Track-SQL: Enhancing Generative Language Models with Dual-Extractive Modules for Schema and Context Tracking in Multi-turn Text-to-SQL

Bingfeng Chen; Shaobin Shi; Yongqi Luo; Boyan Xu; Ruichu Cai; Zhifeng Hao

cs.CLMar 6, 2026

Track-SQL: Enhancing Generative Language Models with Dual-Extractive Modules for Schema and Context Tracking in Multi-turn Text-to-SQL

Authors: Bingfeng Chen, Shaobin Shi, Yongqi Luo, Boyan Xu, Ruichu Cai, Zhifeng Hao

Organizations: School of Computer Science, Guangdong University of Technology · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) · Peng Cheng Laboratory · College of Science, Shantou University

Abstract

Generative language models have shown significant potential in single-turn Text-to-SQL. However, their performance does not extend equivalently to multi-turn Text-to-SQL. This is primarily due to generative language models' inadequacy in handling the complexities of context information and dynamic schema linking in multi-turn interactions. In this paper, we propose a framework named Track-SQL, which enhances generative language models with dual-extractive modules designed to track schema and contextual changes in multi-turn Text-to-SQL. Specifically, Track-SQL incorporates a Semantic-enhanced Schema Extractor and a Schema-aware Context Extractor. Experimental results demonstrate that Track-SQL achieves state-of-the-art performance on the SparC and CoSQL datasets. Furthermore, detailed ablation studies reveal that Track-SQL significantly improves execution accuracy in multi-turn interactions by 7.1% and 9.55% on these datasets, respectively. Our implementation will be open-sourced at https://github.com/DMIRLAB-Group/Track-SQL.

Figures & tables

Figure 1: The overall framework of Dual-Extractive Modules for Schema and Context Tracking. The framework trains a schema item classification model and an SQL generator. Based on the former, we construct a Semantic-enhanced Schema Extractor and a Schema-aware Context Extractor . The extraction results from these two extractors are utilized for subsequent training of the SQL generation model. The core idea of Track-SQL is to reduce the gap between the input and the target SQL before entering the multi-turn SQL generation phase by means of dynamic schema linking and context information extraction.

Model	SparC						CoSQL
	QM			IM			QM			IM
	EM	EX	TS	EM	EX	TS	EM	EX	TS	EM	EX	TS
In-Context Learning Approach
ACT-SQL Zhang et al. (2023)	51.0	63.8	56.9	24.4	38.9	29.6	46.0	63.7	55.2	13.3	30.7	21.5
CoE-SQL Zhang et al. (2024a)	56.0	70.3	63.3	36.5	50.5	41.9	52.4	69.6	60.6	23.9	39.6	30.4
Fine-tuned Model

Table 1: Performance of Track-SQL and previous works on the SparC and CoSQL dev set.

	QM		IM
	EX	TS	EX	TS
Track-SQL	75.39	69.16	57.81	50.71
w/o SESE	68.57( $\downarrow{6.82}$ )	62.42( $\downarrow{6.74}$ )	51.42( $\downarrow{6.39}$ )	45.26( $\downarrow{5.45}$ )
w/o ACID	74.56( $\downarrow{0.83}$ )	67.91( $\downarrow{1.25}$ )	56.39( $\downarrow{1.42}$ )	48.57( $\downarrow{2.14}$ )
w/o SACE	72.73( $\downarrow{2.66}$ )	67.16( $\downarrow{2}$ )	51.89( $\downarrow{5.92}$ )	44.78( $\downarrow{5.93}$ )
w/o SACE & SESE	71.40( $\downarrow{3.99}$ )	65.08( $\downarrow{4.08}$ )	50.71( $\downarrow{7.10}$ )	43.36( $\downarrow{7.35}$ )

Table 2: Ablation result on SparC dev set.

	QM		IM
	EX	TS	EX	TS
Track-SQL	70.60	62.26	43.67	32.76
w/o SESE	64.94( $\downarrow{5.66}$ )	58.98( $\downarrow{3.28}$ )	36.86( $\downarrow{6.81}$ )	31.74( $\downarrow{1.02}$ )
w/o ACID	68.22( $\downarrow{2.38}$ )	61.37( $\downarrow{0.89}$ )	39.24( $\downarrow{4.43}$ )	32.42( $\downarrow{0.34}$ )
w/o SACE	68.81( $\downarrow{1.79}$ )	61.46( $\downarrow{0.8}$ )	37.88( $\downarrow{5.79}$ )	29.01( $\downarrow{3.75}$ )
w/o SACE & SESE	66.03( $\downarrow{4.57}$ )	58.88( $\downarrow{3.38}$ )	34.12( $\downarrow{9.55}$ )	26.96( $\downarrow{5.8}$ )

Table 3: Ablation result on CoSQL dev set.

