NL-Based Database Administration for Handling Heterogeneous
Datasets Using Finetuned LLM
Pradnya Sawant
a
and Kavita Sonawane
b
St. Francis Institute of Technology, Mumbai, India
fi
Keywords:
Database Administrator Queries, Data Definition Language, Data Manipulation Language, Data Control
Language, Large Language Model, NLIDB.
Abstract:
Translating natural language (NL) questions into structured query language (SQL) queries is becoming in-
creasingly important for making databases easier to use and manage. Different large language models (LLMs)
have been used for this translation in recent years. These models are mostly trained and evaluated on datasets
covering a few types of data manipulation language(DML) queries like projection, selection, aggregate func-
tions, joins, etc. However, these datasets failed to contain queries required for Database Administrator(DBA)
operations such as creating and modifying database schema, managing user permissions, etc. This paper
presents an approach to help database administrators (DBAs) and end users interact with databases more in-
tuitively by generating SQL queries from natural language inputs. As no such dataset is publicly available,
we have created a specialized dataset called DBASQL, which includes common DBA operations addressing
data definition language(DDL), data manipulation language(DML), and data control language(DCL) related
natural language questions like creating tables, views, or indexes; inserting values; updating data types or
values; renaming tables or columns; granting or revoking user permissions, paired with their corresponding
SQL queries. For experimentation, we have finetuned Text-to-Text Transfer Transformer (T5) Large on our
customized DBASQL dataset, aiming to improve the accuracy of these translations. Our evaluation shows
that this approach effectively translates NL to SQL that addresses DBA operations, making it easier to han-
dle DDL, DML, and DCL database operations without requiring extensive SQL knowledge. This research
highlights the potential of NLP models to improve the efficiency of natural language to SQL translation by
enabling smarter database interfaces for DBA as well. Also, the proposed DBASQL dataset can be integrated
with any heterogeneous datasets, such as single-domain and cross-domain, for the translation of natural lan-
guage to SQL queries. Hence, covering the border range of SQL queries that can be used by both end users
and database administrators.
1 INTRODUCTION
With data growing quickly, managing databases ef-
ficiently is more important than ever. However, ac-
cessing database information often requires knowing
query languages like SQL, which can be challenging
for non-technical users.
Natural Language Processing (NLP) provides a
solution by enabling non-technical users to ask ques-
tions in natural language rather than knowing com-
plex SQL commands. NLP models convert these
natural language questions into appropriate SQL
queries, making it easier for both Database Admin-
istrators(DBA) and non-technical users to work with
a
https://orcid.org/0000-0003-3982-4077
b
https://orcid.org/0000-0003-0865-6760
databases.
Current NL to SQL systems have shown success
in general-purpose applications, but adapting these
models for DBA-specific operations presents unique
challenges. The systems already available have im-
proved a lot in converting natural language into SQL,
but most datasets mainly focus on a few Data Manip-
ulation Language (DML) queries. On the other hand,
queries for Data Definition Language (DDL), such
as CREATE and ALTER, Data Manipulation Lan-
guage(DML) like CREATE, INSERT, UPDATE, and
Data Control Language (DCL), like GRANT and RE-
VOKE, are not well-represented. DBA-specific nat-
ural language questions involve complex queries to
modify schema structures, update data types and val-
ues, grant and revoke permissions, etc., that require
precise SQL commands tailored to each database’s
56
Sawant, P. and Sonawane, K.
NL-Based Database Administration for Handling Heterogeneous Datasets Using Finetuned LLM.
DOI: 10.5220/0013608700004664
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 3, pages 56-64
ISBN: 978-989-758-763-4
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

configuration. This paper addresses these challenges
by fine-tuning a recent large language model (LLM),
T5-large, specifically for DBA-related SQL queries.
