PaperlayerView exact revision on arXiv

Source-linked AI summary

TI-CNN: Convolutional Neural Networks for Fake News Detection

Yang Yang, Lei Zheng, Jiawei Zhang, Qingcai Cui, Zhoujun Li, Philip S. Yu

arXiv:1806.00749v3cs.CLcs.SI

TL;DR

Fake-news detection is difficult because deceptive content spreads widely while public datasets and existing automated methods remain limited. The paper proposes TI-CNN, which jointly learns explicit and latent features from text and images in a unified space. Experiments on real-world datasets report that TI-CNN successfully identifies fake news.

  • Problem

    Fake-news identification is difficult because few large-scale public datasets exist and deceptive language is strategically written to avoid detection.

  • Method

    TI-CNN jointly models text and image information by combining explicit features with latent features learned through convolutional neural networks.

  • Results

    TI-CNN successfully identifies fake news using explicit features and the combined text-image approach in experiments on real-world datasets.

  • Takeaways & Limitations

    Combining text and image information provides an effective approach for recognizing fake news from large volumes of online information.

Abstract

from arXiv · show

With the development of social networks, fake news for various commercial and political purposes has been appearing in large numbers and gotten widespread in the online world. With deceptive words, people can get infected by the fake news very easily and will share them without any fact-checking. For instance, during the 2016 US president election, various kinds of fake news about the candidates widely spread through both official news media and the online social networks. These fake news is usually released to either smear the opponents or support the candidate on their side. The erroneous information in the fake news is usually written to motivate the voters' irrational emotion and enthusiasm. Such kinds of fake news sometimes can bring about devastating effects, and an important goal in improving the credibility of online social networks is to identify the fake news timely. In this paper, we propose to study the fake news detection problem. Automatic fake news identification is extremely hard, since pure model based fact-checking for news is still an open problem, and few existing models can be applied to solve the problem. With a thorough investigation of a fake news data, lots of useful explicit features are identified from both the text words and images used in the fake news. Besides the explicit features, there also exist some hidden patterns in the words and images used in fake news, which can be captured with a set of latent features extracted via the multiple convolutional layers in our model. A model named as TI-CNN (Text and Image information based Convolutinal Neural Network) is proposed in this paper. By projecting the explicit and latent features into a unified feature space, TI-CNN is trained with both the text and image information simultaneously. Extensive experiments carried on the real-world fake news datasets have demonstrate the effectiveness of TI-CNN.

I. INTRODUCTION

Fake news is difficult to detect because it spreads widely, is strategically written, and provides limited public training data. TI-CNN addresses these challenges by jointly modeling text and image information with explicit and learned latent features.

  • Fake news spreads rapidly through social networks and can significantly affect political events such as the 2016 US presidential election.
  • Detection is challenging because few large-scale public datasets exist and fake news is deliberately written to avoid detection.
  • Text-only representations often isolate words from context and struggle with sparse, high-order features.
  • Fake news contains detectable textual and visual leakages, including differences in lexical diversity, irrelevant or altered images, and mismatches between text and images.
  • TI-CNN combines explicit and convolutionally learned latent features from text and images in a unified space for fake-news identification.
  • The paper contributes a high-quality dataset, text analysis, effective image features, and a unified convolutional model for fake-news recognition.

II. RELATED WORK

Prior work approaches deception detection through linguistic and network information, while deep learning provides representations and nonlinear models for text and image-based classification. The paper formalizes fake-news detection as predicting real or fake labels from extracted text-image features.

  • II. RELATED WORK: Deception detection research includes linguistic approaches and network approaches covering scientific fraud, fake news, and false tweets.
  • II. RELATED WORK: Word-analysis methods have been extended with sentiment, lexical, content, stylistic, semantic, and syntactic information for deceptive-text detection.
  • II. RELATED WORK: CNNs use filters to capture features, supporting deep-learning representations in natural-language processing and related applications.
  • III. PROBLEM DEFINITION: The problem represents each article as a text-image tuple and assigns labels [1, 0] for real news or [0, 1] for fake news.
  • III. PROBLEM DEFINITION: Explicit and latent features are extracted from article text and images, and a model f maps the resulting feature representation to news labels.

