Advancing Abductive Reasoning in Knowledge Graphs through Complex Logical Hypothesis Generation

TL;DR

The paper addresses the problem of abductive reasoning over knowledge graphs by introducing complex logical hypothesis generation as a formal task. It proposes a generation-based pipeline (RLF-KG) that leverages sampling, supervised transformer training, and reinforcement learning with KG feedback to produce explanations that align with observations, even on unseen data. Empirical results on FB15k-237, WN18RR, and DBpedia50 show state-of-the-art performance in terms of Jaccard similarity and competitive Smatch scores, with RLF-KG consistently improving explanations. The work demonstrates the practical potential of integrating KG structure into abductive reasoning and highlights trade-offs in reward design and computational efficiency, pointing to future work on dynamic knowledge graphs and broader domains.

Abstract

Abductive reasoning is the process of making educated guesses to provide explanations for observations. Although many applications require the use of knowledge for explanations, the utilization of abductive reasoning in conjunction with structured knowledge, such as a knowledge graph, remains largely unexplored. To fill this gap, this paper introduces the task of complex logical hypothesis generation, as an initial step towards abductive logical reasoning with KG. In this task, we aim to generate a complex logical hypothesis so that it can explain a set of observations. We find that the supervised trained generative model can generate logical hypotheses that are structurally closer to the reference hypothesis. However, when generalized to unseen observations, this training objective does not guarantee better hypothesis generation. To address this, we introduce the Reinforcement Learning from Knowledge Graph (RLF-KG) method, which minimizes differences between observations and conclusions drawn from generated hypotheses according to the KG. Experiments show that, with RLF-KG's assistance, the generated hypotheses provide better explanations, and achieve state-of-the-art results on three widely used KGs.

Figures (6)

• Figure 1: This figure shows some examples of observations and inferred logical hypotheses, expressed with discrepancies in interpretations.
• Figure 2: The figure demonstrates the tokenization process for hypotheses. We uniformly consider logical operations, relations, and entities as individual tokens, establishing a correspondence between the hypotheses and a sequence of tokens. For a more detailed explanation, please refer to the Appendix \ref{['sec:algo-sampling']}.
• Figure 3: The first two steps of training a hypothesis generation model: In Step 1, we randomly sample logical hypotheses with diverse patterns and perform graph searches on the training graphs to obtain observations. These observations are then tokenized. In Step 2, a conditional generation model is trained to generate hypotheses based on given tokenized observations.
• Figure 4: In Step 3, we employ RLF-KG to encourage the model to generate hypotheses that align more closely with the given observations from the knowledge graph. RLF-KG helps improve the consistency between the generated hypotheses and the observed evidence in the knowledge graph.
• Figure 5: Our task involves considering thirteen distinct types of logical hypotheses. Each hypothesis type corresponds to a specific type of query graph, which is utilized during the sampling process. By associating each hypothesis type with a corresponding query graph, we ensure that a diverse range of hypotheses is sampled.