Worked within the SenticNet Group under the guidance of Zihao Huang on source-grounded financial question answering, studying how different Retrieval-Augmented Generation (RAG) architectures retrieve and use financial evidence for complex reasoning.
The work combined multimodal financial information from earnings-call transcripts and investor presentations with structured knowledge representations and a controlled evaluation framework.
Implemented and compared multiple retrieval paradigms for source-grounded financial question answering, including dense vector, graph-enhanced, Knowledge Graph-based, and higher-order hypergraph-based approaches.
Worked with complementary financial evidence from earnings-call transcripts and investor presentation slides, preserving links between source material, extracted information, and provenance.
Developed a structured evaluation workflow to examine retrieval quality, contextual relevance, evidence sufficiency, answer quality, semantic agreement, and engineering behavior across the evaluated RAG architectures. Comparative evaluation was performed across retrieval quality, contextual quality, evidence sufficiency, answer quality, semantic agreement, and engineering behavior.
Financial events were extracted and structured to support graph-based and higher-order representations, providing a common foundation for Knowledge Graph and Hypergraph-based retrieval. Financial events combine entities, financial metrics, temporal information, numerical values, and source evidence.