Recent advancements in Natural Language Processing have improved extracting structured semantic representations from unstructured text with Frame Semantic Role Labeling (FSRL). This paper introduces the RCIF approach for frame detection, using Retrieval-Augmented Generation models. RCIF does not require an explicit target span and consists of three stages: generating frame embeddings, retrieving candidate frames, and identifying the most suitable frames. Experiments across various configurations show that the retrieval component reduces complexity and improves performance on FrameNet 1.5 and 1.7. Additionally, the method enhances generalization for translating natural language questions into SPARQL queries.
https://arxiv.org/abs/2502.12210