Model Description
Graphix-T5 is a graph-aware semi-pretrained text-to-text PLM specifically designed to improve multi-hop reasoning for the complex text-to-SQL task. This novel architecture enhances the structural encoding capabilities of the T5 model while preserving its powerful contextual encoding ability. The experimental results demonstrate the effectiveness of GRAPHIX-T5 and underscore the importance of incorporating structural information in text-to-text PLMs for tackling intricate text-to-SQL challenges. The smaller gap in performance between the dev and test sets indicates the stronger generalization capability of Graphix-T5.
Training Data
Graphix-3B is trained based on SPIDER, a cross-domain text-to-SQL benchmark. And it's evaluated in vanilla SPIDER dev, test, and other variants: SPIDER-SYN, SPIDER-DK, SPIDER-REALISTIC without additional training. This model will continue to be fine-tuned on more complex text-to-SQL data, i.e. BIRD to deal with harder but more real applications
To Begin With
You can use this model directly with a pipeline for text generation. This example generates a different sequence each time it's run:
from transformers import AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained("patrickNLP/Graphix-3B")
model = AutoModel.from_pretrained("patrickNLP/Graphix-3B")
Performance
Graphix-3B w/ Picard maintains state-of-the-art (SOTA) semantic parsing capabilities, as demonstrated by its performance on the SPIDER
leaderboard. Its only submission achieves 74.0% on EM and 77.6% on EX in the testing dataset.
Please see Graphix Official Implementation
for details.
Reference
Graphix-T5: Mixing Pre-Trained Transformers with Graph-Aware Layers for Text-to-SQL Parsing
Can LLM Already Serve as A Database Interface? A BIg Bench for Large-Scale Database Grounded Text-to-SQLs
Spider: A Large-Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Task
PICARD: Parsing Incrementally for Constrained Auto-Regressive Decoding from Language Models
Citation
@misc{li2023graphixt5,
title={Graphix-T5: Mixing Pre-Trained Transformers with Graph-Aware Layers for Text-to-SQL Parsing},
author={Jinyang Li and Binyuan Hui and Reynold Cheng and Bowen Qin and Chenhao Ma and Nan Huo and Fei Huang and Wenyu Du and Luo Si and Yongbin Li},
year={2023},
eprint={2301.07507},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
- Downloads last month
- 16