	SparC		CoSQL
	$[email protected]$	$[email protected]$	$[email protected]$	$[email protected]$
Track-SQL	8.50	21.46	12.74	25.90
w/o com-enh	9.57	21.05	13.62	29.93
w/o com&col-enh	9.44	28.18	13.38	27.42

Table 4: Ablation results of SESE. com-enh represents comment-enhanced layer, and col-enh represents column-enhanced layer used in RESDSQL.

	SESE		SQL Generator
	Train	Inference	Train	Inference
SparC	30.9±2.8(h)	0.20(s)	1.5±0.2(h)	1.15(s)
CoSQL	27.5±2.4(h)	0.21(s)	1.6±0.3(h)	1.15(s)

Table 5: Time performance of the Track-SQL framework, Train employs the number of hours required to achieve the optimal model as the performance metric, while Inference bases its evaluation on the time consumed to perform inference on a single batch.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

	Easy (483)			Medium (441)			Hard (145)			Extra (134)
	EM	EX	TS	EM	EX	TS	EM	EX	TS	EM	EX	TS
Track-SQL	81.8	84.9	83.0	64.6	72.8	66.4	44.8	64.8	53.8	35.1	54.5	38.8
w/o SESE	77.6	78.7	77.6	56.5	68.0	59.4	42.8	59.3	46.2	30.6	44.0	35.1
w/o ACID	79.3	85.5	82.8	62.8	72.8	65.8	44.8	63.4	49.7	39.6	53.0	41.0
w/o SACE	79.5	84.1	81.8	63.9	71.9	66.4	39.3	60.0	48.3	32.8	48.5	37.3
w/o SACE & SESE	79.1	80.7	78.5	63.3	71.7	64.2	43.4	59.3	46.9	37.3	50.0	39.6

Appendix

Table 6: Ablation results of Track-SQL on SparC validation set by difficulty. The numbers in brackets () indicate the number of samples.

	Easy (417)			Medium (320)			Hard (163)			Extra (107)
	EM	EX	TS	EM	EX	TS	EM	EX	TS	EM	EX	TS
Track-SQL	80.6	85.4	82.5	51.6	67.5	55.6	36.2	59.5	46.0	24.3	39.3	28.0
w/o SESE	75.1	79.4	76.7	50.0	60.9	51.6	36.2	53.4	46.0	25.2	38.3	31.8
w/o ACID	78.2	83.5	81.1	52.8	65.0	56.2	31.3	55.8	44.2	22.4	37.4	26.2
w/o SACE	78.2	83.5	80.3	50.0	61.6	51.9	35.6	62.0	48.5	29.0	43.9	36.4
w/o SACE & SESE	75.8	77.7	75.8	48.4	60.9	51.6	38.0	66.3	52.8	16.8	35.5	24.3

Appendix

Table 7: Ablation results of Track-SQL on CoSQL validation set by difficulty. The numbers in brackets () indicate the number of samples.

	Turn 1 (422)			Turn 2 (422)			Turn 3 (270)			Turn 4 (88)			Turn > 4 (1)
	EM	EX	TS	EM	EX	TS	EM	EX	TS	EM	EX	TS	EM	EX	TS
Track-SQL	74.2	79.6	76.5	65.9	75.6	67.8	57.8	67.4	59.6	51.1	68.2	60.2	0	100	100
w/o SESE	70.1	75.6	71.8	60.9	69.2	61.4	51.1	60.4	53.0	40.9	58.0	52.3	0	0	0
w/o ACID	72.5	80.6	77.3	64.2	75.1	66.6	56.7	67.8	59.3	54.5	64.8	56.8	0	0	0
w/o SACE	72.3	79.1	75.8	63.0	72.3	64.9	55.9	65.9	59.6	51.1	65.9	60.2	0	0	0
w/o SACE & SESE	71.3	77.0	72.7	63.7	70.9	64.0	58.9	67.0	58.5	51.1	61.4	54.5	0	0	0

Appendix

Table 8: Ablation results of Track-SQL on SparC validation set by turn. The numbers in brackets () indicate the number of samples.