As there is no such dataset available publicly for han-
dling all types of DBA queries, we created a cus-
tomized ”DBASQL” (Database Administrator Struc-
tured Query Language) dataset containing natural lan-
guage queries commonly used by database adminis-
trators. This dataset includes natural language ques-
tions such as “Modify datatype of custid column from
integer to number”, “Allow insert operation on user
table to the manager”, and “Create table employee
having id, name, age”. These questions are paired
with corresponding SQL commands, forming a robust
dataset for training and fine-tuning advanced LLMs to
achieve accurate and reliable translation. This dataset
can be used in combination with other existing het-
erogeneous datasets, like single-domain and cross-
domain to cover the range of SQL operations through
natural language questions. The proposed DBASQL
dataset is publicly available at https://www.kaggle.
com/datasets/pradnyasawant/dbasql. Also, the pro-
posed model can be used to handle complex tasks like
referencing schemas and updating database content
and table schema, and managing user permissions on
database objects.
Through experiments, we show that our model
performs with high accuracy, highlighting the poten-
tial of T5 Large in automating database administra-
tion tasks. By minimizing the need for deep SQL
knowledge, this research helps create a smarter Natu-
ral language Interface to Database (NLIDB) for man-
aging DBA operations as well.
2 RELATED WORK
Converting natural language (NL) queries into SQL
has been studied for a long time. Early methods relied
on rule-based systems, but these were limited because
they were rigid and couldn’t handle a wide variety of
queries or complex database structures cite Kumar14.
Neural network-based models brought big improve-
ments to converting natural language into SQL, espe-
cially with sequence-to-sequence (Seq2Seq) models.
For example, Seq2SQL used reinforcement learning
to create SQL commands, solving issues with query
structure and accuracy (V. Zhong and Socher, 2017).
Later models like SyntaxSQLNet (T. Yu, 2018b), F-
SemtoSql (Q. Li and Zhong, 2020) and TypeSQL
(T. Yu and Radev, 2018) became even more accu-
rate by adding rules and using information about data
types. They also introduced attention mechanisms,
which helped the models focus on the important parts
of the input, making it easier to handle more compli-
cated queries.
Along with NLIDB, NL2VIS(Natural Language to
Visualizations) systems like NL4DV(A. Narechania
and Stasko, 2021), Advisor(C. Liu and Yuan, 2021),
ncNet(Y. Luo and Qin, 2022)are becoming popular as
non-technical users can generate business insights us-
ing charts, graphs, etc. from the underlying database.
There are different benchmarks available for generat-
ing visualizations through natural language questions
(K. Z. Hu and et al., 2019)(Y. Luo and Qin, 2021).
The Transformer architecture completely changed
NLP-to-SQL tasks, with models such as BERT(J. De-
vlin and Toutanova, 2018), RoBERTa(K. Ahkouk
and Ennaji, 2021), XLNet(Q. Li and Zhong, 2020),
T5(Y. Li and Zhang, 2023), and Codex(Trummer,
2022) providing better context understanding, which
improved performance in generating SQL queries.
Fine-tuning models like T5 on SQL datasets made
NL-to-SQL translation more reliable and adaptable,
while OpenAI’s Codex model showed great abil-
ity in generating SQL commands across many dif-
ferent types of queries. (T. Yu, 2018a) intro-
duced a large, cross-domain dataset with complex
multi-table SQL queries, which became an impor-
tant benchmark for evaluating recent large language
models(LLMs)(M. A. Khan and Azam, 2024)(N. T.
K. Le and Teshebaev, 2023)(C. Raffel and Liu, 2020).
This dataset has driven the development of advanced
models that can be generalized effectively. The
Spider dataset is a popular benchmark in NLP-to-
SQL research. It contains natural language ques-
tions linked to complex SQL queries for many dif-
ferent database schemas. Spider is great for testing
how well models can work with new database struc-
tures and handle multi-table joins and nested queries.
However, it mainly covers DML queries and does not
include database administration-related DDL, DML,
and DCL queries. WikiSQL is another well-known
dataset that is simpler to use. It focuses on single-
table queries created from Wikipedia tables. How-
ever, since it only includes SELECT queries and
doesn’t cover the DDL and DCL queries (T. Yu,
2018a).
