Balancing diversity with strict EGRA constraint adherence in Arabic story generation

Determine whether automatic generation of Arabic Early Grade Reading Assessment (EGRA)-aligned stories can achieve consistent narrative diversity while simultaneously maintaining strict adherence to EGRA constraints, including controlled passage length (≤60 words), exactly one proper noun, present-tense usage, a clear narrative structure with introduction, middle dilemma, and resolution, gender balance, avoidance of stereotypes, and age-appropriate, locally suitable vocabulary.

Background

A central challenge highlighted in the paper is the trade-off between increasing output diversity and maintaining the strict pedagogical and linguistic constraints mandated by EGRA for early-grade Arabic reading assessments. Output-level stochastic decoding methods often boost diversity at the expense of constraint adherence, risking assessment validity.

Prior work has explored inspiration-based prompting and iterative self-refinement to enhance diversity and quality, yet a general solution that reliably preserves EGRA constraints while ensuring diverse narratives, especially in Arabic, has not been established. The authors frame this as an unresolved issue motivating their investigation of internal noise steering methods.