	Turn 1(293)			Turn 2 (285)			Turn 3 (244)			Turn 4 (114)			Turn > 4 (71)
	EM	EX	TS	EM	EX	TS	EM	EX	TS	EM	EX	TS	EM	EX	TS
Track-SQL	64.8	76.5	70.3	58.2	72.3	63.9	56.6	68.0	58.2	55.3	64.9	55.3	40.8	57.7	47.9
w/o SESE	61.4	70.0	64.5	57.9	67.4	62.8	50.8	64.3	54.5	52.6	56.1	51.8	42.3	50.7	47.9
w/o ACID	65.2	73.4	68.6	55.1	69.1	62.8	54.5	66.8	58.6	51.8	62.3	53.5	42.3	57.7	47.9
w/o SACE	66.2	75.8	69.3	52.6	66.7	59.6	57.8	69.3	59.4	52.6	64.0	58.8	42.3	54.9	47.9
w/o SACE & SESE	62.8	71.0	65.2	55.1	68.1	62.5	52.5	65.2	55.7	50.0	58.8	50.0	35.2	52.1	43.7

Appendix

Table 9: Ablation results of Track-SQL on CoSQL validation set by turn. The numbers in brackets () indicate the number of samples.

Figure 2: The results of the ablation study on the Codellama 7B+Track-SQL model on the SparC and CoSQL dev sets (calculated using the multi-turn TS metrics).

Figure 3: The results of the ablation study on the DeepSeek 7B+Track-SQL model on the SparC and CoSQL dev sets (calculated using the multi-turn TS metrics).

Figure 4: The results of the ablation study on the Mistral 7B+Track-SQL model on the SparC and CoSQL dev sets (calculated using the multi-turn TS metrics).

	Table-Weighted Avg			Table-Macro Avg			Column-Weighted Avg			Column-Macro Avg
	Precision	Recall	F1	Precision	Recall	F1	Precision	Recall	F1	Precision	Recall	F1
SESE	96.48	96.43	96.45	95.74	96.42	96.07	97.09	96.93	96.99	90.95	94.34	92.55
w/o com-enh	96.18	96.11	96.12	95.32	96.16	95.72	96.91	96.68	96.76	90.11	94.24	92.03
w/o com-enh & col-enh	96.58	96.57	96.57	96.09	96.31	96.20	96.70	96.67	96.69	91.31	92.05	91.68

Appendix

Table 10: The general precision index score of the semantic-enhanced schema extractor under the SparC dataset

	Table-Weighted Avg			Table-Macro Avg			Column-Weighted Avg			Column-Macro Avg
	Precision	Recall	F1	Precision	Recall	F1	Precision	Recall	F1	Precision	Recall	F1
SESE	95.33	95.28	95.30	94.29	94.87	94.57	96.85	96.71	96.77	89.70	92.65	91.11
w/o com-enh	95.32	95.30	95.31	94.47	94.66	94.57	96.81	96.79	96.80	90.90	91.24	91.07
w/o com-enh & col-enh	94.65	94.39	94.45	92.83	94.64	93.64	96.40	95.98	96.13	86.97	92.78	89.59

Appendix

Table 11: The general precision index score of the semantic-enhanced schema extractor under the CoSQL dataset

Case#1	Question#1 Question#2 Question#3	What is every student’s id? Of those ids, which correspond to those who own cats as pets? List all the other ids.
	Gold	SELECT stuid FROM student EXCEPT SELECT T1.stuid FROM student AS T1 JOIN has_pet AS T2 ON T1.stuid = T2.stuid JOIN pets AS T3 ON T3.petid = T2.petid WHERE T3.pettype = ’cat’
	w/o SACE	SELECT stuid FROM has_pet except SELECT stuid FROM has_pet JOIN pets ON has_pet.petid = pets.petid WHERE pettype = ’cat’
	Track-SQL	SELECT stuid FROM student except SELECT student.stuid FROM student JOIN has_pet ON student.stuid = has_pet.stuid JOIN pets ON has_pet.petid = pets.petid WHERE pets.pettype = ’cat’
	$SQL_{base}$	SELECT student.stuid FROM student JOIN has_pet ON student.stuid = has_pet.stuid JOIN pets ON has_pet.petid = pets.petid WHERE pets.pettype = ’cat’
Case#2	Question#1 Question#2 Question#3	How many car models are produced in total? How many in Germany? How about in Japan?
	Gold	SELECT count(*) FROM MODEL_LIST AS T1 JOIN CAR_MAKERS AS T2 ON T1.Maker = T2.Id JOIN COUNTRIES AS T3 ON T2.Country = T3.CountryId WHERE T3.CountryName = ’japan’

Appendix

Table 12: Case analysis of ablation experiments on the SParC dataset

	SESE(Inference)		SQL Generator(Inference)		SESE(Train)		SQL Generator(Train)
	Total time(s)	Avg Time(s)	Total time(s)	Avg Time(s)	Total time(h)	Avg Time(s)	Total time(h)	Avg Time(s)
SparC	240.348±1.45	0.20±0.00013	1386.98±2.13	1.152±0.025	30.93h±2.78	3.88±0.43	1.45±0.21	11.23±2.31
CoSQL	214.456±2.56	0.21±0.0012	1170.41±1.56	1.15±0.012	27.49h±2.43	3.92±0.29	1.61±0.33	11.43±2.51

