基于事实信息的核查是目前不实信息核查研究的主流方法,但现有研究成果还存在文档检索中抽取的文档内容与待检测声明相关度不高、证据检索中忽略了句子间的内在联系以及声明验证中小语言模型的逻辑推理能力不足等问题。基于此,提出了一种融合大语言模型和证据抽取的事实核查模型。为提高待检测声明与文档内容的相关度,提出了DRCV(Document Retrieving for Claim Verification)文档检索算法;为了从文档中提取与声明最相关的句子作为证据,构建了“文档-声明对”训练证据检索模型,提出了基于关键词-注意力机制的证据检索方法;为增强模型的逻辑推理能力和提高事实核查的准确度,开发了基于大语言模型的声明验证模型,该模型选取参数量从5亿至1 300亿的七款大语言模型对声明进行验证,并利用其逻辑推理能力核查声明的事实性。基于真实数据集对提出的事实核查模型进行仿真实验,结果表明该模型进行事实核查的准确率比仅使用大语言模型高0.1%~34.0%,且比现有效果最好的模型准确率高1.8%。
Abstract
Fact-based verification is currently the mainstream method used in research on misinformation verification. However, existing research results have issues such as the low relevance between the document content extracted during document retrieval and the statements to be verified, the neglect of the intrinsic connections between sentences in evidence retrieval, and the insufficient logical reasoning ability of small language models in statement verification. To address these issues, a fact-checking model that integrates large language models and evidence extraction has been proposed. To enhance the relevance between the statements to be verified and the document content, a document retrieval for claim verification(DRCV) document retrieval algorithm was introduced. To extract the most relevant sentences from documents as evidence, a “document-statement pair” training evidence retrieval model was constructed, along with a keyword-attention mechanism-based evidence retrieval method. To improve the model’s logical reasoning capability and increase the accuracy of fact-checking, a statement verification model based on large language models was developed. This model selects seven large language models with parameter sizes ranging from 500 million to 130 billion to verify statements and utilizes their logical reasoning abilities to check the facts of statements. Simulation experiments was conducted on real datasets to evaluate the proposed fact-checking model. The results indicate that the accuracy of fact-checking using this model is 0.1% to 34.0% higher than that of models using only large language models, and it surpasses the accuracy of the currently best-performing model by 1.8%.
声明验证是利用证据检索得到的句子中包含的客观事实与声明中的内容进行对比,以核查该声明的事实性进行核查。Hanselowski等[15]将检索出的句子集合中的每条句子作为证据和声明进行简单拼接形成一个“证据-声明对”,将其输入到预训练语言模型作为三分类任务进行训练和测试,但该方法忽视了证据间的联系;Jiang等[13]和Stammbach等[17]使用小语言模型将句子集合中的所有句子和声明集中拼接形成“证据组-声明对”进行训练和测试;Krishna等[21]利用序列到序列(Sequence to Sequence)模型生成基于人类逻辑的推理信息作为证据,在对声明进行分类的基础上提供了可解释性。但这些方法或模型缺乏利用检索到的证据对声明进行事实核查过程的可解释性和逻辑推理能力。因此,如果将抽取到的证据作为一个整体输入到大语言模型中,既保留了证据间的逻辑关系,还能利用其逻辑推理能力对待检测声明进行分类、为事实核查任务提供可解释性和进一步提高事实核查的准确率。
SHUK, MAHUDESWARAND, WANGS H, et al. FakeNewsNet: A data repository with news content, social context, and spatiotemporal information for studying fake news on social media[J]. Big Data, 2020, 8(3): 171-188. DOI: 10.1089/big.2020.0062 .
SILVAA, LUOL, KARUNASEKERAS, et al. Embracing domain differences in fake news: Cross-domain fake news detection using multi-modal data[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2021, 35(1): 557-565. DOI: 10.1609/aaai.v35i1.16134 .
[4]
LIUZ W, ZHANGT L, YANGK L, et al. Emotion detection for misinformation: A review[J]. Information Fusion, 2024, 107: 102300. DOI: 10.1016/j.inffus.2024.102300 .
[5]
YUEZ, ZENGH, ZHANGY, et al. MetaAdapt: Dmain adaptive few-Shot misinformation detection via meta learning[C]//61st Annual Meeting of the Association for Computational Linguistics, ACL 2023. Association for Computational Linguistics (ACL), 2023: 5223-5239. DOI: https://doi.org/ 10.18653/v1/2023.acl-long.286 .
YANGY Z, ZHOUY M, YINGQ C, et al. A survey on fake news detection based on fact verification[J]. Journal of Communication University of China (Science and Technology), 2023, 30(6): 28-36. DOI: 10.16196/j.cnki.issn.1673-4793.2023.06.007(Ch ).
[8]
LIY P, HEH R, BAIJ, et al. MCFEND: A multi-source benchmark dataset for Chinese fake news detection[C]//Proceedings of the ACM Web Conference 2024. New York: ACM, 2024: 4018-4027. DOI: 10.1145/3589334.3645385 .
