Investigating the Impact of Context on Zero-Shot Entity Resolution: Applications in Chronic Lower Back Pain

In order to assist chronic pain researchers, knowledge graphs have been proposed to help model the dense, complex, and trans-disciplinary nature of chronic pain, facilitating information processing, clinical diagnosis, and accelerating research. At the heart of knowledge graph construction, entity resolution is a critical step in normalizing data, promoting a robust representation of the information embedded in the literature by combining and de-duplicating identical and similar terms. To improve entity resolution, we attempt to augment pre-existing methods that leverage Large Language Models by integrating the sentence from which the entity is extracted. In our experiments, we demonstrate that integrating context into zero-shot methods is a non-trivial task, illustrating how including context may not always be required. As such, we evaluate the impact of including context and offer methods to strategically inject additional information when required, achieving an F1-Score of 0.66 when applied to our entity resolution dataset.