IV. DATA ANALYSIS

The paper analyzes election-related real and fake news using textual and image information to identify quantitative clues for detection. The dataset combines fake-news material from many websites with real news from authoritative sources.

  • The analysis examines real and fake 2016 US presidential-election news through computational linguistic, sentiment, psychological, and image-related perspectives.
  • The dataset contains 20,015 news articles: 11,941 fake and 8,074 real.
  • Fake-news data were scraped from more than 240 websites, while real news came from authoritative outlets including The New York Times and The Washington Post.
  • Fake-news titles frequently contain ‘notitle’, capitalized words, and meaningless numbers representing special characters.

1) Computational Linguistic:

The analysis identifies consistent linguistic differences between real and fake news in article length, punctuation, capitalization, cognitive wording, pronouns, and motion verbs.

  • 1) Computational Linguistic:: 4,360 words versus 3,943, and 84 sentences versus 69, are reported for real and fake news, respectively.Real news averages 51.9 words per sentence, compared with 57.1 for fake news; fake-news sentence lengths also show greater variance.
  • 1) Computational Linguistic:: Fake news uses more question marks, with rhetorical questions used to emphasize ideas and intensify sentiment.
  • 1) Computational Linguistic:: Fake-news exclamation counts are generally low, but their box-plot inner fence is much larger than real news’s.The analysis associates exclamations with strong commands or emotional outbursts that can fan specific emotions.
  • 1) Computational Linguistic:: Fake news contains substantially more capital letters, which the analysis links to attracting readers’ attention and emphasizing ideas.
  • 1) Computational Linguistic:: Real news uses more negations and exclusive words, while fake news uses fewer first- and second-person pronouns, more third-person pronouns, and fewer motion verbs.The paper presents these patterns as cognitive and psychological differences in how deceptive information is written.

3) Lexical Diversity:

The paper examines lexical diversity, sentiment, and image properties as distinguishing characteristics of real and fake news.

  • 3) Lexical Diversity:: Lexical diversity measures how many different words a text uses, while lexical density measures the proportion of lexical items.
  • 3) Lexical Diversity:: 2.2e-06 versus 1.76e-06, the reported lexical-diversity values for real and fake news, respectively, indicate greater diversity in real news.
  • 4) Sentiment Analysis:: Real news is reported as more positive than negative, whereas fake news shows more negative sentiment and greater variation in negative sentiment.
  • Image Analysis: 0.366 faces versus 0.299, the average counts reported for real and fake news images, respectively, show more faces in real-news images.Real-news images are also reported to have better resolution than fake-news images.
  • Model Architecture: TI-CNN combines explicit and latent textual and visual features by projecting them into a shared feature space and fusing the resulting representations.The architecture uses parallel CNNs to extract latent features from textual and visual information for fake-news detection.

A. Text Branch

The text branch combines explicit statistics of news text with latent features learned by a CNN over word embeddings.

  • A. Text Branch: Textual explicit features encode statistics such as news length, sentence count, question marks, exclamations, and capital letters.These statistics form a fixed-size vector that is transformed through a fully connected layer.
  • A. Text Branch: The latent text subbranch applies convolutional filters to adjacent word embeddings to learn local word-level features.CNNs produce local features around words and combine them into a fixed-sized representation for fake-news detection.
  • A. Text Branch: The news is represented as a matrix of word embeddings, with shorter articles padded to a maximum length n.
  • A. Text Branch: Each convolutional filter processes a window of h adjacent words to generate a feature map through a nonlinear transformation.The feature map is formed by applying the filter across all possible word windows.
  • A. Text Branch: Max pooling retains the largest convolutional response, and the pooled results feed a fully connected layer for textual latent features.The paper describes max pooling as improving robustness by retaining the most important convolutional results.

B. Image Branch

The image branch combines hand-crafted visual features with CNN-learned latent features. Convolution and max-pooling extract position-robust representations from raw news images.

