Jointly Learning Evidence Retrieval and Claim Verification

This direction is merged into another direction: joint retrieval-verification training. Its last version stays readable and citable here.

Version 1 of 117 papers25 conceptsfrozen 2026-10-04

Traditional automated fact-checking systems operate like a relay race. First, a retrieval module finds relevant documents; then, a selection module picks out specific sentences; finally, a verification module decides if those sentences support or refute a claim.

The problem with this "pipeline" approach is error propagation. If the retriever picks the wrong document, the verifier never sees the truth. If the selector misses a crucial nuance, the verifier makes a judgment based on incomplete data. Researchers are moving away from these brittle sequences toward joint multi-task learning, where the act of finding evidence and the act of verifying a claim happen simultaneously within a single, integrated model.

The Pipeline Problem and the Multi-task Solution

In a standard pipeline, the components are "blind" to each other's needs. A retriever might find a document that is topically relevant (e.g., about "Climate Change") but lacks the specific evidentiary power needed to verify a nuanced claim.

As James Thorne and colleagues noted in the FEVER Shared Task, combining large-scale retrieval with natural language inference is a formidable challenge. To solve this, models like EMBA, developed by Sizhen Li and Bo Cheng, treat retrieval and classification as a joint task. By sharing architectural components, the model leverages mutual signals: the features used to identify relevant context are forced to align with the requirements of the final classification.

Bidirectional Feedback Loops

The intuition behind joint optimization is that the "answer" should help find the "proof." In a pipeline, information only flows forward. In a joint system, the verification stage provides a feedback signal that refines the retrieval stage.

Wenpeng Yin and Dan Roth demonstrated this with TwoWingOS (TwoWingOS: A Two-Wing Optimization Strategy for Evidential Claim Verification). By allowing the two "wings" of the model—evidence identification and claim verification—to train together, the system learns which types of evidence are most useful for making a final judgment. Similarly, Eduard Dragut and Arjun Mukherjee showed that using entailment signals can push the most definitive sentences to the top of the search results, significantly enhancing the quality of evidence.

From Sentences to Paragraphs

While early work focused on isolated sentences, recent breakthroughs emphasize broader context. Xiangci Li, Nanyun Peng, and Gully Burns introduced paragraph-level multi-task models for scientific fact-checking.

By using "compact paragraph encoding," these models avoid the loss of context that occurs when sentences are stripped from their surroundings. This holistic representation allows the system to better distinguish between relevant evidence and distracting information, a critical requirement for complex domains like medicine or physics.

Latent Variables and Interpretability

One risk of joint models is that they can become "black boxes." To counter this, Martin Fajcik and his team developed Claim-Dissector (Claim-Dissector: An Interpretable Fact-Checking System with Joint Re-ranking and Veracity Prediction).

They treated the relevance of individual evidence pieces as latent variables. This means the model doesn't just output a "True" or "False" label; it internally calculates the probability that each piece of evidence is relevant to that label. This creates a built-in rationale, providing clear evidence rankings at the document and sentence level while maintaining the performance benefits of joint training.

Cross-Domain Robustness

The benefits of this integrated approach extend beyond general fact-checking into technical domains. Wenhao Yu and Lingfei Wu applied joint training to Technical Question Answering, showing that when the reading comprehension component provides feedback to the retrieval component, the system identifies actionable solutions rather than just simple keyword matches.

By treating retrieval and reasoning as two sides of the same coin, these researchers have created systems that are not only more accurate but also more resilient to the complexities of real-world information.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_jointly_learning_evidence_retrieval_and__26e4240b,
  title = {Jointly Learning Evidence Retrieval and Claim Verification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/jointly-learning-evidence-retrieval-and-claim-verification/rdr_26e4240b370e438c95d14300ce3f9042},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_jointly_learning_evidence_retrieval_and__26e4240b_v1,
  title = {Jointly Learning Evidence Retrieval and Claim Verification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/jointly-learning-evidence-retrieval-and-claim-verification/rdr_26e4240b370e438c95d14300ce3f9042/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}