When a search engine retrieves a document, the most useful description is often not written by the document's author. It comes from other people: the words on hyperlinks that point to a page (anchor text), or the sentences in other papers that cite it. Recent work shows these third-party descriptions still help strongly with modern neural retrievers. The catch is that they are distributed very unevenly. A few popular documents have thousands of incoming links, most have one or two, and new documents have none. This note proposes a retriever that keeps third-party text as its own document section (a "field") and decides how much to trust that field by looking at two things together. The first is the query, as earlier adaptive retrievers already do. The second is the link state of each candidate document: how many referrals it has, how varied they are, and whether they agree with the document's own content.
The motivating failure is simple. Suppose a retriever learns, from the query alone, that "this looks like a known-item query, so put weight on the citation field." It then lowers the score of every relevant document whose citation field is empty, which in practice means penalizing exactly the new and long-tail documents. We claim that adaptive weighting over fields only becomes safe once it is conditioned on the document side as well, and that this conditioning should matter most for the documents that are poorly linked. The proposal is built so that this claim can fail: a cheap fix, renormalizing weights over whichever fields are present, may turn out to be enough, and the experiments are designed to show whether it is.
Where the prior work leaves the gap
Three lines of work meet here. The first, multi-field retrieval, treats a document as separate sections rather than one string. Early web language-model systems built one model per field and mixed them with fixed or per-task weights (Combining document representations for known-item search; Language Models for Searching in Web Corpora). Neural rankers later took fields as native input (Neural Ranking Models with Multiple Document Fields; GLOW). The most direct precursor is Multi-Field Adaptive Retrieval (MFAR). MFAR indexes every field with both BM25 and a dense encoder, and a small network reads the query embedding and predicts a weight for each field-and-scorer pair. Removing that query conditioning cost 22.6% Hit@1 on the STaRK benchmark. MFAR supplies our method, but its fields are all internal metadata, and its weights depend only on the query, so every candidate document is weighted the same way.
The second line is anchor text. Classical TREC systems showed that anchor text dominates homepage and named-page finding but adds little depth for topical queries (TREC-10 Web Track Experiments at MSRA; Microsoft Research Asia at the Web Track of TREC 2003). Referral Augmentation for Zero-Shot Information Retrieval (RAR) revived the idea for neural retrieval. It appends citing passages to the document for BM25, or mean-pools their embeddings for dense models, and more than doubles Recall@10 on ACL paper retrieval. RAR motivates this proposal, because it shows the signal is large. It also constrains it, because it beats generated expansions (DocT5Query, Query2Doc) by wide margins, so synthetic referrals are not a free substitute. RAR applies the same flat merge to every query and every document, and the authors note that it has no answer for documents without links.
The third line is query-dependent model selection. Rule-based switching between anchor-based and content-based pipelines by query type (University of Glasgow at TREC 2004; Finding "Abstract Fields" of Web Pages and Query Specific Retrieval) gave the best mixed-query result at TREC 2004. Learned selectors improved on fixed blends (Learning to select rankers; Investigating Retrieval Method Selection with Axiomatic Features). All of these choose per query. None conditions the choice on properties of the individual candidate document, such as how well linked it is.
Core Claim
Let $s_f(q,d)$ be the score of document $d$ for query $q$ under field-and-scorer pair $f$, for example "referral field, dense". Let $z_d$ be a small vector of link statistics for $d$: log referral count, the spread of referral embeddings, the cosine between the mean referral embedding and the document's own content embedding, and the age of the newest referral. We score
$$\text{score}(q,d)=\sum_{f\in F(d)} \frac{\exp\big(g_f(q)+h_f(q,z_d)\big)}{\sum_{f'\in F(d)}\exp\big(g_{f'}(q)+h_{f'}(q,z_d)\big)}\, s_f(q,d).$$
Here $F(d)$ is the set of fields that are non-empty for $d$. The term $g_f(q)$ is MFAR's query-only weight. The term $h_f(q,z_d)$ is a small bilinear correction that lets the trust placed in a field depend on that document's link state.
The claim has two parts. First, keeping referrals as a separate field with this two-sided weighting beats RAR's flat merge and beats MFAR with a referral field added under query-only weights. Second, the gain comes mainly from poorly linked and very heavily linked documents, where a single per-query weight is wrong in opposite directions. Renormalizing over $F(d)$ alone, that is, setting $h=0$, is the cheap fix that must be beaten.
Research Program
We build the system on MFAR's recipe: separate BM25 and Contriever indices per field, top-$k$ shortlists from each index, and contrastive training of the weighting network. The one change to the index is a referral field. On ACL we use RAR's citation contexts. On DBPedia, Natural Questions and HotpotQA we use Wikipedia inbound-link passages, following RAR. On STaRK-MAG we use the titles and abstracts of citing papers, taken from its citation relations. On TREC 2004 .GOV we use inbound anchor text.
