AutoRegressive Knowledge Base Completion

Yannick Brunink, Jacopo Urbani, Michael Cochez

Keywords: Knowledge Graphs, Probabilistic Reasoning

Abstract:

Despite their large sizes, many Knowledge Graphs (KGs) remain highly incomplete. This problem has motivated numerous approaches to $\textit{complete}$ the KGs by embedding them in a latent space to find the missing links. Although these methods show promising performance, a general limitation is that the scores given to possible links are uncalibrated and cannot be interpreted across different queries. Hence, we say they are $\textit{local}$ as they relate to a specific context. This limitation makes it non-trivial to deduce the truth value of the links and to answer complex queries. Another limitation is that their learning depends on negative sampling, which is challenging due to the Open World Assumption (OWA).

To solve this problem, we propose a novel auto-regressive generative model that learns a joint distribution of the entities and relations of the KG without resorting to negative sampling. This distribution can be used to infer the probability that a link is sampled from the KG, which allows us to return a $\textit{global}$ score that is interpretable in different contexts. Moreover, our method has the additional advantage that it offers probabilistic semantics for complex reasoning and knowledge base completion, achieving state-of-the-art performance on link prediction with consistent scores across the entire KG.

Primary Area: learning on graphs and other geometries & topologies

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11048