Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fact Checking Beyond Training Set

Payam Karisani, Heng Ji

TL;DR

This work tackles cross-domain generalization in automatic fact checking by analyzing the retriever-reader pipeline under distribution shift. It introduces two core advances: adversarially trained target encoders for robust evidence retrieval and a representation-alignment reader augmented with reversal-order cues and a CORAL-based discrepancy loss to stabilize veracity prediction across domains. The authors repurpose MultiFC and Snopes into eight domain-transfer scenarios and show that their methods achieve state-of-the-art performance against strong baselines, including GPT-4-based synthetic data pipelines. The result is a more robust, domain-transferable fact-checking system with practical implications for real-world deployment where domain gaps are common.

Abstract

Evaluating the veracity of everyday claims is time consuming and in some cases requires domain expertise. We empirically demonstrate that the commonly used fact checking pipeline, known as the retriever-reader, suffers from performance deterioration when it is trained on the labeled data from one domain and used in another domain. Afterwards, we delve into each component of the pipeline and propose novel algorithms to address this problem. We propose an adversarial algorithm to make the retriever component robust against distribution shift. Our core idea is to initially train a bi-encoder on the labeled source data, and then, to adversarially train two separate document and claim encoders using unlabeled target data. We then focus on the reader component and propose to train it such that it is insensitive towards the order of claims and evidence documents. Our empirical evaluations support the hypothesis that such a reader shows a higher robustness against distribution shift. To our knowledge, there is no publicly available multi-topic fact checking dataset. Thus, we propose a simple automatic method to re-purpose two well-known fact checking datasets. We then construct eight fact checking scenarios from these datasets, and compare our model to a set of strong baseline models, including recent domain adaptation models that use GPT4 for generating synthetic data.

Fact Checking Beyond Training Set

TL;DR

This work tackles cross-domain generalization in automatic fact checking by analyzing the retriever-reader pipeline under distribution shift. It introduces two core advances: adversarially trained target encoders for robust evidence retrieval and a representation-alignment reader augmented with reversal-order cues and a CORAL-based discrepancy loss to stabilize veracity prediction across domains. The authors repurpose MultiFC and Snopes into eight domain-transfer scenarios and show that their methods achieve state-of-the-art performance against strong baselines, including GPT-4-based synthetic data pipelines. The result is a more robust, domain-transferable fact-checking system with practical implications for real-world deployment where domain gaps are common.

Abstract

Evaluating the veracity of everyday claims is time consuming and in some cases requires domain expertise. We empirically demonstrate that the commonly used fact checking pipeline, known as the retriever-reader, suffers from performance deterioration when it is trained on the labeled data from one domain and used in another domain. Afterwards, we delve into each component of the pipeline and propose novel algorithms to address this problem. We propose an adversarial algorithm to make the retriever component robust against distribution shift. Our core idea is to initially train a bi-encoder on the labeled source data, and then, to adversarially train two separate document and claim encoders using unlabeled target data. We then focus on the reader component and propose to train it such that it is insensitive towards the order of claims and evidence documents. Our empirical evaluations support the hypothesis that such a reader shows a higher robustness against distribution shift. To our knowledge, there is no publicly available multi-topic fact checking dataset. Thus, we propose a simple automatic method to re-purpose two well-known fact checking datasets. We then construct eight fact checking scenarios from these datasets, and compare our model to a set of strong baseline models, including recent domain adaptation models that use GPT4 for generating synthetic data.
Paper Structure (15 sections, 8 equations, 4 figures, 11 tables)

This paper contains 15 sections, 8 equations, 4 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Proposed Model
  4. Problem Statement
  5. Adversarial Training for Evidence Retrieval
  6. Representation Alignment for Veracity Prediction
  7. Training and Testing Procedures
  8. Experimental Setup
  9. Results and Analysis
  10. Conclusions
  11. Limitations
  12. Complementary Reports About the Datasets
  13. Complementary Information About the Training Setup
  14. Complementary Related Work
  15. Complementary Experiments

Figures (4)

  • Figure 1: On August 2, 2023, the Associated Press (and other news outlets) reported that Donald Trump was indicted. The questions were asked from the LLMs on August 17, 2023. As of December 2023, these models are still unable to verify this claim.
  • Figure 2: \ref{['fig:pipeline']}) Commonly used fact checking (FC) pipeline consists of a retrieval model (called the retriever), and a veracity prediction model (called the reader). \ref{['fig:indomain-fc-f1']}) The performance (Macro F1) of the pipeline across two domains (Misc vs Politics) in two scenarios (in-domain vs out-of-domain). \ref{['fig:indomain-retriever-ndcg']}) The performance (NDCG@10) of the retriever across the two mentioned domains. \ref{['fig:indomain-reader-f1']}) The performance (Macro F1) of the reader across the two domains.
  • Figure 3: \ref{['fig:retriever-a']}) The source retriever is a bi-encoder, and uses dot product as the loss function. \ref{['fig:retriever-b']}) We fix the parameters of the source claim encoder, and adversarially train the target claim encoder to mimic the source model. This step is done using unlabeled data in the two domains. \ref{['fig:retriever-c']}) Next, we fix the parameters of the source document encoder, and adversarially train the target document encoder. Similarly, this step does not need labeled data. \ref{['fig:retriever-d']}) The two trained target encoders can be used for the retrieval task in the target domain. The components that have gray outline show the inputs, outputs, and objective terms. The rest are neural networks. The parameters of the components that have gray background are fixed during training.
  • Figure 4: \ref{['fig:reader']}) The reader model. Dashed line indicates shared parameters. S/D, S/C, T/D, and T/C stand for source documents, source claims, target documents, and target claims respectively. The symbol $\mathbin\Vert$ is the concatenation operator. \ref{['fig:multifc-scatter']}-\ref{['fig:snopes-scatter']}) The 2D projection of the claims in the MultiFC and Snopes datasets (using t-SNE). The vectors are the outputs of a BERT classifier, after being trained to predict the domains. Figure best viewed in color.