Heavily linked documents need special handling, because mean-pooling hundreds of referrals washes out their meaning. For these documents we keep a capped set of referral vectors, selected by clustering, and score the referral field by the best match among them. This borrows RAR's "shortest path" option, but bounds index growth.
Training uses STaRK-MAG's training queries and the TREC 2004 training topics. ACL, DBPedia, NQ and HotpotQA are evaluated zero-shot, which tests whether the link-state correction transfers across corpora. Every result is also reported by in-degree bucket: zero, one to three, four to fifty, and above fifty referrals.
Baselines
The naive reference point has two parts. One is content-only retrieval with BM25 and Contriever over the flattened document with no referrals. The other is content-only MFAR. Together they tell us how much any use of links adds.
The strongest existing referral method is RAR, run as its paper specifies: concatenation for BM25 and mean pooling for dense encoders, over the same referral text and the same Contriever backbone. On ACL, RAR reports Recall@10 of 0.505 with BM25 and 0.355 with SimCSE (Referral Augmentation). We will rerun it and treat our rerun as the bar. Beating it means a statistically significant gain in Recall@10 on ACL, with no loss in the zero-referral bucket relative to content-only retrieval.
The strongest adaptive method is MFAR with the referral field added and query-only weights, trained with the same data, number of steps and hyperparameter search budget as our model. The file gives MFAR's relative ablation drop but no absolute STaRK numbers: no reported result found on Lacuna for MFAR's absolute Hit@1 on STaRK-MAG. We therefore rerun it.
On TREC 2004 the reference is the rule-based THUIRmix043 run, which reports MAP 0.5245 (Finding "Abstract Fields"). This comparison is indicative only, because our system uses different scorers.
Three ablations isolate the key ingredient. The first is renormalization only ($h=0$). The second is $h$ conditioned on referral count alone. The third is the full model with the content-agreement feature, which tests whether the network uses agreement to discount noisy or spam-like referrals.
Why not just…?
Why not just append referrals to the document, as RAR does? Flat merging works well on average and is cheap, so it is a real competitor. It has two problems. It forces citation text into every query's match, which earlier TREC work found unhelpful for topical queries (TREC-10 Web Track Experiments at MSRA). And pooling over hundreds of referrals blurs heavily linked documents. If our model fails to beat RAR in the topic-distillation and over-fifty-referral buckets, the separate-field design is not justified.
Why not just add a referral field to MFAR and let its query-only weights handle it? This is the obvious combination, and it is our main ablation. A single per-query weight cannot simultaneously trust referrals for a hub document and ignore an empty field for a new one. We predict this model loses recall in the zero-referral bucket relative to content-only MFAR. If it does not, the document-side correction is unnecessary.
Why not just renormalize weights over the fields a document actually has? This is the most serious threat, because it already removes the penalty for empty fields. What it cannot do is down-weight a field that is present but unreliable, such as two vague referrals or off-topic link text. The $h=0$ ablation settles this. If renormalization matches the full model in every bucket, we will report that the simple fix suffices and narrow the contribution to the referral-field design.
Why not just fill empty referral fields with generated expansions such as DocT5Query? RAR found human referrals far ahead of generated ones on ACL, by 37 points absolute over DocT5Query. Generated text still costs nothing to add, so we include DocT5Query fill-in as an extra baseline for the zero-referral bucket only.
Decisive Experiments
The central test is a time split. On ACL and STaRK-MAG, we build referral fields only from citations published before a cutoff, and evaluate queries whose targets are papers published after it, which therefore have no referrals. We compare these against queries targeting older papers. The prediction is that query-only weighting with a referral field falls below content-only retrieval on new targets, while the link-aware model stays at or above it and keeps RAR-level gains on old targets.
A second test uses the TREC 2004 mixed query set. Plotting the learned referral weight against query type should show weights near zero for topic distillation and high for homepage finding, without any query-type labels in training.
A third test injects off-topic anchor text into a random 10% of documents. The agreement feature should drive those documents' referral weight down, which RAR and MFAR cannot do.
Metrics are Hit@1, Recall@20 and MRR on STaRK; Recall@10 on ACL and the RAR corpora; and MAP by query type on TREC 2004. Each is broken down by in-degree bucket, with paired significance tests. We also report index size and query latency relative to MFAR.
References
- Combining document representations for known-item search
- Language Models for Searching in Web Corpora
- Neural Ranking Models with Multiple Document Fields
- GLOW : Global Weighted Self-Attention Network for Web Search
- Multi-Field Adaptive Retrieval
- TREC-10 Web Track Experiments at MSRA
- Microsoft Research Asia at the Web Track of TREC 2003
- Referral Augmentation for Zero-Shot Information Retrieval
- University of Glasgow at TREC 2004: Experiments in Web, Robust, and Terabyte Tracks with Terrier
- Finding "Abstract Fields" of Web Pages and Query Specific Retrieval - THUIR at TREC 2004 Web Track
- Learning to select rankers
- Investigating Retrieval Method Selection with Axiomatic Features
- STaRK: Benchmarking LLM Retrieval on Textual and Relational Knowledge Bases