Open-domain question answering (QA) systems are frequently evaluated on a dual requirement: they must not only identify a factually correct answer string, but they must also provide provenance—a specific supporting document from a trusted, restricted target corpus that proves the answer is correct.
This setup creates a fundamental tension between information discovery and formal verification. To solve this, researchers developed cross-corpus answer projection and validation techniques. The core intuition is simple: it is far easier to discover a fact in a massive, highly redundant corpus (like the open Web) and map it back to a local collection than it is to find that fact within the restricted collection alone.
The Power of Data Redundancy
Why is finding an answer in a small, static corpus difficult? A restricted corpus has low data redundancy. If a question asks for an obscure fact, the local dataset might contain only a single sentence addressing it. To extract the answer, a system must perfectly parse that sentence's complex syntax and grammar.
Conversely, the open Web is massively redundant. As Brill et al. argued, leveraging massive amounts of data can substitute for complex linguistic processing. On the Web, a fact is stated in thousands of different ways. Simple surface patterns or keyword matches are highly likely to succeed because someone, somewhere, has phrased the answer exactly the way the system expects.
To exploit this asymmetry, researchers split open-domain QA into two main cross-corpus paradigms:
- Discovery-First (Web to Corpus): Systems query the open Web or a massive proxy corpus to harvest candidate answers easily, leveraging high data redundancy. They then project these candidates back into the restricted target corpus to locate an authoritative supporting document. Lin et al. implemented this in the Aranea system, treating the Web as a high-recall discovery engine before filtering for local provenance. Similarly, Qiu et al. used the open Web to generate candidate answers that were subsequently mapped back to a restricted target corpus for verification.
- Verification-First (Corpus to Web): Systems extract candidate answers from the local corpus first, then validate or rerank them by checking how frequently they appear in external web search results. For instance, Ittycheriah et al. used web search frequency as a verification filter to boost the confidence of local candidates. Adafre et al. demonstrated that high agreement between local extractions and external web snippets serves as a robust sanity check against internal retrieval errors. In community forums, Nakov demonstrated that automatically generating search queries from question-answer pairs to retrieve external web evidence significantly improves answer veracity checks.
The Alignment Bottleneck
While discovering a fact on the Web is straightforward, mapping it back to a restricted corpus introduces a severe performance drop known as the answer projection bottleneck.
When a system finds a correct answer string (e.g., "1969") on the open Web, it must locate a document within the static target corpus that explicitly justifies this answer. If the target corpus lacks a matching passage, or if the passage uses highly divergent phrasing, the system cannot provide a valid document reference. In benchmark evaluations, an answer that is globally true but locally ungrounded is penalized as "unsupported."
[User Query] ──> Search Open Web (High Redundancy) ──> Found Fact: "1969"
│
(Answer Projection)
│
▼
Search Target Corpus (Low Redundancy)
├── Match Found? ──> Output Answer + Doc ID
└── No Match? ──> Penalty: "Unsupported Answer"
Katz et al. identified this answer projection process as a primary bottleneck in hybrid QA systems, noting that it is highly prone to failure when the target corpus lacks the exact phrasing or context found online. Huang et al. and Jijkoun et al. independently observed this gap: identifying a correct fact is fundamentally easier than locating its specific authoritative source within a limited collection. As Cucerzan et al. summarized, the core difficulty shifts from information discovery to evidence mapping.
Anchoring with Structured Knowledge Bases
To mitigate the noise of open-web projection, systems can use external structured knowledge bases, dictionaries, and encyclopedias to act as semantic anchors.
Instead of projecting raw, unstructured text snippets, systems project known semantic relationships. Hildebrandt et al. demonstrated that definitions mined from external dictionaries can be converted into targeted search queries. These queries project structured semantic definitions onto unstructured target text, making it easier to isolate and validate supporting sentences.
For complex queries, Ren et al. validated local candidates by calculating their cosine similarity to definitions retrieved from high-authority external encyclopedias like Wikipedia. By using structured external knowledge as a reference standard, systems bridge the gap between open-domain discovery and strict local verification.
Key Foundational Work
Extracting Answers from the Web Using Data Annotation and Knowledge Mining Techniques (2002) — contributed 4 concept elements to this direction.
Data-Intensive Question Answering (2001) — contributed 4 concept elements to this direction.
IBM's Statistical Question Answering System-TREC 11 (2002) — contributed 2 concept elements to this direction.