Document Type : Original/Review Paper

Authors

• omid hajipoor ¹
• Ahmad Nickabadi ¹
• Mohammad Mehdi Homayounpour ²

¹ Amirkabir University of Technology Department of Computer Engineering, Tehran, Iran

² amirkabir university

Abstract

Conditional text generation is crucial in natural language processing but often struggles with the high computational costs of Large Language Models (LLMs) and training instability in Generative Adversarial Networks (GANs). In this paper, we introduce the Guidance-Transformer Generative Adversarial Network (GTGAN), a framework that generates text by working within a smooth, continuous hidden (latent) space rather than outputting discrete words directly, bypassing common optimization bottleneck errors. To enforce precise control, GTGAN utilizes a dual-guidance classifier mechanism to organize this latent space and direct the generation process. Additionally, we employ a global discriminator (to maintain overall sentence coherence) and a local discriminator (to verify small word groups), which together prevent repetitive text (mode collapse) and training errors (exposure bias). Tested on Yelp and Amazon reviews, GTGAN improves text quality (BLEU-5) by up to 18.9% and category retention accuracy by 8.78% over preceding GANs. Notably, GTGAN achieves these results while using under 50M parameters (less than 1% of the size of standard 7B parameter LLMs), demonstrating a highly efficient, controllable, and lightweight solution for conditional text generation.

Keywords

• Conditional Text Generation
• Large Language Models (LLMs)
• Generative Pretrained Transformer (GPT)
• Llama
• Generative Adversarial Networks (GANs)

Main Subjects

• Document and Text Processing