مجلة آمنير للدراسات والبحوث
Volume 3, Numéro 1, Pages 79-89
2026-07-26

Defining And Measuring Quality In Automatically Generated Language: A Systematic Review

Authors : Tahri Abderrahmen .

Abstract

The rapid proliferation of natural language generation (NLG) systems has intensified the need for robust frameworks to define and assess the quality of their content. This systematic review synthesizes research published between 2018 and 2023 to examine how quality is conceptualized and operationalized across key NLG tasks, including machine translation, summarization, dialogue systems and creative text generation. A total of 87 peer-reviewed studies were analysed, revealing three dominant evaluation paradigms: human-centred approaches (e.g., fluency, adequacy, and coherence), automated metrics (e.g., BLEU and BERTScore), and task-specific composite measures. The findings highlight persistent discrepancies between human judgments and automated scores, an overreliance on surface-level metrics, and a lack of standardized taxonomies that incorporate domain-specific and ethical dimensions, such as factual consistency and bias. In response, this review proposes a multilayered evaluation framework that integrates quantitative metrics with qualitative human assessment. This study provides a foundation for developing more reliable and comprehensive quality assurance practices for applied natural language processing systems.

Keywords

Natural language generation ; Evaluation metrics ; Content quality assessment ; Human–AI collaboration ; Systematic review