Claim Verification Using a Multi-GAN Based Model

Proceedings of Recent Advances in Natural Language Processing, pages 494–503 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_056 494 Claim Verification using a Multi-GAN based Model Amartya Hatua University of Houston 4800 Calhoun Rd, Houston, TX 77004 amartyahatua@gmail.com Arjun Mukherjee University of Houston 4800 Calhoun Rd, Houston, TX 77004 arjun@uh.edu Rakesh M. Verma University of Houston 4800 Calhoun Rd, Houston, TX 77004 rmverma@cs.uh.edu Abstract This article describes research on claim verifi- cation carried out using a multiple GAN-based model. The proposed model consists of three pairs of generators and discriminators. The generator and discriminator pairs are responsi- ble for generating synthetic data for supported and refuted claims and claim labels. A the- oretical discussion about the proposed model is provided to validate the equilibrium state of the model. The proposed model is applied to the FEVER dataset, and a pre-trained language model is used for the input text data. The syn- thetically generated data helps to gain informa- tion that improves classification performance over state of the art baselines. The respective F1 scores after applying the proposed method on FEVER 1.0 and FEVER 2.0 datasets are 0.65±0.018 and 0.65±0.051. 1 Introduction Misleading claims and news are becoming perva- sive in our lives. Sometimes these are extremely difficult to identify. As a result, they can cause se- rious problems. This makes the research on claim verification essential. Fake news can be broadly classified into three categories (Rubin et al., 2015): i) Serious fabrications (uncovered in mainstream or participant media, yellow press or tabloids); ii) Large-scale hoaxes; and iii) Humorous fakes (news satire, parody, game shows). To solve this problem, research on this subject has evolved from knowledge-base oriented methods to sophisticated deep learning-based techniques. Related Work In (Mihalcea and Strapparava, 2009), the authors used natural language processing (NLP) techniques to detect fake news. They used tokenization and stemming for preprocessing the data and applied Naıve Bayes and Support Vector Machine (SVM) algorithms for classification. In recent research, the linguistic style and text source are considered the most critical factors to decide the genuineness of a fact or claim (Rashkin et al., 2017), (Baly et al., 2018), (Pérez-Rosas et al., 2017). Sometimes multiple sources of particular claims are used as external resources for claim verification. In (Rashkin et al., 2017), researchers compared the linguistic characteristics of real news with satire, hoaxes, and propaganda. They presented a case study based on the data collected by PolitiFact.com, where they used Glove for embedding, and Long Short Term Memory (LSTM) for prediction. To im- prove their result, they concatenate the Linguistic Inquiry and Word Count (LIWC) features (Pen- nebaker et al., 2001) with LSTM output vectors before the activation layer. LIWC features have played a vital role in claim verification research. LIWC extracts essential words that are part of psycho-linguistic categories and help in content analysis according to (Krippen- dorff, 2018; Neuendorf and Kumar, 2015). Their re- search work was extended by Kashyap et al. (Popat et al., 2018), who proposed an end-to-end frame- work for credibility analysis. This …

Content truncated to 500 words