The research introduces EAR, an Entity-Aware Partitioning approach designed for multiple-choice question answering within retrieval-augmented generation systems. The core challenge addressed is the selection of appropriate chunks for retrieval, with fixed-size chunks often failing to effectively connect to the question's context. EAR operates by extracting normalized surface anchors from questions, answer options, and the corpus. These anchors are then used to retrieve local windows around matching corpus segments. A larger parent passage can be attached through an extractive summary if necessary.
Evaluations were conducted using Mistral, Gemma, and DeepSeek models on a subset of 153 questions derived from a cleaned version of the Massive Multitask Language Understanding (MMLU) benchmark. The evaluation protocol utilized top-k retrieval sweeps of 3 and 8. Results demonstrated a reduction in retrieved words of 37.5-40.2% compared to standard chunking methods. Associated accuracy changes were observed, with gains of +5.2 to +5.9 points at top-k=3 and top-k=8, respectively, alongside corresponding decreases in accuracy for some models.
The methodology focuses on creating compact and inspectable retrieval units. The experiment observed statistically insignificant differences in accuracy between the entity-window approach and standard chunking across the models tested. The anchor extraction rule-based system is designed to be domain-specific and requires separate validation before transfer to other domains.
The research highlights the potential for optimizing RAG systems by controlling the granularity of retrieved information. The EAR approach offers a method for reducing the volume of data processed during retrieval, which can improve inference speed and reduce computational costs. Source: https://arxiv.org/abs/2609.12268
