FUSION
2025 IEEE First International Conference on Innovations in Engineering and Next-Generation Technologies for Sustainability (ICINVENTS)
10.1109/ICINVENTS64613.2025.11401395Structured data
Company filings, financial data
Unstructured text
Market commentary, narrative sources
Fusion layer
My contribution
Large language model
Analysis drawing on multiple sources
The question
Financial information is held in different formats: structured datasets such as prices and company financials on one side, narrative sources such as analyst opinions and earnings-call transcripts on the other. FUSION examines how large language models can guide navigation and integration across those formats for more coherent, context-aware analysis.
My contribution
I built the fusion layer of the system, the part concerned with connecting structured financial data with narrative and other unstructured sources.
What the paper reports
According to the abstract on IEEE Xplore, vectorization combined with re-ranking significantly improved both the relevance and the clarity of the model's outputs.
- A multi-stage pipeline: tailored data ingestion, financial-specific text chunking and a dual index combining keyword and vector search
- Four retrieval strategies tested: keyword search, pure vector retrieval, hybrid methods, and vector retrieval with re-ranking
- A curated dataset of 315 financial documents and 20 benchmark questions, cross-validated with FinQA
- Evaluated with ROUGE scores and expert feedback