Portal vs. Generative AI in Digital News: Sentiment Neutralization and Information Fidelity Across Summarization Sources

초록

This study examines whether summarization sources are associated with systematic differences in affective tone and information retention in Korean news. Using 50 articles sampled across five domains, we compare portal-based extractive summaries from Naver IRIS with abstractive summaries generated by ChatGPT-4o under identical input conditions (150 texts in total). Sentiment preservation is measured using changes in KoBERT-based polarity probabilities and sentiment-word ratios, while information retention is assessed using ROUGE-L(F1-Score). Stylistic features, including sentence length and lexical indicators, are additionally analyzed through morphological tokenization. Differences are tested using paired t-tests and two-way ANOVA. The results show that GPT summaries preserve more positive sentiment and attenuate negative sentiment more strongly than portal summaries, with clearer differences in society and lifestyle topics. In contrast, ROUGE-L advantages are modest and not consistently significant, indicating that source choice relates more closely to affective reframing than to lexical overlap. Implications are discussed for topic-sensitive evaluation and lightweight quality control in automated news summarization.

키워드