PAGE: Prompt Augmentation for text Generation Enhancement

In recent years, natural language generative models have shown outstanding performance in text generation tasks. However, when facing specific tasks or particular requirements, they may exhibit poor performance or require adjustments that demand large amounts of additional data. This work introduces PAGE (Prompt Augmentation for text Generation Enhancement), a framework designed to assist these models through the use of simple auxiliary modules. These modules, lightweight models such as classifiers or extractors, provide inferences from the input text. The output of these auxiliaries is then used to construct an enriched input that improves the quality and controllability of the generation. Unlike other generation-assistance approaches, PAGE does not require auxiliary generative models; instead, it proposes a simpler, modular architecture that is easy to adapt to different tasks. This paper presents the proposal, its components and architecture, and reports a proof of concept in the domain of requirements engineering, where an auxiliary module with a classifier is used to improve the quality of software requirements generation.

💡 Research Summary

The paper addresses a practical limitation of modern large language models (LLMs): while they excel at generic text generation, adapting them to specific tasks or strict requirements often demands large amounts of task‑specific data and costly fine‑tuning. To overcome this, the authors propose PAGE (Prompt Augmentation for text Generation Enhancement), a lightweight, modular framework that improves generation quality and controllability by enriching the input prompt with auxiliary information derived from simple, non‑generative models.

PAGE’s architecture consists of three main components. First, an auxiliary module—such as a classifier, entity extractor, or relation detector—processes the raw input and produces a concise inference (e.g., a label, a set of keywords, or a structured tag). These modules are deliberately kept lightweight; they can be trained on small labeled datasets or even reused off‑the‑shelf, avoiding the massive computational overhead of training another generative model. Second, the auxiliary output is transformed into a textual or token‑based “meta‑information” string (for example, “