Rating is a elementary and fashionable drawback in search. Nevertheless, present rating algorithms normally limit the granularity of rating to full passages or require a selected dense index for every desired stage of granularity. Such lack of flexibility in granularity negatively impacts many purposes that may profit from extra granular rating, reminiscent of sentence-level rating for open-domain question-answering, or proposition-level rating for attribution. On this work, we introduce the thought of any-granularity rating which leverages multi-vector approaches to rank at various ranges of granularity whereas sustaining encoding at a single (coarser) stage of granularity. We suggest a multi-granular contrastive loss for coaching multi-vector approaches, and validate its utility with each sentences and propositions as rating models. Lastly, we show the applying of proposition-level rating to post-hoc quotation addition in retrieval-augmented era, surpassing the efficiency of prompt-driven quotation era.
![](https://mlr.cdn-apple.com/media/Fig2_1f7421a515.png)