LSW-Net: A Linguistic-Signal Wavelet Network for Fake-News Detection in News Text

The proliferation of fake news makes preserving information integrity on the Internet essential, yet most detectors learn opaque, model-specific representations. We introduce LSW-Net, which represents a document as a one-dimensional linguistic signal whose amplitude at each token is an explicit function of a part-of-speech prior, sentiment polarity and TF-IDF salience. This signal is decomposed with a discrete wavelet transform (DWT; Daubechies db4, three levels) into interpretable multi-resolution sub-bands that a bidirectional LSTM reads; in parallel, a denoising-autoencoder latent and a contextual sentence embedding are fused with the BiLSTM state at the classification head. On the GonzaloA/fake_news benchmark LSW-Net reaches 96.65% accuracy, 96.92% F1 and 0.994 AUC; on this lexically separable corpus it is competitive with, but does not exceed, a TF-IDF + SVM baseline (98.2% accuracy), and an ablation shows the wavelet channel adds no accuracy there. To test where the representation matters, we add the PHEME rumour corpus under a leave-one-event-out protocol: on unseen events the wavelet channel improves macro-F1 in six of eight events, and LSW-Net degrades far less when moving from in-distribution to unseen-event evaluation (-29 macro-F1 points) than the TF-IDF + SVM baseline (-37 points). LSW-Net thus trades a small amount of in-distribution accuracy for a representation that is both inspectable and more robust to topic shift. Code is publicly available at https://github.com/Ghaith-Alawady/lsw-net.