Fig 1.
Example of comparison between previous retrieval methods (top) and our proposed KGMP (bottom).
Fig 2.
An overview of KGMP: a KBQA framework leveraging fine-tuned LLMs with a retrieval-first approach before query generation.
Fig 3.
Example of training data. Each piece of training data is divided into two parts, traversal and answer, with the data on the left side representing instructions or database query results, and the data on the right side representing the KGMP outputs.
Table 1.
Comparison of StructGPT+KGMP and ChatKBQA+KGMP with other benchmarks on the WebQSP and CWQ datasets.
Fig 4.
Comparison of the proportion of different error types on WebQSP before and after joining KGMP.
Fig 5.
Comparison of LLM accuracy in predicting multi-hop queries. The results of ChatKBQA+KGMP(without Reranker) only counted data with subgraph sizes not excedding 4096 Tokens, which accounts for, only 53.29% of WebQSP data and 26.51% of CWQ data were counted.
Fig 6.
Comparison of average input Tokens length between ChatKBQA and ChatKBQA+KGMP.
Table 2.
ChatKBQA and ChatKBQA+KGMP FLOPs with different tokens inputs.
Fig 7.
Comparison of LLM inference elapsed time and database GQL execution elapsed time with different memory footprint reductions.