CoSQL (T. Yu and Su, 2019a)and SParC(T. Yu
and Su, 2019b) are based on the Spider dataset,
but add conversational and multi-step queries, where
users improve their queries gradually. However, like
Spider, they mostly focus on DML queries and do not
cover DDL or DCL queries in detail. While current
NLP-to-SQL datasets provide a strong foundation for
developing models that handle DML queries, there is
a clear gap in datasets representing DBA-related DDL
and DCL queries. Addressing this gap is important
NL-Based Database Administration for Handling Heterogeneous Datasets Using Finetuned LLM
57

for developing a robust NL-to-SQL system that can
handle all database management queries.
Even with these advancements, applying mod-
els to database administration (DBA) operations is
an area of growing research. DBA operations, like
managing permissions, manipulating structures of
schema, changing database contents, and making
complex schema references, which general-purpose
models struggle to handle. Models like RAT-
SQL (B. Wang and Richardson, 2020) and SmBoP
(Z. Zhao and Liang, 2021) focus on schema encoding
and semantic parsing of a few DML queries, but they
do not address queries related to DBA operations.
Hence there is a need to fine-tune LLM for DBA op-
erations, which help better manage these administra-
tion operations, creating a more user-friendly inter-
face for DBAs with limited SQL knowledge. This pa-
per builds on these advancements by fine-tuning the
T5 large model and developing a specialized dataset
for schema and database content manipulation, han-
dling user permissions, etc. Our work demonstrates
the model’s capability in generating accurate SQL
queries for DBA needs, broadening NLP-to-SQL ap-
plications, and addressing gaps in handling special-
ized DBA operations, ultimately supporting the de-
velopment of smarter, more accessible NLIDB.
3 PROPOSED ARCHITECTURE
Figure 1 shows the proposed architecture for NL-
based database administration. In this, a natural lan-
guage question addressing DBA queries is first pre-
processed using tokenization, padding, etc. Then the
NLP module will be responsible for converting this
preprocessed text into the appropriate SQL query.
The highlighted part in the figure shows that we
have fine-tuned the T5 large model for NL processing
specifically for DBA-related queries. This generated
SQL query is executed against the relational database
through which the output can be produced. For Fine-
Tuning the T5 large model, we have not considered
the table schema as input which can reduce the com-
plexity of the model. The finetuning of the T5 model
is explained in Section 3.1.
3.1 Finetuning of T5 Large
The model is initialized with pre-trained weights from
T5-large, and the input is formatted to include the nat-
ural language questions. The output is the target SQL
query. Both input and output are tokenized to fit the
model vocabulary, with careful attention to sequence
truncation and padding. The loss is calculated on
Figure 1: Proposed Architecture for NL-based Database
Administration.
the generated SQL sequences compared to the ground
truth. By optimizing the model using techniques like
learning rate scheduling and validation on held-out
examples, T5-large learns to generalize its transla-
tion capabilities, even for complex SQL queries. For
training the model we have used our own customized
”DBASQL” dataset, which is explained in detail in
Section 3.2.
3.2 Proposed Dataset: DBASQL
Dataset
There is no publicly available dataset specifically de-
signed for NL to SQL translation that comprehen-
sively handles Data Definition. Language(DDL),
Data Manipulation Language(DML), and Data Con-
trol Language(DCL) queries for database adminis-
tration (DBA) requirements. Most existing datasets,
such as Spider, focus on SQL query generation for
various databases but tend to emphasize SELECT
statements (DML) and not the full range of DBA-
related tasks, such as creating, altering, or manag-
ing database structures (DDL). Developing a dataset
tailored to DBA requirements would require incorpo-
rating queries that handle schema modifications, and
other administrative operations. Hence, we have cre-
ated a ”DBASQL” dataset having around 2500 pairs
of natural language and SQL queries covering DDL,
DML, and DCL queries for DBA requirements.
DBASQL Dataset contains natural language ques-
tions and corresponding SQL queries addressing
DBA tasks like schema creation and modifications,
updation of table contents, managing user permis-
sions, etc. as listed in Table 1. These queries are
divided into three different categories: DBA-related
INCOFT 2025 - International Conference on Futuristic Technology
58

DDL, DML, and DCL queries. This dataset also in-
cludes natural language questions that do not explic-
itly mention SQL clause names, making it easier for
users to understand and interact with. By avoiding di-
rect references to SQL clauses, the DBA experience
will be improved. As a result, this dataset helps cre-
ate a more natural and effective interaction between
the DBA and the database system.