Appendix

Table 13: Inference time performance of the Track-SQL framework. Avg Time(s) in terms of training represents the time consumption of a single batch, and Total time(h) represents the time required to obtain the best model. The SQL generator is based on Deepseek 7B

	SESE(Inference)	SQL Generator(Inference)	SESE(Train)	SQL Generator(Train)
Graphics Memory(GB)	2.235	16.477	20.997	62.194

Appendix

Table 14: Memory Costs of Training and Inference in the Track-SQL Framework. In the training stage, the batch size is set to 6, and in the inference stage, the batch size is set to 1.The SQL generator is based on Deepseek 7B

Input for generating annotated descriptions of database columns

You are a database schema designer who specializes in generating concise descriptions for table columns based on the provided column names, types, and sample values. The descriptions should be brief, adjective-noun phrases that reflect the nature of the data in the column.

Table Name: {table_name}

Column: {column_name}

Type: {column_type}

Sample Values: {values}

Appendix

Table 15: This table lists LLM input sequences used for generating annotations for database columns and tables

Format for fine-tuning SQL generation

Input :

You are a SQL query generator that converts multi-turn questions along with associated database schema information into corresponding SQL statements. The multi-turn questions will be concatenated using the ‘

\&

’ symbol, and you should generate the SQL statement that answers the current turn of the question based on the provided database schema.

Each database schema is provided in the following format:

Table name : Column name1, Column name2 Different tables are separated by the ‘|’ symbol, and the order of table names and column names is relevant to the current question; those appearing earlier are more closely related.

Base SQL: {

SQL_{base}

}

Appendix

Table 16: Format for fine-tuning SQL generation

Text-to-SQL

Text-to-SQL translates natural-language questions into executable database queries, linking user intent to schemas, values, and SQL logic.

See all 173 papers

QDA-SQL: Questions Enhanced Dialogue Augmentation for Multi-Turn Text-to-SQL
Jun 15, 2024Yinggang Sun, Ziming Guo, Haining Yu +5Synthetic Data Generation Question Answering
School of Cyberspace Science · Harbin Institute of Technology · Harbin, China +4
Rose-SQL: Role-State Evolution Guided Structured Reasoning for Multi-Turn Text-to-SQL
May 5, 2026Le Zhou, Feng Yao, Fengcai Qiao +3Schema Linking Text-to-SQL
College of Systems Engineering, National University of Defense Technology · College of Advanced Interdisciplinary Studies, National University of Defense Technology · Institute of Automation, Chinese Academy of Sciences +1
DualSQL: Text-to-SQL with Multi-Agent Reinforcement Learning
Sep 16, 2026Shijie Chen, Yu Gan, Yeounoh Chung +6Multi-Agent LLM Systems Text-to-SQL
The Ohio State University · Google LLC

Jun 15, 2024J/K move · Enter open · S save

QDA-SQL: Questions Enhanced Dialogue Augmentation for Multi-Turn Text-to-SQL

Yinggang Sun, Ziming Guo, Haining Yu, Chuanyi Liu+4

School of Cyberspace Science · Harbin Institute of Technology · Harbin, China+4

Synthetic Data Generation Question Answering Text-to-SQL

Open paper AskPDF arXiv

Fine-tuning large language models (LLMs) for specific domain tasks has achieved great success in Text-to-SQL tasks. However, these fine-tuned models often face challenges with multi-turn Text-to-SQL tasks caused by ambiguous or unanswerable questions. It is desired to enhance LLMs to handle multiple types of questions in multi-turn Text-to-SQL tasks. To address this, we propose a novel data augmentation method, called QDA-SQL, which generates multiple types of multi-turn Q&A pairs using LLMs. In QDA-SQL, we introduce a method incorporating validation and correction mechanisms to handle complex multi-turn Text-to-SQL tasks. Experimental results demonstrate that QDA-SQL enables fine-tuned models to exhibit higher performance on SQL statement accuracy and enhances their ability to handle complex, unanswerable questions in multi-turn Text-to-SQL tasks. The generation script and test set are released at https://github.com/mcxiaoxiao/QDA-SQL

Abstract

Text-to-SQL

Explore similar work

QDA-SQL: Questions Enhanced Dialogue Augmentation for Multi-Turn Text-to-SQL

Rose-SQL: Role-State Evolution Guided Structured Reasoning for Multi-Turn Text-to-SQL

DualSQL: Text-to-SQL with Multi-Agent Reinforcement Learning