Referral Augmentation for Zero-Shot Information Retrieval

Anonymous

Abstract:

We propose Referral-Augmented Retrieval (RAR), a simple technique that concatenates document indices with referrals, i.e. text from other documents that cite or link to the given document, to provide significant performance gains for zero-shot information retrieval. The key insight behind our method extends an intuition from classical web retrieval: referrals provide a more complete, multi-view representation of a document, much like incoming page links in PageRank provide a comprehensive idea of a webpage's importance. We formulate this classically-rooted intuition as a general augmentation and find that it empirically works across various new domains and retrieval methods, outperforming modern generative text expansion techniques such as DocT5Query and Query2Doc — a 37% and 21% absolute improvement on ACL paper retrieval Recall@10, respectively, while also eliminating expensive model training and inference. We also analyze different methods for multi-referral aggregation and show that RAR enables up-to-date information retrieval without re-training.
We believe RAR can help revive and re-contextualize this classic information retrieval intuition in the age of neural retrieval, unlocking new retrieval gains by combining untapped corpus structure with the semantic advantages of modern pretrained transformers.

Paper Type: long

Research Area: Information Retrieval and Text Mining

Languages Studied: English

Consent To Share Submission Details: On behalf of all authors, we agree to the terms above to share our submission details.