Large Language Models (LLMs) have demonstrated strong performance on many Natural LanguageProcessing (NLP) tasks. Following widespread adoption, research focused on improving their capabilities and finding solutions for drawbacks. One notable drawback is hallucination: the phenomenonwhere a LLM generates false information with high confidence when asked about something that wasnot present in its training data. Retrieval Augmented Generation (RAG) addresses this by groundingLLMs with information from an external knowledge source. In traditional RAG systems, embedding models use pure semantic understanding to generate a vector for a similarity search against aknowledge base. However, this overlooks the emotional and temporal nuances that are required foraffective computing. In this paper, we propose the Affective Link Score (ALS), a computation that represents the strengthof a connection between events as a function of semantic, emotional, and temporal components.Wefurther introduce Deep Recall, a framework for Affective RAG that uses ALS as a re-ranker tosurface emotional content in a context pool. Our empirical results demonstrate that 1) an ALS modelcan achieve 100% recall in causal-affective link prediction and 2) Deep Recall surfaces 22% moreemotional events than standard RAG. This suggests that incorporating affective and temporal signalsinto the information retrieval process allows for a more robust reconstruction of narrative continuitythan semantics alone.
Joseph Evans (Wed,) studied this question.