A hierarchical review of multi-level sentiment analysis in financial texts: Advances in contextual embeddings and deep learning

People depend more than ever on unstructured financial texts, news stories, earnings calls, regulatory filings, and even social media to make decisions. That’s pushed sentiment analysis into the spotlight. But old-school sentiment analysis just flattens everything out and misses the layers and nuance in financial language. Now, with advances in deep learning and contextual embeddings, researchers can dig into sentiment at different levels of text, capturing way more detail. This review pulls together what’s new in hierarchical and multi-level sentiment analysis for financial texts, zeroing in on contextual embedding models and the latest deep learning setups. We systematically looked through the last five years of peer-reviewed studies, collecting papers from top databases like Scopus, Web of Science, IEEE Xplore, ACM Digital Library, and SpringerLink. We only included studies that focused on hierarchical sentiment modeling, contextual embeddings, and deep learning within financial text analysis. Then we broke down the data, comparing models, embeddings, datasets, and how people measured results. What stands out? There’s a big move away from lexicon-based and classic machine learning approaches transformer-based contextual embeddings now lead the way. Finance-focused models like FinBERT and Financial-RoBERTa are especially strong. Hierarchical and hybrid architectures almost always beat their flat counterparts, mainly because they keep track of sentiment connections across different levels and clear up context much better. Still, there are hurdles: adapting models to new domains, making them explainable, scaling them up, and the fact that there aren’t standard benchmarks to compare results across different levels. This review points out why hierarchical modeling and contextual embeddings matter for financial sentiment analysis. It also calls out what’s missing and where researchers should go next like building unified, explainable, and efficient sentiment systems. Bottom line: these insights give future researchers and anyone building AI-driven financial tools a solid place to start.