To verify whether a claim is true or false, an automated system must solve two distinct problems: it must find the relevant evidence within a massive corpus (retrieval), and it must reason over that evidence to determine a verdict of support, refutation, or insufficient information (verification).
Traditionally, these tasks have been treated as a sequential pipeline. A retrieval module filters out a small set of candidate sentences, which are then handed off to a separate verification module. However, as noted by Hongbin Lin et al. in Heterogeneous-Graph Reasoning and Fine-Grained Aggregation for Fact Checking, isolating these steps introduces a fundamental vulnerability: error propagation. If the retriever fails to select the exact semantic rationale required for verification, or if it introduces distracting noise, the downstream verifier has no way to recover. As argued by Xiangci Li et al. in A Paragraph-level Multi-task Learning Model for Scientific Fact-Verification, Nanyun Peng 0001 et al. in A Paragraph-level Multi-task Learning Model for Scientific Fact-Verification, and Martin Fajcik et al. in Claim-Dissector: An Interpretable Fact-Checking System with Joint Re-ranking and Veracity Prediction, a breakdown in the first stage inevitably dooms the second.
The Intuition of Joint Optimization
The core insight of joint retrieval-verification training is that finding evidence and evaluating truth are not independent tasks; they are deeply complementary. Instead of optimizing each component in isolation, a joint framework trains both modules simultaneously using a shared loss function or shared internal representations.
Why does this work? In a standard pipeline, the retriever is typically trained on a proxy objective, such as lexical overlap or general semantic similarity. However, a sentence can be highly topically relevant to a claim without containing the precise logical nugget needed to prove or disprove it.
By unifying the tasks, the final veracity prediction loss can backpropagate all the way into the retrieval layers. This creates a crucial feedback loop: the model learns to select evidence because it is useful for making the final judgment, not just because it looks similar on the surface. Wenpeng Yin et al. conceptualize this in TwoWingOS: A Two-Wing Optimization Strategy for Evidential Claim Verification as a "two-wing" optimization strategy, where feedback from the verification stage actively refines and sharpens the selection of relevant evidence. Similarly, Christopher Hidey et al. demonstrate in Team SWEEPer: Joint Sentence Extraction and Fact Checking with Pointer Networks that a joint multi-task objective encourages the network to prioritize inferentially rich evidence over mere topical matches.
Shared Representations and Mutual Signals
At an architectural level, joint training typically relies on multi-task learning with shared encoders. As established during the foundational Fact Extraction and VERification (FEVER) shared task described by James Thorne et al. in The Fact Extraction and VERification (FEVER) Shared Task, sharing representations between the retrieval and reasoning components ensures that the features used to identify context are inherently aligned with the requirements of the classification task.
This alignment creates a powerful synergy. For instance, Sizhen Li et al. introduced an end-to-end framework in An End-to-End Multi-task Learning Model for Fact Checking that leverages these mutual signals to match the performance of complex pipelines while drastically simplifying the underlying architecture. This cross-task benefit holds true even in highly specialized domains. For example, Juraj Vladika et al. found in HealthFC: Verifying Health Claims with Evidence-Based Medical Fact-Checking that concurrently learning evidence selection and veracity prediction yields superior results when navigating complex medical literature, where the interplay between nuanced clinical evidence and health claims is exceptionally tight.
Mechanisms for Information Flow
How exactly do these two components communicate during training? Researchers have developed several distinct mechanisms to facilitate this bidirectional information flow:
- Latent Evidence Selection: One elegant approach is to treat the selected evidence as a hidden or latent variable. In the framework proposed by Xuming Hu et al. in CHEF: A Pilot Chinese Dataset for Evidence-Based Fact-Checking, the model does not rely on a fixed, pre-defined ranker; instead, the veracity prediction task directly guides the latent retrieval component to discover which sentences are truly informative. The authors of Claim-Dissector: An Interpretable Fact-Checking System with Joint Re-ranking and Veracity Prediction similarly utilize a latent variable model to build a linear ensemble of per-evidence probabilities, ensuring the final judgment is mathematically bound to the ranked evidence.
- Attention Modulation and Refinement: Rather than hard selection, some architectures use soft attention to bridge the gap. Canasai Kruengkrai et al. developed a multi-level attention model in A Multi-Level Attention Model for Evidence-Based Fact Checking where sentence selection scores are used to directly modulate the evidence representations passed to the veracity predictor. Taking this a step further, Jean-Flavien Bussotti et al. show in Refining Attention for Explainable and Noise-Robust Fact-Checking with Transformers that concurrently training an evidence labeling module explicitly refines transformer attention weights, making the model robust to noise while providing faithful, explainable rationales.
Beyond Fact-Checking
The paradigm of joint retrieval-verification training is not confined strictly to fact-checking benchmarks; it reflects a broader shift in open-domain text processing. For instance, in technical question answering, Wenhao Yu et al. demonstrated in Technical Question Answering across Tasks and Domains that jointly training document retrieval and reading comprehension modules allows the reader to provide feedback to the retriever, steering it toward documents containing actionable solutions rather than simple keyword matches.
Similarly, Zhiwei Zhang et al. framed scientific claim verification as a machine reading comprehension task in Abstract, Rationale, Stance: A Joint Model for Scientific Claim Verification, allowing document retrieval and evidence selection to benefit from a continuous, bidirectional flow of information. By breaking down the walls between finding information and understanding it, these joint architectures ensure that the "why" of an explanation is always fundamentally linked to the "what."
Key Foundational Work
Team SWEEPer: Joint Sentence Extraction and Fact Checking with Pointer Networks (2018) — contributed 2 concept elements to this direction.
An End-to-End Multi-task Learning Model for Fact Checking (2018) — contributed 2 concept elements to this direction.
TwoWingOS: A Two-Wing Optimization Strategy for Evidential Claim Verification (2018) — contributed 2 concept elements to this direction.