  • Visual explicit features encode image resolution and the number of faces through a fully connected layer.
  • A separate CNN learns latent visual features directly from raw images because explicit features are hand-crafted rather than data-driven.
  • Convolutional filters scan image regions with shared parameters, producing translation-invariant feature maps.
  • Max-pooling selects maximum activations over local regions, provides position invariance, and downsamples images by Kα and Kβ.

C. Rectified Linear Neuron

The model uses ReLU in the image branch to address gradient problems associated with sigmoid and tanh activations, while regularization and likelihood-based training support generalization.

  • C. Rectified Linear Neuron: ReLU is added to the image branch because sigmoid and tanh may cause exploding or vanishing gradients.
  • C. Rectified Linear Neuron: ReLU computes a simple gradient and sets negative elements to zero, which can speed neural-network training.
  • D. Regularization: Dropout and l2-norms are used to prevent overfitting, while early stopping is also treated as regularization.
  • E. Training: The network minimizes negative log likelihood, transforms label scores into conditional probabilities with a sigmoid, and uses RMSprop with back-propagation.

VI. EXPERIMENTS

The experiments describe a case study, data split, repeated runs, and model specifications for TI-CNN. The reported setup combines textual and visual branches with explicit and latent subbranches.

  • A. Case study: Two example fake-news articles involve a fake CNN page, an irrelevant searchable image, a false Wikileaks attribution, and a low-quality composite image.
  • B. Experimental Setup: The data are divided into 80% training, 10% validation, and 10% testing, with every experiment run at least 10 times.
  • B. Experimental Setup: The textual latent branch uses 100-dimensional word2vec embeddings, a 10-word context, (3, 3) filters, and 10 filters.
  • B. Experimental Setup: Table III specifies TI-CNN’s architecture and identifies dropout settings Dα = 0.5 and Dβ = 0.8.

2) Image branch:

The experiments evaluate TI-CNN against baselines and examine parameter sensitivity. Text and image information together outperform the compared baselines, while selected embedding, batch, hidden-layer, dropout, and filter settings affect performance.

  • 2) Image branch:: The visual latent branch reshapes images to 50 × 50 and uses three convolutional layers with 32 filters, ReLU, and (2, 2) max-pooling.
  • 2) Image branch:: The TI-CNN model contains 7,509,980 total parameters and 7,509,176 trainable parameters.
  • C. Experimental Results: TI-CNN with text and image information significantly outperforms all compared baseline methods, whereas image-only information is insufficient.
  • D. Sensitivity Analysis: Embedding dimensions improve precision, recall, and F1-measure from 50 to 100, after which precision and recall decrease.
  • D. Sensitivity Analysis: Batch sizes 32 and 64 are best; F1 rises from batch size 8 to 32 and drops through 128, whose epochs take more than 10 minutes.
  • D. Sensitivity Analysis: A hidden-layer dimension of 128 performs best, while performance drops at 256 because of overfitting.
  • D. Sensitivity Analysis: The model performs well for Dα in [0.1,0.5] and Dβ in [0.1,0.8], with selected dropout probabilities (0.5,0.8).
  • D. Sensitivity Analysis: With filter size (3,3), the model’s F1-measure is 0.92-0.93.

VII. CONCLUSIONS AND FUTURE WORK

TI-CNN combines text and image information with explicit and convolutionally learned latent features to identify fake news. The study reports successful identification on a dataset focused on American presidential-election news, while outlining broader future directions.

  • TI-CNN unifies text and image information with explicit features and latent features learned by convolutional neurons.The model is designed to combine both modalities and can absorb other news features.
  • The experiments use a dataset collected before the presidential election and report successful fake-news identification by TI-CNN.
  • The dataset focuses on American presidential-election news, motivating investigation of other languages through additional French national-election data.
  • Future work includes incorporating social-network structures and user behaviors into fake-news identification.
  • Headline–news-text relevance and image–text relevance, including image-generated captions, are identified as promising research directions.
Loading 1806.00749v3…