Source-linked AI summary
We used Neural Networks to Detect Clickbaits: You won't believe what happened Next!
Ankesh Anand, Tanmoy Chakraborty, Noseong Park
TL;DR
Existing clickbait-detection methods rely on hand-crafted features and linguistic resources. The paper applies recurrent neural networks to distributed word and character representations, reporting improvements over the best-performing cited baseline.
Problem
Existing clickbait-detection methods rely on hand-crafted features and language-specific NLP tools and lexicons, making adaptation challenging and feature engineering time-consuming.
Source: §1 · p. 2Method
The approach uses distributed word and character embeddings as features for recurrent neural network models that detect clickbait.
Source: §1 · p. 2Results
More than 5% improvement in accuracy and F1-score and more than 2% in ROC-AUC over the best-performing Chakraborty et al. SVM baseline.
Source: §3 · p. 6Takeaways & Limitations
The authors report that their models improve on state-of-the-art clickbait detection without relying on heavy feature engineering.
Source: §4 · p. 6
Abstract
from arXiv · showhide
Online content publishers often use catchy headlines for their articles in order to attract users to their websites. These headlines, popularly known as clickbaits, exploit a user's curiosity gap and lure them to click on links that often disappoint them. Existing methods for automatically detecting clickbaits rely on heavy feature engineering and domain knowledge. Here, we introduce a neural network architecture based on Recurrent Neural Networks for detecting clickbaits. Our model relies on distributed word representations learned from a large unannotated corpora, and character embeddings learned via Convolutional Neural Networks. Experimental results on a dataset of news headlines show that our model outperforms existing techniques for clickbait detection with an accuracy of 0.98 with F1-score of 0.98 and ROC-AUC of 0.99.
1 Introduction
The paper proposes recurrent neural networks using distributed word and character representations to detect clickbait, addressing existing methods’ reliance on hand-crafted features. On news headlines, the authors report significant improvement over state-of-the-art results across accuracy, F1-score, and ROC-AUC.
1 Introduction
- Clickbait headlines use provocative, catchy content to attract clicks while withholding article information and sometimes exaggerating what readers should expect.Source: §1 · p. 1
- Clickbaits exploit curiosity gaps and contribute low-quality content to social media streams while violating journalistic ethics.The authors connect the headline to the knowledge gap and the click to the urge to fill it.Source: §1 · pp. 1–2
- Prior clickbait-detection work was limited, and existing methods relied on hand-crafted features and language-specific resources that are challenging to adapt and time-consuming to engineer.Source: §1 · p. 2
- The authors propose distributed word embeddings for lexical and semantic features alongside character embeddings for orthographic and morphological features.Source: §1 · p. 2
- They explore LSTM, GRU, and standard RNN architectures to capture context beyond individual words or fixed-size word windows.Source: §1 · p. 2
- The proposed model significantly improves on state-of-the-art results in accuracy, F1-score, and ROC-AUC on a dataset of clickbait and non-clickbait headlines.The passage does not give the metric values.Source: §1 · p. 2
2 Model
The model combines distributed word and character-level embeddings with a bidirectional recurrent network to classify headlines. It explores RNN, LSTM, and GRU architectures for capturing contextual information.
2 Model
- The model concatenates word and character embeddings, feeds them to a bidirectional RNN, then classifies its fixed-size representation with a sigmoid output node.The architecture diagram depicts this pipeline.Source: §2 · pp. 2–3, Figure 1 · p. 3
2.1 Features
- Distributed word embeddings capture semantic and syntactic properties, while character-level embeddings capture orthographic and morphological features and can represent out-of-vocabulary words.The word vectors are pre-trained 300-dimensional word2vec embeddings; character representations are learned with three 1-dimensional CNN layers and max pooling.Source: §2.1 · p. 3
2.2 Recurrent Neural Network Models
- RNNs maintain sequential history, and the paper uses bidirectional variants to capture context in both forward and backward directions.A standard RNN computes its internal state from the current input and previous hidden state, with a nonlinear function such as tanh.Source: §2.2 · pp. 3–4
- LSTM gating addresses difficulty preserving long-range dependencies in standard RNNs, while GRUs use gates to modulate information without a separate memory cell.The LSTM uses input, forget, and output gates; the GRU uses update and reset gates and a candidate activation.Source: §2.2 · p. 4
3 Evaluation
The evaluation compares recurrent architectures and embedding combinations on a balanced headline dataset, then benchmarks the strongest model against existing methods. BiLSTM(CE+WE) exceeds the best-performing baseline on accuracy, F1-score, and ROC-AUC.
3 Evaluation
- The evaluation uses 15,000 headlines, evenly split between clickbait and non-clickbait, with 10-fold cross-validation.Non-clickbait headlines come from Wikinews; clickbait headlines come from five named online publishers.Source: §3 · p. 5
- Training uses mini-batch gradient descent with batch size 64, ADAM, binary cross-entropy, and dropout at rate 0.3; character embeddings are updated for the task.The implementation uses Keras with a TensorFlow backend.Source: §3 · p. 5
- BiLSTM(CE+WE) slightly outperforms the other models; BiLSTM generally performs better than BiGRU and BiRNN, and combined word and character embeddings give the best results within each architecture.Word embeddings alone are the next-best feature set.Source: §3 · p. 5
- More than 5% improvement in accuracy and F1-score, and more than 2% in ROC-AUC, over Chakraborty et al. (SVM).The comparison is reported for BiLSTM(CE+WE) against the best-performing baseline.Source: §3 · pp. 5–6, Table 2 · p. 6
4 Conclusion
The paper presents three bidirectional recurrent neural network variants for clickbait detection using distributed word and character-level embeddings. The authors report significant improvement over the state of the art without heavy feature engineering, and propose visualizing internal states and incorporating attention in future work.
4 Conclusion
- The paper introduces three bidirectional recurrent neural network variants using distributed word and character-level embeddings for clickbait detection.Source: §4 · p. 6
- The models achieve significant improvement over the state of the art without relying on heavy feature engineering.Source: §4 · p. 6
- Future work will qualitatively visualize the models’ internal states and incorporate an attention mechanism.Source: §4 · p. 6