A novel pipeline framework for multi-hop complex KGQA task.
This framework mainly contains two modules, answering_filtering_module and relational_chain_reasoning_module
And this two module should be trained independently, at reference step, question and KG load into answering_filtering_module ad inputs, then get the top-K candidates ,and retrieval these candidates` relational chain in KG, and let relational_chain_reasoning_module provide the final answer to USERS.
overall pipeline architecture See model
answering_filtering_module See Module1
relational_chain_reasoning_module See Module2
Statistical Performance Comparsion:
Experimental results on three subsets of MetaQA. The first group of results was taken from papers on recent methods. The values are
reported using hits@1.
| Model | 1-hop MetaQA | 2-hop MetaQA | 3-hop MetaQA ||
| :-----| ----: | :----: ||
| EmbedKGQA | 97.5 | 98.8 | 94.8 ||
| SRN | 97.0 | 95.1 | 75.2 ||
| KVMem | 96.2 | 82.7 | 48.9 ||
| GraftNet | 97.0 | 94.8 | 77.7 ||
| PullNet | 97.0 | 99.9 | 91.4 ||
| Our Model | 98.3 | 99.7 | 97.9 ||
Experiment results compared with SOTA methods on WebQuestionsSP-tiny test set. All QA pairs in WebQuestionsSP-tiny are 2-hop relational questions.
| Model | WebQuestionsSP-tiny hit@1 ||
| EmbedKGQA | 66.6 ||
| SRN | - ||
| KVMem | 46.7 ||
| GraftNet | 66.4 ||
| PullNet | 68.1 ||
| Our Model | 70.4 ||
Hope you enjoy it !!! Arxiv link: https://arxiv.org/abs/2110.12679