Open-World Knowledge Graph Completion with Linguistic Noise

Reasoning over Knowledge Bases is a wellestablished task in Artificial Intelligence, but when it comes to perform reasoning in an open-world scenario, in which entities involved in logical statements may be unknown to the reasoner, exact reasoning does not apply. In this work we deal with the case of Open-World Knowledge Graph Completion (KGC): a set of tasks concerning the assessment of the validity of new triples with unseen entities given a pre-existing Knowledge Graph. The use of Large Language Models to represent triples in a vector space is widespread, but most approaches either focus on injecting knowledge in the Language Model at training stage or adopt the LLMs to perform the tasks altogether. The main drawback of such approaches is the wrong prioritization of the LLM's internal knowledge over the information present in the KG, which is also less subject to various forms of linguistic noise. In this paper we study a method to counteract this issue by adopting the Logic Tensor Network framework to learn predicate groundings, defined as standalone neural models, and refine entity embeddings, previously obtained through an LLM. We provide a preliminary investigation of two main properties of such a framework: its potential in performing specifically the Triple Validation task, and its robustness to linguistic noise injected in the entity embeddings. We compare our LTN-based method with a Deep Learning fully supervised approach and an LLM-based approach, finding promising results for further development.