Publication Date

6-4-2026

Document Type

Article

Publication Title

Iciss 2025 Proceedings of the 8th International Conference on Information Science and Systems

DOI

10.1145/3803722.3803729

First Page

43

Last Page

49

Abstract

While existing language embeddings achieve high accuracy on benchmark tasks, recent studies suggest they often struggle to distinguish true from false statements. This work argues that new representation learning and aggregation strategies are needed, especially those that adapt to misinformation's evolving patterns and incorporate explainability. As a first step, we explore two directions: (1) aggregation strategies inspired by signal processing, where word embeddings are transformed into spectrogram-like representations for classification to overcome context window limitations, and (2) future extensions using dynamic neural architectures such as Neural ODEs, Liquid Time-Constant Networks (LTCs), Kolmogorov-Arnold Networks (KANs), and Legendre Memory Units (LMUs). Preliminary experiments using spectrogram aggregation with VGG16 on the ISOT dataset yield promising accuracy (86.47%). While results are limited to a single dataset, they demonstrate the feasibility of this approach. We present these ideas as a position paper to stimulate further exploration of representation learning methods tailored to misinformation detection.

Keywords

Kolmogorov-Arnold networks, Legendre Memory Unit, Liquid Time-Constant Networks, Signal Processing, Spectrogram

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Department

Applied Data Science

Download

Share

COinS