[9]
MUY D, BONTCHEVAK, ALETRASN. It’s about time: Rethinking evaluation on rumor detection benchmarks using chronological splits[C]//Findings of the Association for Computational Linguistics: EACL 2023. Stroudsburg: Association for Computational Linguistics, 2023: 736-743. DOI: 10.18653/v1/2023.findings-eacl.55 .
[10]
SUJ Y, ZHUOT Y, MANSUROVJ, et al. Fake news detectors are biased against texts generated by large language models[EB/OL]. 2023: arXiv: 2309.08674.
[11]
HUANGK H, MCKEOWNK, NAKOVP, et al. Faking fake news for real fake news detection: Propaganda-loaded training data generation[C]//Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Stroudsburg: Association for Computational Linguistics, 2023: 14571-14589. DOI: 10.18653/v1/2023.acl-long.815 .
[12]
THORNEJ, VLACHOSA, CHRISTODOULOPOULOSC, et al. FEVER: A large-scale dataset for fact extraction and VERification[C]//Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). Stroudsburg: Association for Computational Linguistics, 2018: 809-819. DOI: 10.18653/v1/n18-1074 .
[13]
LINY J, LINC Y, YEHC J, et al. CFEVER: A Chinese fact extraction and VERification dataset[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38(17): 18626-18634. DOI: 10.1609/aaai.v38i17.29825 .
[14]
JIANGK, PRADEEPR, LINJ. Exploring listwise evidence reasoning with T5 for fact verification[C]//Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers). Online. Stroudsburg: Association for Computational Linguistics, 2021: 402-410. DOI: 10.18653/v1/2021.acl-short.51 .
[15]
DEHAVENM, SCOTTS. BEVERS: A general, simple, and performant framework for automatic fact verification[C]//Proceedings of the Sixth Fact Extraction and VERification Workshop (FEVER). Stroudsburg: Association for Computational Linguistics, 2023: 58-65. DOI: 10.18653/v1/2023.fever-1.6 .
[16]
HANSELOWSKIA, ZHANGH, LIZ L, et al. UKP-athene: Multi-sentence textual entailment for claim verification[C]//Proceedings of the First Workshop on Fact Extraction and VERification (FEVER). Stroudsburg: Association for Computational Linguistics, 2018: 103-108. DOI: 10.18653/v1/w18-5516 .
[17]
SOLEIMANIA, MONZC, WORRINGM. BERT for evidence retrieval and claim verification[C]//Advances in Information Retrieval. Cham: Springer, 2020: 359-366.10.1007/978-3-030-45442-5_45. DOI: 10.1007/978-3-030-45442-5_45 .
[18]
STAMMBACHD. Evidence selection as a token-level prediction task[C]//Proceedings of the Fourth Workshop on Fact Extraction and VERification (FEVER). Dominican Republic. Stroudsburg: Association for Computational Linguistics, 2021: 14-20. DOI: 10.18653/v1/2021.fever-1.2 .
[19]
DUZ X, QIANY J, LIUX, et al. GLM: General language model pretraining with autoregressive blank infilling[C]//Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Stroudsburg: Association for Computational Linguistics, 2022: 320-335. DOI: 10.18653/v1/2022.acl-long.26 .
[20]
ZENGA H, LIUX, DUZ X, et al. GLM-130B: An open bilingual pre-trained model[EB/OL]. 2022: arXiv: 2210.02414.
[21]
KAPLANJ, MCCANDLISHS, HENIGHANT, et al. Scaling laws for neural language models[EB/OL]. 2020: arXiv: 2001.08361.
[22]
KRISHNAA, RIEDELS, VLACHOSA. ProoFVer: Natural logic theorem proving for fact verification[J]. Transactions of the Association for Computational Linguistics, 2022, 10: 1013-1030. DOI: 10.1162/tacl_a_00503 .
[23]
HUANGJ, CHANGK C C. Towards reasoning in large language models: A survey[C]//Findings of the Association for Computational Linguistics: ACL 2023. Stroudsburg: Association for Computational Linguistics, 2023: 1049-1065. DOI: 10.18653/v1/2023.findings-acl.67 .
[24]
PANL M, WUX B, LUX Y, et al. Fact-checking complex claims with program-guided reasoning[C]//Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Stroudsburg: Association for Computational Linguistics, 2023: 6981-7004. DOI: 10.18653/v1/2023.acl-long.386 .
[25]
ZHANGX, GAOW. Towards LLM-based fact verification on news claims with a hierarchical step-by-step prompting method[C]//Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). Stroudsburg: Association for Computational Linguistics, 2023: 996-1011. DOI: 10.18653/v1/2023.ijcnlp-main.64 .
[26]
HUB Z, SHENGQ, CAOJ, et al. Bad actor, good advisor: Exploring the role of large language models in fake news detection[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38(20): 22105-22113. DOI: 10.1609/aaai.v38i20.30214 .
[27]
CUIY M, CHEW X, LIUT, et al. Pre-training with whole word masking for Chinese BERT[J]. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021, 29: 3504-3514. DOI: 10.1109/TASLP.2021.3124365 .
[28]
WEIJ, TAY Y, BOMMASANIR, et al. Emergent abilities of large language models[EB/OL]. 2022: arXiv: 2206.07682.
[29]
HUANGL, YUW J, MAW T, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions[EB/OL]. 2023: arXiv: 2311.05232.