The count of queries is sufficient to train any LLM
model which is given in Table 2, covering all DDL,
DML, and DCL queries.
The screenshot of the DBASQL dataset is shown
in Figure 2.
Figure 2: DBASQL Dataset.
4 IMPLEMENTATION DETAILS
AND EXPERIMENTAL SETUP
The following is the experimental setup and the
model parameters used for fine-tuning T5 large.
Customized the DBASQL dataset for DBA-
related queries and DataLoader classes tailored for
the NLP to SQL task, which includes specific padding
and tokenization requirements that can impact model
performance. Training includes inference functions
to infer query types from SQL queries, which are spe-
cific to the NL-to-SQL task. Finetuning of the T5
large model uses the following architecture parame-
ters.
• Learning Rate: 1e-4 (0.0001),
• Number of Epochs: 20,
• Optimizer: AdamW optimizer,
• Scheduler: Learning rate scheduler with linear
warm-up and linear decay,
• Batch Size: 4,
• Loss Function: The Cross-Entropy loss function
is computed based on the output of the T5 model
during training,
• Early Stopping: Patience of 5 epochs is used for
early stopping. If validation loss does not improve
for 5 consecutive epochs, the training stops.
The performance of finetuned and modified T5 is
measured based on the following two parameters.
Exact Match Accuracy (EMA): It compares the
expected query with the predicted SQL query to
check whether they match each other. It is concerned
with the syntactical correctness of the generated
SQL. The EMA is calculated using Equation 1 as
follows:
ACCema = Nema/n (1)
Logical Accuracy(LA) : It checks whether the gener-
ated SQL retrieves the correct data semantically, even
if the SQL structure differs. The LA is calculated
using Equation 2 as follows:
ACCla = Nla/n (2)
Where n is the number of examples.
Nema- Number of predicted queries that are syntac-
tically similar to the expected SQL query.
Nla- Number of predicted queries that are logically
similar to the expected SQL query.
5 RESULTS AND DISCUSSION
We have proposed the novel DBASQL dataset for
Database administration queries with the intention
of handling the full breadth of SQL queries effec-
tively. Also, experimenting with the proposed T5
model on a variety of NL questions will justify the
strength of the proposed model. The training and val-
idation loss graph for the fine-tuned T5 model is pre-
sented in Figure 3. The comparison of exact match
accuracy(EMA) and logical accuracy(LA) for DDL,
DML, and DCL queries is shown in figures Figure 4,
Figure 5, and Figure 6, respectively. The proposed
model is observed to be better at predicting DBA
queries even without using the schema of the under-
lying tables.
The logical accuracy for DDL queries like CRE-
ATE TABLE and RENAME TABLE is more than the
exact match accuracy as the primary keys are assigned
automatically by the model as shown in Figure 4.
Also, the renaming of the tables is done dynamically
by the model.
NL-Based Database Administration for Handling Heterogeneous Datasets Using Finetuned LLM
59

Table 1: Description of Queries (DBA Related).
Category Query Type Description
DDL
CREATE Used to create database objects like tables, indexes, and views.
ALTER
Modifies the structure of an existing database object. Common opera-
tions include:
ADD: Adds a new column or primary key to an existing table.
DROP COLUMN: Removes an existing column from a table.
RENAME COLUMN: Renames a column in a table.
RENAME TABLE: Renames a table.
MODIFY COLUMN: Changes the data type or size of an existing column.
DROP Deletes an existing database object such as a table or an index.
TRUNCATE Removes all rows from a table but retains its structure.
DML
SELECT Retrieves data from one or more tables.
INSERT Adds new rows of data to a table.
UPDATE Modifies existing data in a table.
DELETE FROM Removes rows from a table.
DCL
GRANT Allowing users or roles to perform specific actions like SELECT, IN-
SERT, DELETE, and UPDATE on database objects.
REVOKE Restricting users from accessing or modifying database objects after
permissions are revoked.
Table 2: Count of Queries(DBA Related).
