Measuring sentiment from financial news is crucial for economics and finance. Existing dictionary-based methods have limitations in capturing context, negation, and semantic structure. This paper introduces a framework for creating daily news mood indices using transformer-based language models. The study evaluates if these models better represent sentiment than traditional dictionary approaches. They analyzed 143,755 financial news articles from Factiva, classifying sentiment at the sentence level using FinBERT. These predictions were then aggregated into article-level and daily sentiment measures. The developed indices were compared against existing benchmark measures. A significant contribution is the validation of these new measures against human judgments. An incentivized exercise involved 444 participants rating a subsample of financial news articles. These consensus human ratings served as an external benchmark for automated measures. Transformer-based measures demonstrated stronger agreement with human judgments and better performance in distinguishing sentiment categories. The findings indicate that transformer models, by incorporating contextual information, produce sentiment measures more aligned with human assessments of financial news.