Context
A company can raise its outlook, cancel an acquisition, or announce an investment without using obviously positive or negative language. Understanding the news requires identifying what happened and who was involved. SENTiVENT grew from Gilles’s doctoral research at Ghent University, with collaboration at VUB. The research addressed fine-grained event extraction and sentiment in economic news. The task brings financial interpretation into the annotation design: participants, event categories, and factual qualifications become explicit objects of analysis instead of remaining implicit in a broad sentiment judgment.
Approach
The work joined data collection, task definition, annotation, model development, and error analysis. A domain-specific event typology helped turn financial news into structured examples. Event mentions were connected to participants and attributes, including distinctions such as negation and modality. The original research publication account gives historical context. This approach preserves more of a sentence’s meaning than a single document-level sentiment label, especially when several companies or events occur together.
Quality controls
Annotation guidelines and an adjudicated reference set made the task inspectable. Difficult disagreements were part of defining the evidence against which systems could be assessed. The independent model evaluation project later brought the public resource into contemporary structured-output workflows. Financial knowledge graphs address a related representation problem in production work, but that commercial delivery is separate from the academic research and its public results.
Outcomes
The research produced peer-reviewed publications and an expert-annotated public dataset. Its contribution is a reusable task definition and reference resource for financial language processing. It should not be read as evidence of a trading system, a forecasting result, or an investment-return claim.
- Published research on company-specific events and financial sentiment.
- Released annotated data supporting supervised extraction and evaluation.
