Scoring Bridge Concepts by Conditional Gain for Zero-Overlap Retrieval

Many relevant documents share no words with the question that should find them. On WikiQA, 20.3% of correct answer sentences have zero content words in common with the question (WikiQA). On BRIGHT, where a coding question is answered by documentation of an algorithm it never names, a strong embedding model falls from 59.0 on a standard benchmark to 18.3 (BRIGHT). Such a document becomes findable only through an intermediate concept, such as a theorem, a mechanism, or a data structure. That intermediate concept, taken alone, looks no more similar to the question than anything else in the corpus. Relational learning met the same problem long ago. A greedy learner discards a "bridge" variable because adding it gives no immediate gain, even though it is what makes the next step informative (Relational information gain).

We propose to treat vocabulary mismatch as this bridge problem and to score candidate bridges with a dependency-based notion of information. A bridge is worth expanding if conditioning on it makes the documents it surfaces look relevant to the query more than it makes arbitrary documents look relevant, and if those documents are not already in the candidate pool. Put simply, a language model proposes intermediate concepts. We ground each concept in a real corpus passage, keep only the concepts whose passage creates a query-specific dependency between the query and new documents, and spend the expensive relevance judge only on those. The aim is to get zero-overlap relevant documents into the first-stage candidate pool at a fixed, small judging budget. The method should keep working when the language model has not memorized the answer, which is the regime where current LLM query expansion is known to fail.

Core Claim

The claim is testable. At a matched number of relevance-judge calls, greedy selection of bridges by an optimism-corrected conditional gain should raise first-stage Recall@100 on BRIGHT over two alternatives: uniform expansion with the same bridges, and chain-of-thought query rewriting. It should also stay at or above BM25 on the "unmatched" FEVER claims, where LLM expansion collapses. Recall@100 matters here because a second-stage reranker, however strong, cannot rescue a document the first stage never retrieved. Retro* states this limitation explicitly (Retro*).

The idea that "information is dependency" comes from two older lines of work. The first is Greiner and Genesereth's partial-interpretation novelty (What's New?). A statement is "about" a concept if, under some assignment of context, it narrows that concept's possible states. Their platypus example tells a system about hair without mentioning hair, which is exactly how a zero-overlap document is relevant. The second is the rule-interest test in A New Approach to Rule Interest Measures: a composite rule is interesting only if it breaks conditional independence, so that joint evidence says more than the pieces separately. Both works are purely formal. The first is intractable over possible worlds, and the second is stated for association rules over discrete tables. We borrow the criterion, not the machinery. The context becomes a retrieved corpus passage, and the "narrowing" becomes a shift in a learned relevance judge that is contrasted against a control set.

The Scoring Rule

Let $q$ be the query, $P$ the current candidate pool, and $b$ a candidate bridge. A bridge is represented by a corpus passage retrieved for an LLM-proposed concept name, and never by LLM-generated text. Let $S_b$ be the top documents that $b$ retrieves with BM25 or a dense retriever. Let $J(q, d \mid c)$ be the probability, under a fixed cross-encoder judge, that document $d$ answers $q$ when context passage $c$ is shown alongside it. We score

$$ G(b \mid P) \;=\; \frac{1}{m}\sum_{d \in \mathrm{top}_m(S_b \setminus P)} J(q,d \mid b) \;-\; \frac{1}{m}\sum_{d' \in R_m} J(q,d' \mid b), $$

where $R_m$ is a set of $m$ control documents drawn from the corpus and length-matched to $S_b$.

The first term asks whether conditioning on $b$ makes new documents look relevant. Restricting it to documents outside $P$ enforces novelty relative to what is already known, which follows Greiner and Genesereth. The second term is the conditional-independence control. It targets bridges that are "persuasive" in general, making the judge rate everything as relevant, rather than bridges that tie the query to specific documents. It also addresses the known weakness of relational information gain: taking an optimistic maximum over subsets overrates noisy variables (Relational information gain). We further replace the sample means with a lower confidence bound over the $m$ samples.

Search proceeds greedily. At each step we add the pool of the highest-scoring bridge and recompute $G$ against the enlarged pool. We stop when no bridge exceeds a threshold $\tau$ or when the budget of judge calls is spent. Each bridge costs $2m$ judge calls, so the budget is explicit and can be matched across methods.

Relation to Existing Work

Several systems already bridge vocabulary gaps with language models, and they motivate the design. HyDE-style hypothetical documents, as used in HerO, write text in the corpus's vocabulary. InteR alternates LLM expansion and retrieval and reaches 50.0 MAP on TREC DL'19, but it degrades beyond two rounds (Synergistic Interplay). Think-Then-Rewrite trains a rewriter with a contrastive mutual-information reward between the rewritten query and the gold document embedding (Think Then Rewrite). RaFe trains rewriters from reranker feedback (RaFe). All of these produce and merge expansions without asking whether a particular expansion adds a query-specific dependency beyond what is already retrieved. Think-Then-Rewrite's objective relates only two quantities, query and document, and it needs training data. Our rule is a test-time selection criterion that contrasts conditioned relevance against a control set and against the current pool, and it requires no training.

One paper constrains the design more than any other. Hypothetical Documents or Knowledge Leakage? shows that LLM expansions succeed mainly when the model already memorized evidence entailed by the gold passage. On unmatched FEVER claims, expansion Recall@5 is 23.8, below BM25 at 31.0. This finding is why our bridges must be corpus passages. The judge is never shown LLM-authored text, so a bridge earns credit only for documents the corpus actually contains.

