Fact-Anchored Split Reasoning

Abstract

We propose Fact-Anchored Split Reasoning (FASR), a layered NLParchitecture that improves machine reasoning on semantically contradictory andcomplex sentences. Unlike conventional transformer models that process fullsentences as a single unit, FASR first detects a known fact anchor embedded inthe input — a clause whose truth is supported by the model's training knowledge— then splits the sentence into two parts at that anchor point. Each partindependently retrieves a concept set, after which a bridging layer identifies theshared conceptual medium connecting both parts, and a reasoning chain isconstructed toward a conclusion. Experiments on five contradictorycommonsense sentences demonstrate that FASR achieves a mean semanticalignment score of 0.6387 against expected reasoning links, compared to abaseline score of 0.4628 — a 38.0% relative improvement — winning all fivetest cases (5–0).