LESCP

This is the code of our AAAI 2025 oral paper "Filling Memory Gaps: Enhancing Continual Semantic Parsing via SQL Syntax Variance-Guided LLMs Without Real Data Replay".

---

Environment

Main dependencies

• Python 3.8+
• transformers
• vllm
• nltk
• pandas
• tqdm
• sqlparse

You can also install our environment via running this code:

pip install -r requirements.txt

Prepare

Model

Put Mixtral-8x7B-Instruct-v0.1 into the 'model' directory. You can download the model here.

Put codet5-base into the 'model' directory. You can download the model here.

Data

Put spider and spider_task_stream into the 'data' directory. You can get the original Spider dataset here.

Put combine1_perm_1 and wikisql into the 'data' directory. You can get the original Spider dataset here.

Running Code

Make multi-tables consisting of tables from WikiSQL

cd combine_prompt_make
python prompt_make.py

Run the embedding and cluster code to preprocess

Corresponding to sections "Domain Information Elimination" and "Component Bias Analysis" in the paper:

cd embedding_cluster
python step1_generate_VALUES.py
python step2_synthesis.py

Synthetic data (Intra-Task Memory)

Corresponding to section "Intra-Task Memory Reinforcement" in the paper:

cd intra
python fstep1_My_generate_VALUES.py
python fstep2_Make_data_concanated.py
python istep1_paraphrase.py
python istep2_paraphrased_process.py
python istep3_remove_alias.py

Attention: when execute istep, you should process every task using --task_id

Synthetic data (Inter-Task Memory)

Corresponding to section "Inter-Task Memory Completion" in the paper.

Generate data for single table:

cd inter
python ostep0_inference_makesingle.py --start_task_id 0 --end_task_id 6  --batch_size 32 --inference_type V1 --input_max_length 2400 --order 0
python ostep1_outdomain_process_makesingle.py
python ostep2_outdomain_judge_makesingle.py
python ostep3_judged_process_makesingle.py

Generate data for multi table:

cd inter
python ostep0_inference.py --start_task_id 0 --end_task_id 6 --batch_size 32 --inference_type V1 --input_max_length 2400 --order 0
python ostep1_outdomain_process.py
python ostep2_outdomain_judge.py
python ostep3_judged_process.py

Merge data

At last, running the code to merge all the data generated:

cd data_merge
python merge.py

Training

bash run.sh

---

Usage Notes

• Each script is independent. Please check the comments in each file for specific usage and parameter settings.
• It is recommended to set up the required Python environment and dependencies before running any scripts.

Acknowledgements

We gratefully acknowledge the following works that inspired and contributed to this project:

1. SSCL-Text2SQL
2. C3
3. Spider
4. CodeTask-CL

---

Citation

@article{Liu_Zhang_Song_Zhang_Yang_2025, 
    title={Filling Memory Gaps: Enhancing Continual Semantic Parsing via SQL Syntax Variance-Guided LLMs Without Real Data Replay}, 
    volume={39}, 
    url={https://ojs.aaai.org/index.php/AAAI/article/view/34644}, 
    DOI={10.1609/aaai.v39i23.34644}, 
    abstractNote={Continual Semantic Parsing (CSP) aims to train parsers to convert natural language questions into SQL across tasks with limited annotated examples, adapting to dynamically updated databases in real-world scenarios. Previous studies mitigate this challenge by replaying historical data or employing parameter-efficient tuning (PET), but they often violate data privacy or rely on ideal continual learning settings. To address these issues, we propose a new Large Language Model (LLM)-Enhanced Continuous Semantic Parsing method, named LECSP, which alleviates forgetting while encouraging generalization, without requiring real data replay or ideal settings. Specifically, it first analyzes the commonalities and differences between tasks from the SQL syntax perspective to guide LLMs in reconstructing key memories and improving memory accuracy through calibration. Then, it uses a task-aware dual-teacher distillation framework to promote the accumulation and transfer of knowledge during sequential training. Experimental results on two CSP benchmarks show that our method significantly outperforms existing methods, even those utilizing data replay or ideal settings. Additionally, we achieve generalization performance beyond upper limits, better adapting to unseen tasks.}, 
    number={23}, 
    journal={Proceedings of the AAAI Conference on Artificial Intelligence}, 
    author={Liu, Ruiheng and Zhang, Jinyu and Song, Yanqi and Zhang, Yu and Yang, Bailong}, 
    year={2025}, 
    month={Apr.}, 
    pages={24641-24649} 
    }