NSnet's lookup into a knowledge base of tuples also bridges gaps with external facts, but it has no criterion for whether a looked-up fact adds a necessary constraint (Bridging Knowledge Gaps). Its knowledge-base ablation costs only 2.1% on SciTail, which is a caution that retrieved background facts often turn out to be trivia. Generative retrieval (MINDER, DSI analysis) memorizes query-document associations in model weights, which does not help at inference time with bridges never seen in training. We therefore do not compete with it.

Research Program

We will build the pipeline from off-the-shelf parts using the BRIGHT codebase (xlang-ai/BRIGHT). An open instruction-tuned LLM proposes $K$ concept names per query; $K=16$ is the default. Each name is sent to BM25 and to a dense retriever, and the top passage becomes $b$. A single fixed cross-encoder serves as the judge $J$ in every method. The initial pool is the top 100 from BM25 plus the dense retriever.

The analysis has three parts. First, we measure recall against judge calls for each method. Second, we split BRIGHT and FEVER queries by whether the gold evidence is entailed by the LLM's closed-book answer, reusing the matched/unmatched protocol of Yoon et al., to test robustness to leakage. Third, we analyze accepted bridges on MISCIPLUS, where an undermining limitation passage should enter the pool only through a methodology bridge such as "in-vitro study" or "animal model." This third part tests whether the conditional term picks up the claim-methodology dependency that surface retrieval misses (Grounding Fallacies).

Baselines

The naive reference is first-stage BM25 with no expansion. The strongest plain lexical baseline also adds classical pseudo-relevance feedback. BRIGHT reports a BM25 average nDCG@10 of 14.5 (BRIGHT). On FEVER, BM25 Recall@5 is 31.0 (Knowledge Leakage). Any method that falls below these numbers on unmatched queries fails.

The strongest first-stage expansion competitor is LLM chain-of-thought query pre-reasoning. BRIGHT reports gains of up to 12.2 points from it. We also run HyDE-style expansion and a two-round InteR loop using the same LLM, the same retrievers, and the same judge. InteR's 50.0 MAP on DL'19 is the reported ceiling for iterative expansion in web search (Synergistic Interplay). The best standard retriever on BRIGHT, reported at 22.5 nDCG@10, is included as a dense-only reference.

For end-to-end numbers, the strongest system is Retro. It reports 36.6 nDCG@10 at 7B and 40.6 at 32B with 16 reasoning samples (Retro*). We do not try to beat Retro as a reranker. Instead we feed it our pool and the baseline pools, and we count a win if our pool raises its final nDCG@10 at equal reranking cost. We found no reported first-stage Recall@100 on BRIGHT on Lacuna, so we compute that number ourselves for every method.

Four ablations isolate the key ingredient. The first uses the same $K$ bridges, but each receives an equal share of the judge budget and all their pools are merged; this isolates selection. The second is a marginal variant scored by $J(q,d \mid \varnothing)$ over new documents; this isolates conditioning on the bridge. The third drops the control term; this isolates the conditional-independence contrast. The fourth passes the LLM's own concept description as context instead of a corpus passage; this isolates grounding and should reproduce the leakage failure. Every method gets the same judge-call budgets, from 64 to 1024 calls per query, and the same hyperparameter search on a held-out BRIGHT split. That search covers $\tau$ and $m$ for us and the number of expansions for the baselines.

Why not just…?

Why not just rerank a much larger first-stage pool? Retrieving the top 1000 and running the judge over all of them is the most serious threat. If the judge is cheap enough, a large pool can recover zero-overlap documents without any bridges. The recall-versus-calls curve settles this directly. The large-pool baseline sits on the same axis, and if it matches our recall at equal calls, the selection criterion adds nothing and we will report that. We expect it to lose on BRIGHT and MISCIPLUS because a target with zero lexical and semantic proximity may rank below 1000 in both retrievers. The large-pool baseline also judges $d$ without the bridge context that makes its relevance visible.

Why not just let the LLM reason about the query and rewrite it? Chain-of-thought pre-reasoning is cheap and already gives up to 12.2 points on BRIGHT. But its gains come from parametric knowledge, and Yoon et al. show that this kind of expansion falls below BM25 when the model has not memorized the evidence. Our matched/unmatched split tests whether conditional-gain selection closes that gap. If chain-of-thought rewriting equals us on unmatched queries as well, our grounding argument fails.

Why not just run InteR for more rounds? InteR is the natural iterative competitor, but it peaks at two rounds because nothing filters out redundant or drifting expansions. Our novelty term relative to the pool and the threshold $\tau$ act as that filter and as a stopping rule. We will run InteR at one to four rounds under the same budget. The check is whether our greedy loop keeps improving where InteR declines.

Decisive Checks

The idea succeeds if three results hold. On BRIGHT, first-stage Recall@100 at 256 judge calls per query must beat both chain-of-thought rewriting and uniform expansion over identical bridges. On unmatched FEVER and SciFact claims, Recall@5 must stay at or above BM25's 31.0, while the ungrounded-context ablation falls below it. On MISCIPLUS, accepted bridges must bring undermining passages into the pool more often than the marginal variant does.

The idea is refuted if the marginal variant or the no-control variant matches the full score. That would mean the dependency structure, the core of the proposal, contributes nothing beyond judging more documents.