DDL Query Type Count DML Query Type Count DCL Query Type Count
CREATE 250 INSERT 176 GRANT 94
ALTER MODIFY 63 UPDATE 164 REVOKE 80
ALTER RENAME TABLE 80 DELETE FROM 150
ALTER RENAME COLUMN 86
ALTER MODIFY 85
ALTER DROP COLUMN 50
ALTER ADD 105
DROP 47
TRUNCATE 60
DESCRIBE 30
OTHER 25
Figure 3: Loss Graph for Finetuned T5 Large.
The LA of DML queries like INSERT is more as
alteration of the datatype and size are automatically
done by the model as presented in Figure 5.
As shown in Figure 6, the EMA and LA are both
the same for all natural language questions address-
ing DCL queries like granting and revoking user per-
missions like select, update, and delete on any table.
Hence, the proposed model is observed to be bet-
ter at predicting DBA queries even without using the
schema of underlying tables.
The sample results addressing DDL, DML, and DCL
queries are presented in Table 3, Table 4, and Table
5, respectively. For the majority of test data, the pre-
dicted query is the same as the expected query.
Also, for some test data, the model predicts the
SQL clause without explicit mention in natural lan-
guage questions, as shown in bold. As presented in
Table 3, the proposed model creates the table with-
out explicit mentions about the datatype and size of
the table fields in the input natural language question.
This makes the proposed model more robust. Also,
for questions about creating the view, the model as-
signs the view names at run time.
Table 4 presents sample results for DML-related
INCOFT 2025 - International Conference on Futuristic Technology
60

Table 3: Sample Test Results on Finetuned T5 Large for DDL Queries.
Input Text Expected Output Finetuned T5 Predicted
Output
Remark
Create a table named
’Emp’ with columns:
emp id (INT), name
(VARCHAR(100)), email
(VARCHAR(255)), dept id
(INT), salary (DECI-
MAL(12, 2)).
CREATE TABLE Emp
(emp id INT, name VAR-
CHAR(100), email VAR-
CHAR(255), dept id INT,
salary DECIMAL(12, 2));
CREATE TABLE Emp
(emp id INT, name VAR-
CHAR(100), email VAR-
CHAR(255), dept id INT,
salary DECIMAL(12, 2));
Predicted is same as
expected output.
Make table named
’Staffnew’ with columns:
staff id, name, email.
CREATE TABLE Staffnew
(staff id INT, name VAR-
CHAR(100), email VAR-
CHAR(100));
CREATE TABLE
Staffnew (staff id INT
PRIMARY KEY, name
VARCHAR(255), email
VARCHAR(255));
Model is generat-
ing SQL queries
without their ex-
plicit mentions in
NL question.
Add an index on the
course id column in the
Course table.
CREATE INDEX
idx course id ON
Course(course id);
CREATE INDEX
idx course id ON
Course(course id);
Predicted is same as
expected output.
Formulate a view that
includes all professor’s
emails.
CREATE VIEW Profes-
sorEmailList AS SELECT
email FROM Professors;
CREATE VIEW Profes-
sorEmailListView AS SE-
LECT professor id, email
FROM Professors;
The view names
are assigned dy-
namically.
Rename the column ’title’
to ’title paper’ in the table
’ResearchPapers’
ALTER TABLE Re-
searchPapers RENAME
COLUMN title TO ti-
tle paper;
ALTER TABLE Re-
searchPapers RENAME
COLUMN title TO ti-
tle paper;
Predicted is same as
expected output.
Change the name of the
table ’ResearchPapers’ to
’ResearchPapers1’
ALTER TABLE Research-
Papers RENAME TO Re-
searchPapers1;
ALTER TABLE Research-
Papers RENAME TO Re-
searchPapers1;
Model is generat-
ing SQL queries
without their ex-
plicit mentions in
NL question.
Modify the datatype of
column ’enrollmentdate’ to
DATE in the table ’Enroll-
ments’
ALTER TABLE Enroll-
ments MODIFY COLUMN
enrollmentdate DATE;
ALTER TABLE Enroll-
ments MODIFY COLUMN
enrollmentdate DATE;
Predicted is same as
expected output.
