Input Matters: Evaluating Input Structure's Impact on LLM Summaries of Sports Play-by-Play

TL;DR

This study investigates how different input formats—row-structured, JSON, and unstructured—affect the factual accuracy of LLM-generated NBA game summaries, finding structured inputs significantly reduce errors.

Contribution

It provides empirical evidence that input structure greatly influences factual accuracy in LLM summaries, quantifying error reductions across formats.

Findings

JSON input reduces errors by up to 69%.

Structured inputs significantly lower hallucination rates.

Input structure explains over 80% of error variance.

Abstract

A major concern when deploying LLMs in accuracy-critical domains such as sports reporting is that the generated text may not faithfully reflect the input data. We quantify how input structure affects hallucinations and other factual errors in LLM-generated summaries of NBA play-by-play data, across three formats: row-structured, JSON and unstructured. We manually annotated 3,312 factual errors across 180 game summaries produced by two models, Llama-3.1-70B and Qwen2.5-72B. Input structure has a strong effect: JSON input reduces error rates by 69% for Llama and 65% for Qwen compared to unstructured input, while row-structured input reduces errors by 54% for Llama and 51% for Qwen. A two-way repeated measures ANOVA shows that input structure accounts for over 80% of the variance in error rates, with Tukey HSD post hoc tests confirming statistically significant differences between all input formats.