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 240–253 November 20–23, 2022. ©2022 Association for Computational Linguistics 240 Towards Unified Representations of Knowledge Graph and Expert Rules for Machine Learning and Reasoning Zhepei Wei1,3, Yue Wang2, Jinnan Li3,5, Zhining Liu1,3, Erxin Yu1,3, Yuan Tian1,3, Xin Wang1,3, Yi Chang1,3,4∗ 1School of Artificial Intelligence, Jilin University 2School of Information and Library Science, University of North Carolina at Chapel Hill 3Key Laboratory of Symbolic Computation and Knowledge Engineering, Jilin University 4International Center of Future Science, Jilin University 5College of Computer Science and Technology, Jilin University {weizp19, jnli21, znliu19, yuex19}@mails.jlu.edu.cn, wangyue@email.unc.edu, {yuantian, xinwang, yichang}@jlu.edu.cn Abstract With a knowledge graph and a set of if- then rules, can we reason about the conclu- sions given a set of observations? In this work, we formalize this question as the cog- nitive inference problem, and introduce the Cognitive Knowledge Graph (CogKG) that unifies two representations of heterogeneous symbolic knowledge: expert rules and rela- tional facts. We propose a general frame- work in which the unified knowledge repre- sentations can perform both learning and rea- soning. Specifically, we implement the above framework in two settings, depending on the availability of labeled data. When no labeled data are available for training, the framework can directly utilize symbolic knowledge as the decision basis and perform reasoning. When labeled data become available, the framework casts symbolic knowledge as a trainable neu- ral architecture and optimizes the connection weights among neurons through gradient de- scent. Empirical study on two clinical diag- nosis benchmarks demonstrates the superior- ity of the proposed method over time-tested knowledge-driven and data-driven methods, showing the great potential of the proposed method in unifying heterogeneous symbolic knowledge, i.e., expert rules and relational facts, as the substrate of machine learning and reasoning models. The source code and data are released online1. 1 Introduction Symbolic reasoning methods such as rule-based expert systems (Buchanan and Shortliffe, 1984) are reliable and interpretable in solving complex inference problems in specialized domains, but are also difficult to generalize because eliciting a comprehensive set of rules from human experts is costly and time-consuming. Recently, knowledge ∗Corresponding Author 1http://github.com/jinnanli/CogKG Number of training examples Data-driven Knowledge-driven Majority Guess Hybrid Performance Figure 1: Illustration of impacts of training exam- ples on different reasoning paradigms (with fixed prior knowledge). Note that the curves start w/o pre-training. graph (KG) as a flexible representation of sym- bolic knowledge has been proven successful for knowledge-based reasoning (Bordes et al., 2013) by utilizing the distributed representations to gen- eralize from known facts to unseen yet probably true facts, which is also known as the knowledge graph completion task (Lin et al., 2015). How- ever, such models can only represent and reason about multi-relational data in the form of (subject, predicate, object) triples (Liben-Nowell and Klein- berg, 2003), not conditional if-then rules. There- fore, current knowledge graph embedding models are not suited to …

Content truncated to 500 words