Delete the column ’email’
from the table ’Students’.
ALTER TABLE Students
DROP COLUMN email;
ALTER TABLE Students
DROP COLUMN email;
Model is generat-
ing SQL queries
without their ex-
plicit mentions in
NL question.
Add a new column ’gradu-
ation year’ to the Students
table
ALTER TABLE Students
ADD graduation year INT;
ALTER TABLE Students
ADD graduation year INT;
Predicted is same as
expected output.
Remove the Grades table,
Drop the Grades table
DROP TABLE IF EXISTS
Grades;
DROP TABLE IF EXISTS
Grades;
Model is generat-
ing SQL queries
without their ex-
plicit mentions in
NL question
Delete all contents of the
Students table
TRUNCATE TABLE Stu-
dents;
TRUNCATE TABLE Stu-
dents;
Model is generat-
ing SQL queries
without their ex-
plicit mentions in
NL question
Add a primary key con-
straint to the ’student id’
column in the ’Students’ ta-
ble.
ALTER TABLE Students
ADD PRIMARY KEY
(student id);
ALTER TABLE Students
ADD PRIMARY KEY
(student id);
Predicted is same as
expected output.
Describe the Grades table DESCRIBE Grades; DESCRIBE Grades; Predicted is the
same as expected
output.
NL-Based Database Administration for Handling Heterogeneous Datasets Using Finetuned LLM
61

CREATE
ALTER MODIFY
ALTER RENAME TABLE
ALTER RENAME COLUMN
ALTER MODIFY
ALTER DROP COLUMN
ALTER ADD
DROP
TRUNCATE
DESCRIBE
70
80
90
100
DDL Operations
EMA (%)
70
80
90
100
LA (%)
Figure 4: Comparison of EMA and LA for DDL queries.
INSERT
UPDATE
DELETE
70
80
90
100
DML Operations.
EMA (%)
70
80
90
100
LA (%)
Figure 5: Comparison of EMA and LA for DML queries.
natural language questions. Here, the model predicts
the correct SQL clauses without explicitly mentioning
the SQL clause in the input natural language question,
e.g. for inserting the values in the database table with-
out an explicit mention of the ”INSERT” word in a
natural language question, the model is still predict-
ing the clause precisely.
Sample results of DCL queries are presented in
Table 5. As presented in Table 5 the model is granting
and revoking user permissions accurately on database
tables.
Hence, the proposed model was found to achieve
an EMA 94.74% and LA 97.2% with early stopping
epoch number 14 using the proposed fine-tuned T5
large model without using the table schema as input.
6 CONCLUSION AND FUTURE
SCOPE
This research work aimed to translate DBA-related
natural language questions into SQL queries. We have
proposed and validated fine-tuned T5 on a diversified
customized DBASQL dataset, and we could achieve
the exact match accuracy 94.74% and logical accu-
racy 97.2% without using the table schema as input.
The proposed model can be effectively used in vary-
ing natural language users as well as DBA needs.
The same is justified and validated by the contribu-
tion in the form of a new proposed DBASQL dataset
covering all varieties of natural language questions
without explicitly mentioning the SQL clause. This
dataset can be easily combined with the available
heterogeneous datasets to cover the full breadth of
SQL operations effectively for performing natural
language to SQL translations.
In the future, we can continue to improve the effi-
ciency of LLMs to handle more complex, ambiguous,
and multi-turn queries.
GRANT
REVOKE
70
80
90
100
DCL Operations
EMA (%)
70
80
90
100
LA (%)
Figure 6: Comparison of EMA and LA for DCL queries.
INCOFT 2025 - International Conference on Futuristic Technology
62

Table 4: Sample Test Results on Finetuned T5 Large for DML Queries.
Input Text Expected Output Finetuned T5 Predicted
Output
Remark
Delete a course having id
21
DELETE FROM Courses
WHERE course id = 21;
DELETE FROM Courses
WHERE course id = 21;
Predicted is same as
expected output.
Change the title of a re-
search paper to ’Advance-
ments for Robotics’, having
paper id 30.
