Corpus-driven analysis using Convolutional Neural Networks with Multi-Head Attention
Laurent Vanni, S Haris, Damon Mayaffre · Amsterdam studies in the theory and history of linguistic science. Series 4, Current issues in linguistic theory · 2025
Abstract This paper addresses challenges associated with the interpretability of deep learning classification models, particularly relevant for researchers in the humanities. A proposed methodological framework integrates corpus-driven approaches and interpretable deep learning architectures, resulting in the development of the Multi-channel Convolutional Transformer (MCT). This model effectively balances performance and interpretability, as demonstrated through a case study in political science examining discursive conditions surrounding immigration as an electoral issue in 21st-century French politics. The MCT emerges as a potent tool for text analysis, offering practical advantages for researchers in various domains.