Integrated knowledge graphs form the foundation of numerous data-driven applications, including search engines, conversational agents, and e-commerce solutions. Declarative link discovery frameworks utilize link specifications to define the conditions necessary for establishing a link between knowledge graphs’ resources. Despite domain expertise, defining such link specifications remains challenging due to their intricate syntax, threshold tuning, and the need to precisely express complex linking logic. To address this challenge, we propose NL2LS, a novel language-driven approach that leverages large language models to automatically translate natural language (NL) into link specifications (LSs), enabling domain experts and practitioners to express correct and complex linking rules more effectively. NL2LS employs three distinct training paradigms to handle the complexity of link specifications: zero-shot learning, one-shot learning and supervised fine-tuning. We evaluated NL2LS using different large language model architectures in comparison with a rule-based baseline model on different multi-lingual datasets. Our evaluation using BLEU, METEOR, ChrF++, and TER metrics demonstrates that NL2LS effectively translates natural language into link specifications, lowering the technical barrier and assisting users in specifying link rules more intuitively.

An increasing number of heterogeneous datasets abiding by the Linked Data paradigm is published everyday. Discovering links between these datasets is thus central to achieving the vision behind the Data Web. Declarative Link Discovery (LD) frameworks rely on complex Link Specification (LS) to express the conditions under which two resources should be linked. Complex LS combine similarity measures with thresholds to determine whether a given predicate holds between two resources. State of the art LD frameworks rely mostly on string-based similarity measures such as Levenshtein and Jaccard. However, string-based similarity measures often fail to catch the similarity of resources with phonetically similar property values when these property values are represented using different string representation (e.g., names and street labels). In this paper, we evaluate the impact of using phonetics-based similarities in the process of LD.

Existing technologies employ different machine learning approachesto predict disasters from historical environmental data. However,for short-term disasters (e.g., earthquakes), historical data alonehas a limited prediction capability. In this work, we consider so-cial media as a supplementary source of knowledge in additionto historical environmental data. Further, we build a joint modelthat learns from disaster-related tweets and environmental data toimprove prediction. We propose the combination of semantically-enriched word embedding to represent entities in tweets with theirsemantics representations computed with the traditionalword2vec.Our experiments show that our proposed approach outperformsthe accuracy of state-of-the-art models in disaster prediction