CONCRETE: Improving Cross-lingual Fact-checking with Cross-lingual Retrieval

Proceedings of the 29th International Conference on Computational Linguistics, pages 1024–1035 October 12–17, 2022. 1024 CONCRETE: Improving Cross-lingual Fact-checking with Cross-lingual Retrieval Kung-Hsiang Huang ChengXiang Zhai Heng Ji Department of Computer Science, University of Illinois Urbana-Champaign {khhuang3, czhai, hengji}@illinois.edu Abstract Fact-checking has gained increasing attention due to the widespread of falsified information. Most fact-checking approaches focus on claims made in English only due to the data scarcity issue in other languages. The lack of fact- checking datasets in low-resource languages calls for an effective cross-lingual transfer tech- nique for fact-checking. Additionally, trustwor- thy information in different languages can be complementary and helpful in verifying facts. To this end, we present the first fact-checking framework augmented with cross-lingual re- trieval that aggregates evidence retrieved from multiple languages through a cross-lingual re- triever. Given the absence of cross-lingual information retrieval datasets with claim-like queries, we train the retriever with our proposed Cross-lingual Inverse Cloze Task (X-ICT), a self-supervised algorithm that creates training instances by translating the title of a passage. The goal for X-ICT is to learn cross-lingual retrieval in which the model learns to identify the passage corresponding to a given translated title. On the X-FACT dataset, our approach achieves 2.23% absolute F1 improvement in the zero-shot cross-lingual setup over prior systems. The source code and data are pub- licly available at https://github.com/ khuangaf/CONCRETE. 1 Introduction Fact-checking is an important task that assesses the veracity of a claim. This task has gained increas- ing attention due to the widespread mis- and dis- information that has a significant socioeconomic impact on our society (Scheufele and Krause, 2019; Pate et al., 2019; Fung et al., 2021; Guo et al., 2022; Wu et al., 2022; Fung et al., 2022; Huang et al., 2022). While fact-checking is mainly con- ducted manually, especially in the journalism indus- try, with more than hundred of millions of social media posts and millions of blog posts published per day (Hoang and Mothe, 2018; Djuraskovic, 2022), manual fact-checking is no longer feasible. Hence, we are urgently in need of reliable auto- mated fact-checking approaches. Most existing work develops fact-checking ap- proaches on English-only corpus (Wang, 2017; Thorne et al., 2018; Augenstein et al., 2019; Wad- den et al., 2020). One reason for this is the scarcity of fact-checking websites in other lan- guages for constructing large enough non-English fact-checking datasets. Therefore, it is even more challenging to build fact-checkers for low-resource languages. A solution is to leverage high-resource languages with zero-shot cross-lingual transfer, where the model is trained on source languages of richer resources and directly tested on target lan- guages of lower resources. Note that the sets of languages in the training set and the test sets are disjoint. With this technique, ground-truth labels for claims in low-resource languages are no longer required. Few studies in the fact-checking literature have explored the cross-lingual setup. One line of work attempts to match an input claim with claims in other languages that have been verified (Kazemi et al., 2021, 2022). However, …

Content truncated to 500 words