[Paper Review] Learning Hierarchical Discourse-level Structure for Fake News Detection

Hamid Reza Karimi, Jiliang Tang|arXiv (Cornell University)|Feb 27, 2019

Misinformation and Its Impacts38 references35 citations

TL;DR

HDSF automatically learns a hierarchical discourse-level dependency structure for documents and uses it to produce structurally rich representations for fake news classification, outperforming baselines.

ABSTRACT

On the one hand, nowadays, fake news articles are easily propagated through various online media platforms and have become a grand threat to the trustworthiness of information. On the other hand, our understanding of the language of fake news is still minimal. Incorporating hierarchical discourse-level structure of fake and real news articles is one crucial step toward a better understanding of how these articles are structured. Nevertheless, this has rarely been investigated in the fake news detection domain and faces tremendous challenges. First, existing methods for capturing discourse-level structure rely on annotated corpora which are not available for fake news datasets. Second, how to extract out useful information from such discovered structures is another challenge. To address these challenges, we propose Hierarchical Discourse-level Structure for Fake news detection. HDSF learns and constructs a discourse-level structure for fake/real news articles in an automated and data-driven manner. Moreover, we identify insightful structure-related properties, which can explain the discovered structures and boost our understating of fake news. Conducted experiments show the effectiveness of the proposed approach. Further structural analysis suggests that real and fake news present substantial differences in the hierarchical discourse-level structures.

Motivation & Objective

• Investigate whether hierarchical discourse-level structure can discriminate fake vs. real news.
• Develop an end-to-end framework that learns discourse dependencies without annotated data.
• Create structurally informed document representations for effective fake news classification.
• Identify structure-related properties that differentiate fake and real news and relate to coherence.

Proposed method

• Represent each sentence with BLSTM-based embeddings derived from word embeddings.
• Learn inter-sentential dependency probabilities to form a discourse dependency tree via an attention-based matrix A and root probabilities r.
• Greedily construct a discourse tree using A and r to obtain a tree structure for each document.
• Compute structurally aware sentence representations p_j and c_j from potential parents and children, then derive g_j.
• Aggregate g_j to form a structurally rich document representation x and perform binary fake/real classification using cross-entropy loss.
• Train the entire framework end-to-end with backpropagation; note that tree construction is post hoc and not differentiated.

Experimental results

Research questions

• RQ1 Does the proposed HDSF framework improve fake news detection performance compared to baselines?
• RQ2 What structure-related properties of discourse trees differentiate fake and real news and how do they relate to coherence?

Key findings

• HDSF significantly outperforms baselines on the combined dataset (82.19% accuracy vs. 80.54% for LSTM[w+s]).
• Document structure-aware representations yield stronger discrimination than content-only features like N-grams or LIWC.
• Discourse dependency trees reveal substantial differences between fake and real news in all three proposed properties, with real news showing higher coherence.
• A post hoc greedy tree construction method uses inter-sentential probabilities to assemble a root and parent-child relations.
• Training and development curves indicate decreasing training error and increasing accuracy during optimization.