This page organizes papers related to artificial intelligence published around the world. This page is summarized using Google Gemini and is operated on a non-profit basis. The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.
Pandora is a novel Unified Structured Knowledge Reasoning (USKR) framework that integrates diverse structured knowledge sources (tables, databases, and knowledge graphs) to answer natural language questions. Existing USKR methods rely on specific task strategies or user-defined representations, making it difficult to leverage knowledge transfer across different SKR tasks or link with prior knowledge from LLM, resulting in limited performance. Pandora leverages Python's Pandas API to build a unified knowledge representation for alignment with LLM pretraining. It uses LLM to generate text-based inference steps and executable Python code for each question, and facilitates knowledge transfer by leveraging a memory of training examples covering a variety of SKR tasks.
Takeaways, Limitations
•
Takeaways:
◦
Improved alignment with LLM pre-training by leveraging integrated knowledge representation using the Pandas API.
◦
LLM-based text inference and executable Python code generation provide flexibility for a variety of SKR tasks.
◦
We achieved performance improvements through knowledge transfer utilizing training example memory.
◦
In four benchmarks, it outperforms existing integration frameworks and competes effectively with specific workflows.
•
Limitations:
◦
High dependency on the Pandas API may limit extensibility to other programming languages or APIs.
◦
Since it depends on the performance of LLM, limitations of LLM may also affect the performance of Pandora.
◦
A balanced composition of training examples across different SKR tasks is important, as data bias can impact performance.
