Before Generation, Align it! A Novel and Effective Strategy for Mitigating Hallucinations in Text-to-SQL Generation

This is the official repository for the paper "Before Generation, Align it! A Novel and Effective Strategy for Mitigating Hallucinations in Text-to-SQL Generation", which has been accepted to ACL Findings 2024.

Overview

In this work, we first study and conclude primary hallucinations presented in the current text-to-SQL frameworks and attribute them to two main categories: schema-based hallucinations and logic-based hallucinations.

We then introduce a novel strategy, Task Alignment (TA), designed to mitigate hallucinations at each stage. We further propose a text-to-SQL framework named TA-SQL, which consists of a Task-Aligned Schema Linking (TASL) module and a Task-Aligned LOGical synthesis (TALOG) module. This repo contains all code for the implementation and evaluation of TA-SQL, utilizing GPT-4 as the backend, on BIRD dev set as stated in our paper.

Environment Setup

• Use the following command to configure local environment:

 $ conda create -n tasql python=3.8.16
 $ conda activate tasql
 $ pip3 install -r requirements.txt

• Set environment variables for the Azure OpenAI API or modify your own OpenAI config in ./src/llm.py:

export OPENAI_API_BASE="YOUR_OPENAI_API_BASE"
export OPENAI_API_VERSION="YOUR_OPENAI_API_VERSION"
export OPENAI_API_KEY="YOUR_OPENAI_API_KEY"

Data Preparation

The BIRD dev set used in the paper could be directly downloaded from the BIRD Leaderboard. Please put the unzipped folder under ./data/ after downloading sucessfully. The dataset under the ./data/dev_databases/ should contain the following resources:

• database: Each database folder should contain
  • database_description: The csv files are manufactured to describe database schema and its values for models to explore or references.
  • sqlite: The database contents in BIRD.
• dev_tables.json: The file contains related information for each database, including db_id, table_names_originial, etc,.
• dev.json: The file contains text-to-SQL paired with the oracle knowledge evidence.

Collect Results

To run this project, you could directly execute the command line by following the instruction (You may need to adjust parameters and paths with your preference):

 $ sh ./run.sh

In this script, as stated in Appendix A.1, we first generate a succinct description for each column to serve as a component that displays database schema information. Then, we run the TA-SQL workflow to generate SQL for each natural language question. Outputs of these two steps are stored in ./outputs/.

column_meaning.json

./outputs/column_meaning.json stores a pythonic dictionary, where the key is composed of database_id|table_name|column_name, and the value is key information about each column and their value summarized from raw CSVs. You could directly use this file or generate succinct descriptions by runing ./src/conclude_meaning.py.

In order to keep the same evaluation setting, we also generate column_meaning.json for train set (in progress) and testing set. If you want to use it in your BIRD submission, please mention in the submission email.

Evaluation

To run the evaluation, you need to put the ground-truth sql file dev_gold.sql in the ./data/ first. Then you could evaluate the results using the following command line:

 $ sh ./run_evaluation.sh

EX Performance on BIRD dev

MODEL	SIM.	MOD.	CHALL.	TOTAL
Closed-Source LLM
GPT4	54.35	34.64	31.70	46.35
+TA-SQL	63.14	48.60	36.11	56.19
GPT4-turbo	59.35	38.92	27.78	50.19
+TA-SQL	60.54	40.86	38.19	52.48
Claude	51.34	30.07	23.24	42.47
+TA-SQL	56.97	39.78	27.78	48.89
ChatGPT	47.60	22.44	18.31	37.22
+TA-SQL	51.57	33.76	25.69	43.74
Open-Source weaker LLM
DeepSeek	51.68	29.03	18.06	41.66
+TA-SQL	53.41	32.04	19.44	43.74
CodeLlama	34.81	15.48	11.11	26.73
+TA-SQL	37.30	13.33	11.11	27.57

Performance on BIRD Mini-Dev

To facilitate efficient and cost-effective development cycles, we provide a Lite version of BIRD dev dataset called Mini-Dev. This dataset results from community feedback, leading to the compilation of 500 high-quality text2sql pairs derived from the original dev dataset. To further enhance the practicality of the BIRD system, we make the Mini-Dev dataset available not only in SQLite, but also in MySQL and PostgreSQL. We also introduce two new evaluation metrics for the Mini-Dev dataset: the Reward-based Valid Efficiency Score (R-VES) and the Soft F1-Score to present a more comprehensive and stable evaluation for the performance. More details about Mini-Dev could be found on Mini-Dev Introduction Page.

Here, we present the performance of TA on Mini-Dev

MODEL	EX	R-VES	Soft F1
Closed-Source LLM
GPT4-turbo	45.80	44.79	50.08
+TA-SQL	58.00	56.44	62.40
GPT35-turbo	38.00	37.33	41.84
+TA-SQL	41.60	40.59	44.25
Open-Source LLM
Llama3-70b	40.08	39.02	44.38
+TA-SQL	42.80	41.37	46.66

Project Structure

├─data/
|  ├─dev_databases # data of BIRD dev databases
├─src/
|  ├─conclude_meaning.py  # Generate suffcient descriptions for columns in the dataset
|  ├─llm.py               # OpenAI API call
|  ├─modules.py           # Modules of TA-SQL framework
|  ├─prompt_bank.py       # prompt templates used in each module of TA-SQL framework
|  ├─utils.py             # utils functions
├─evaluation/
|  ├─evaluation.py        # EX evaluation script
|  ├─evaluation_ves.py    # VES evaluation script
├─outputs/
├─README.md
├─requirements.txt
├─run_evaluation.sh
├─run.py 
├─run.sh 

Citation

Please cite the repo if you think our work is helpful to you.

@inproceedings{qu-etal-2024-generation,
    title = "Before Generation, Align it! A Novel and Effective Strategy for Mitigating Hallucinations in Text-to-{SQL} Generation",
    author = "Qu, Ge  and
      Li, Jinyang  and
      Li, Bowen  and
      Qin, Bowen  and
      Huo, Nan  and
      Ma, Chenhao  and
      Cheng, Reynold",
    editor = "Ku, Lun-Wei  and
      Martins, Andre  and
      Srikumar, Vivek",
    booktitle = "Findings of the Association for Computational Linguistics ACL 2024",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand and virtual meeting",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-acl.324",
    doi = "10.18653/v1/2024.findings-acl.324",
    pages = "5456--5471",
    abstract = "Large Language Models (LLMs) driven by In-Context Learning (ICL) have significantly improved the performance of text-to-SQL. Previous methods generally employ a two-stage reasoning framework, namely 1) schema linking and 2) logical synthesis, making the framework not only effective but also interpretable. Despite these advancements, the inherent bad nature of the generalization of LLMs often results in hallucinations, which limits the full potential of LLMs. In this work, we first identify and categorize the common types of hallucinations at each stage in text-to-SQL. We then introduce a novel strategy, Task Alignment (TA), designed to mitigate hallucinations at each stage. TA encourages LLMs to take advantage of experiences from similar tasks rather than starting the tasks from scratch. This can help LLMs reduce the burden of generalization, thereby mitigating hallucinations effectively. We further propose TA-SQL, a text-to-SQL framework based on this strategy. The experimental results and comprehensive analysis demonstrate the effectiveness and robustness of our framework. Specifically, it enhances the performance of the GPT-4 baseline by 21.23{\%} relatively on BIRD dev and it yields significant improvements across six models and four mainstream, complex text-to-SQL benchmarks.",
}