UPDATE ResearchPapers
SET title = ’Advancements
for Robotics’ WHERE
paper id = 30;
UPDATE ResearchPapers
SET title = ’Advancements
for Robotics’ WHERE pa-
per id = 30;
Model is generat-
ing SQL queries
without their ex-
plicit mentions in
NL question.
Insert a student with ID
11, name ’John’, email
’[e-mail]’, de-
partment id 1, advisor id
101, gpa 3.5.
INSERT INTO Students
(student id, name, email,
department id, advisor id,
gpa) VALUES (11, ’John’,
’[e-mail]’, 1,
101, 3.5);
INSERT INTO Students
(student id, name, email,
department id, advisor id,
gpa) VALUES (11, ’John’,
’[e-mail]’, 1,
101, 3.5);
Predicted is the
same as the ex-
pected output.
Add a student with ID 3,
name ’Ronan’, email ’ro-
[e-mail]’, advi-
sor id 102, gpa 9.5, depart-
ment id 4.
”INSERT INTO Students
(student id, name, email,
department id, advisor id,
gpa) VALUES (3, ’Ronan’,
’[e-mail]’, 4,
102, 9.5);”
INSERT INTO Stu-
dents (student id, name,
email, advisor id, gpa,
department id) VAL-
UES (3, ’Ronan’, ’ro-
[e-mail]’, 102,
9.5, 4);
Model is generat-
ing SQL queries
without their ex-
plicit mentions in
NL question.
Table 5: Sample Test Results on Finetuned T5 Large for DCL Queries.
Input Text Expected Output Finetuned T5 Predicted
Output
Remark
Enable ALL permissions
on the ’inventory’ table for
role ’stock manager’.
GRANT ALL PRIVI-
LEGES ON inventory TO
stock manager;
GRANT ALL PRIVI-
LEGES ON inventory TO
stock manager;
Predicted is the
same as the ex-
pected output
without explicit
mentions in the NL
question.
Allow user ’vidya’ to SE-
LECT data from the ’logs’
table.
GRANT SELECT ON logs
TO vidya;
GRANT SELECT ON logs
TO vidya;
Predicted is same as
expected output.
Take away UPDATE privi-
leges on the ’inventory’ ta-
ble from role ’manager’.
REVOKE UPDATE ON in-
ventory FROM manager;
REVOKE UPDATE ON
inventory FROM manager;
Predicted is the
same as the ex-
pected output
without explicit
mentions in the NL
question.
Remove all access on the
’departments’ table for the
’admin’ role.
REVOKE ALL PRIVI-
LEGES ON departments
FROM admin;
REVOKE ALL PRIVI-
LEGES ON departments
FROM admin;
Predicted is the
same as the ex-
pected output
without explicit
mentions in the NL
question.
NL-Based Database Administration for Handling Heterogeneous Datasets Using Finetuned LLM
63

REFERENCES
A. Narechania, A. S. and Stasko, J. (2021). Nl4dv: A toolkit
for generating analytic specifications for data visual-
ization from natural language queries. IEEE Transac-
tions on Visualization and Computer Graphics.
B. Wang, R. Shin, X. L. Y. P. and Richardson, M. (2020).
Rat-sql: Relation-aware schema encoding and link-
ing for text-to-sql parsers. In Proceedings of the 58th
Annual Meeting of the Association for Computational
Linguistics (ACL), pages 7567–7578, Online.
C. Liu, Y. Han, R. J. and Yuan, X. (April 2021). Advisor:
Automatic visualization answer for natural-language
question on tabular data. In 14th IEEE Pacific Visu-
alization Symposium, PacificVis, page 11–20, Tianjin,
China.
C. Raffel, N. Shazeer, A. R. K. L. S. N. M. M. Y. Z. W. L.
and Liu, P. J. (2020). Exploring the limits of transfer
learning with a unified text-to-text transformer. Jour-
nal of Machine Learning Research.
J. Devlin, M.-W. Chang, K. L. and Toutanova, K. (2018).
Bert: Pre-training of deep bidirectional transformers
for language understanding. arXiv:1810.04805.
K. Ahkouk, M. M. and Ennaji, M. (2021). Data agnos-
tic roberta-based natural language to sql query gener-
ation. In IEEE 6th International Conference for Con-
vergence in Technology (I2CT).
