Article: English reading text generation based on optimised variational autoencoder Journal: International Journal of Information and Communication Technology (IJICT) 2025 Vol.26 No.45 pp.52 - 65 Abstract: To address the critical global demand from 1.3 billion English as a Foreign Language learner for personalised reading materials, this study develops a dual-channel regularised variational autoencoder. The model systematically overcomes conventional limitations in readability control and semantic coherence by establishing dynamic mappings between educational linguistic features and latent space, designing a novel readability-driven regularisation loss that integrates lexical complexity, syntactic simplification, and discourse cohesion, and implementing curriculum learning for progressive optimisation. Comprehensive evaluations on the Newsela benchmark corpus demonstrate statistically significant improvements: 7.2% in BLEU-4, 32.8% reduction in readability errors, and 20.6% enhancement in teacher-assessed quality. This framework provides an efficient solution for adaptive learning systems, advancing intelligent generation and scalable deployment of educational resources with high practical utility. Inderscience Publishers - linking academia, business and industry through research

Title: English reading text generation based on optimised variational autoencoder

Authors: Liu Yang

Addresses: Chengdu College of Arts and Sciences, Chengdu, 610000, China

Abstract: To address the critical global demand from 1.3 billion English as a Foreign Language learner for personalised reading materials, this study develops a dual-channel regularised variational autoencoder. The model systematically overcomes conventional limitations in readability control and semantic coherence by establishing dynamic mappings between educational linguistic features and latent space, designing a novel readability-driven regularisation loss that integrates lexical complexity, syntactic simplification, and discourse cohesion, and implementing curriculum learning for progressive optimisation. Comprehensive evaluations on the Newsela benchmark corpus demonstrate statistically significant improvements: 7.2% in BLEU-4, 32.8% reduction in readability errors, and 20.6% enhancement in teacher-assessed quality. This framework provides an efficient solution for adaptive learning systems, advancing intelligent generation and scalable deployment of educational resources with high practical utility.

Keywords: optimised variational autoencoder; English reading text generation; readability control; integration of educational features; Newsela dataset.

DOI: 10.1504/IJICT.2025.150412

International Journal of Information and Communication Technology, 2025 Vol.26 No.45, pp.52 - 65

Received: 19 Aug 2025
Accepted: 06 Nov 2025
Published online: 12 Dec 2025 *

Title: English reading text generation based on optimised variational autoencoder

Keep up-to-date