What Makes an Ideal Quote? Recommending "Unexpected yet Rational" Quotations via Novelty

A research team from Fudan University and Alibaba Group has published a novel framework, NovelQR, designed to solve a core problem in AI-assisted writing: finding the perfect quote. Moving beyond simple topical relevance, the system is built on the key insight from user studies that people prefer quotations that are 'unexpected yet rational' within a given context. The paper, published on arXiv, formalizes this objective by drawing on defamiliarization theory, aiming to recommend quotes that are contextually novel yet semantically coherent, a balance existing models struggle to achieve.

The NovelQR framework operates in two stages. First, a generative label agent interprets each quotation and its context into multi-dimensional 'deep-meaning' labels, enabling a more sophisticated, label-enhanced retrieval process. Second, a token-level novelty estimator reranks the candidate quotes, specifically designed to mitigate auto-regressive continuation bias that can lead to predictable suggestions. In experiments across bilingual datasets spanning diverse real-world domains, NovelQR's recommendations were judged by humans as significantly more appropriate, novel, and engaging than baseline systems, while also matching or surpassing existing methods in the technical task of novelty estimation itself. This represents a shift from surface-level matching to understanding the deeper semantic and aesthetic properties that make a quote memorable and effective.