K. Z. Hu, S. N. S. G. and et al. (2019). Viznet: Towards
a large-scale visualization learning and benchmarking
repository. In Proceedings of CHI Conference on Hu-
man Factors in Computing Systems.
M. A. Khan, M. S. H. Mukta, K. F. N. M. F. S. S. M. M. J.
M. J. A. M. E. A. and Azam, S. (2024). A review on
large language models: Architectures, applications,
taxonomies, open issues and challenges. IEEE Access.
N. T. K. Le, N. Hadiprodjo, H. E.-A. A. K. and Teshe-
baev, A. (2023). Recent large language models in nlp.
In 22nd International Symposium on Communications
and Information Technologies (ISCIT), Sydney, Aus-
tralia.
Q. Li, L. Li, Q. L. and Zhong, J. (2020). A comprehensive
exploration on spider with fuzzy decision text-to-sql
model. IEEE Transactions on Industrial Informatics,
16(4):April.
T. Yu, R. Zhang, K. Y. D. S. W. I. Z. and Su, M. Y.
(2019a). Cosql: A conversational text-to-sql challenge
towards cross-domain natural language interfaces to
databases. In Proceedings of the 2019 Conference on
Empirical Methods in Natural Language Processing
and the 9th International Joint Conference on Nat-
ural Language Processing (EMNLP-IJCNLP), pages
1962–1979, Hong Kong, China.
T. Yu, R. Zhang, K. Y. M. Y. D. W. Z. L. J. M. I. L. Q. Y. S.
R. Z. Z. D. R. (2018a). Spider: A large-scale human-
labeled dataset for complex and cross-domain seman-
tic parsing and text-to-sql task. In Proc. Conf. Empiri-
cal Methods Natural Lang. Process., page 3911–3921.
T. Yu, M. Yasunaga, K. Y. R. Z. D. W. Z. L. D. R. (2018b).
Syntaxsqlnet: Syntax tree networks for complex and
cross-domain text-to-sql task. In Proc. Conf. Empiri-
cal Methods Natural Lang. Proces., page 1653–1663.
T. Yu, Z. Lin, L. T. D. S. W. I. Z. and Su, M. Y. (2019b).
Sparc: Cross-domain semantic parsing in context. In
Proceedings of the 57th Annual Meeting of the As-
sociation for Computational Linguistics (ACL), pages
4511–4523, Florence, Italy.
T. Yu, Z. Li, Z. Z. R. Z. and Radev, D. (2018). Type-
sql: Knowledge-based type-aware neural text-to-sql
generation. In 16th Annual Conference of the North
American Chapter of the Association for Computa-
tional Linguistics, New Orleans.
Trummer, I. (2022). Codexdb: Generating code for pro-
cessing sql queries using gpt-3 codex. In Proceedings
of the International Conference on Artificial Intelli-
gence.
V. Zhong, C. X. and Socher, R. (2017). Seq2sql: Gener-
ating structured queries from natural language using
reinforcement learning. arXiv:1709.00103.
Y. Li, Z. Su, Y. L. H. Z. S. W. W. W. and Zhang, Y. (2023).
T5-sr: A unified seq-to-seq decoding strategy for se-
mantic parsing. In IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP).
Y. Luo, N. Tang, G. L. C. C. W. L. and Qin, X. (2021). Syn-
thesizing natural language to visualization (nl2vis)
benchmarks from nl2sql benchmark. In Interna-
tional Conference on Management of Data, page June
20–25. ACM.
Y. Luo, N. Tang, G. L. J. T. C. C. and Qin, X. (2022). Natu-
ral language to visualization by neural machine trans-
lation. IEEE Transactions on Visualization and Com-
puter Graphics, 28(1):January.
Z. Zhao, Y. Yang, X. H. and Liang, J. (2021). Smbop:
Semi-autoregressive bottom-up semantic parsing. In
Proceedings of the 59th Annual Meeting of the As-
sociation for Computational Linguistics (ACL), pages
5903–5914, Online.
INCOFT 2025 - International Conference on Futuristic Technology
64