IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 11, No. 4, December 2022, pp. 1261~1269

ISSN: 2252-8938, DOI: 10.11591/ijai.v11.i4.pp1261-1269  1261

Automated dynamic schema generation using knowledge graph

Priyank Kumar Singh1, Sami Ur Rehman1, Darshan J.1, Shobha G.2, Deepamala N.2
1
Department of Computer Science and Engineering, R.V College of Engineering, Bengaluru, India
2
Department of Computer Science and Engineering, Faculty of Computer Science and Engineering, R.V College of Engineering,
Bengaluru, India

Article Info ABSTRACT

Article history: On the internet where the number of database developers is increasing with
the availability of huge data to be stored and queried. Establishing relations
Received Sep 11, 2021 between various schemas and helping the developers by filtering,
Revised Jun 26, 2022 prioritizing, and suggesting relevant schema is a requirement.
Accepted Jul 20, 2022 Recommendation system plays an important role in searching through a
large volume of dynamically generated schemas to provide database
developers with personalized schemas and services. Although many methods
Keywords: are already available to solve problems using machine learning, they require
more time and data to learn. These problems can be solved using knowledge
Database graphs (KG). This paper investigates building knowledge graphs to
Graph neural network recommend schemas.
Knowledge graph
Natural language processing
Schema This is an open access article under the CC BY-SA license.

Corresponding Author:
Priyank Kumar Singh
Department of Computer Science and Engineering
R.V College of Engineering
Banglore-560059, Karnataka, India
Email: [email protected]

1. INTRODUCTION
The database is a very integral part of any software application and relational database is the most
popular type of database system available. Designing a database is a very tedious and critical part of any
software engineering process. A large share of the total time is spent on designing the database. To design a
database, professional database designers are required to avoid any problems in the future. So, in this ever-
advancing world of technology, there is tremendous growth in the field of software engineering and hence
there is great demand for efficient data storage and processing techniques. Hence under this big hood of
software development, there is an implicit need of automating the entire process of designing a database
which is a key support of any software project.
A huge amount of data is generated and stored in a structured and unstructured format. Database
schemas are a structured method of storing the data. Even though the data is structured, there exist multiple
schemas and tables which are related, but the relations cannot be established by simple databases. One of the
methods to represent schemas as a combined knowledge source is to design a knowledge graph (KG). The
term “KG” was known to be used in writing since no less than 1973 in the paper presented by Edward W.
Schneider. Google created its own knowledge graph [1], which was announced in 2012, is a more
sophisticated embodiment of the knowledge graph ever seen in the technology world [2]. Following then,
knowledge graphs have become more important in terms of research and application. As a consequence of
the research that took place, quite a few KG products were developed. Cyc [3], Freebase [4], Google
knowledge vault [5], DBpedia [6], YAGO [7], NELL [8], PROSPERA [9], Microsoft’s Probase [10] are
among the prominent KGs ever created. Many real-world applications like Semantic Web search engine [11],

Journal homepage: http://ijai.iaescore.com

1262  ISSN: 2252-8938

named entity recognition and disambiguation [12], [13], Extracting Information [14], [15], Recommendation
systems [16], and Siri [17], Apple’s personal assistant, Watson [18] from IBM also make use of generic
knowledge graphs. Even in the field of medicines, KG called as Medical KG [19] is gaining lot of attention
recently.
Graph neural network (GNN) is also one of the techniques that are used in graph applications, the
input method in this also works the same as that of the neural network system, but the nodes are converted
into graph embeddings which is an important step involved in any GNN model. The main reason behind
using GNN would be to be able to predict relationships(edges) between completely unrelated nodes,
technically known as Edge prediction [20], [21]. Zhao et al. [22], discusses KG’s construction and
architecture. It also gives an outline of KG's construction and looks at the approaches and difficulties that
were encountered during the process. Building KG requires Semantic relatedness which plays a key role in
linking a word with an existing word in our knowledge graph. It is a quantitative measure of how two words
or concepts are related under a context, without any discrepancy of the syntactical form of the word.
Conceptually speaking, semantic relatedness is based on functional relations like hyponymic, hypernymic,
metronymic [23]. At the point when restricted to hyponymy/hypernymy relations, the action evaluates
semantic similarities between two words or concepts.
This paper discusses a method to build a robust knowledge graph and query the same to build a
database by suggesting tables and columns in an automated way. The proposed system is designed to be
based on a knowledge graph, where it creates a smaller graph for each of the nuclear schemas and then
merges this graph with the global KG intelligently based on the similarity between the nodes by comparing
their word vectors. So, the proposed algorithm not only suggests database schemas but also serves as an
efficient tool to store information while minimizing redundancy.

2. METHOD
This section describes the approach that was applied to produce a knowledge graph of different
entities. The workflow of the pipeline is shown in Figure 1. In short, the framework includes the following
steps:
− Schema to knowledge graph: convert each of the schemas to the knowledge graph
− Executing the Algorithm: which exploits the link prediction using multiple natural language processing
and machine learning tools.
− Storing the data: in which the final knowledge graph is stored in the TinkerPop database.
− Interaction with the System: in which multiple questions are asked to the user to get complete insights.
− Query refining and output: in which keywords are extracted using natural language processing tools and
the final schema is given to the user.

Figure 1. Shows the flowchart of the AI-based models and experimental methods applied

Int J Artif Intell, Vol. 11, No. 4, December 2022: 1261-1269

Int J Artif Intell ISSN: 2252-8938  1263

2.1. Schema to knowledge graph

Given the multiple schemas, the first step is to convert the given schema into a knowledge graph.
Firstly, each table is converted to a node [24], table attributes are added as the properties of node and relation
as an edge for the graph and a complete knowledge graph is built. Table attributes are converted to properties
of nodes so that it is easier to keep the properties of the original schema intact.

2.2. Executing the algorithm

This is the core step of the complete architecture. For the algorithm as shown in Table 1 to work at
least, two knowledge graphs are required. For each node n1 in graph G1, iterate over each node n2 of graph
G2. Using the synset similarity [25] function (from NLTK library) the percentage of similarity is calculated
between the node n1 and n2, if the similarity is 100% (this mean both the nodes are the same) both the nodes
are merged to form a single node otherwise if the similarity percentage is more than the threshold value, a
new node n is created i.e., the hypernym of both nodes. Once the new node is created it is then connected to
the node n1 and n2 by adding new edges from node n1 to n and from n2 to n.

Table 1. Algorithm to merge two schemas based on similarity

Algorithm 1: Join Graph
Data: G1: graph1, G2: graph2
Result: single merged graph
for each node n1ϵ nodes(G1) do
for each node n2ϵ nodes(G2) do
if similarity(n1,n2) == 1 then
combine both nodes to one;
else
if similarity (n1,n2) ≥ thresholdvalue then
create-new-node(n);
add-edge(n1,n);
add-edge(n2,n);
end
end
end
end

2.3. Storing the data

Once the final knowledge graph is generated, it is then stored in a graph database i.e., Tinker Pop.
After storing the knowledge graph in the database, the server is started and now the knowledge graph is ready
to query. As of now, Gremlin query language [26] has been used to query the graph.

2.4. Interaction with the system

Users interact with the system through a chatbot and a sequence of questions is asked to get a
complete insight into what a user wants. Based on the insights the system will provide the possible nodes
(tables) to choose from. Once the nodes (table) are selected, the final schema is given to the user with all the
relevant attributes and relations.

2.5. Query refining and output

Once the queries are received from the user, it is then processed to remove the unwanted words like
articles, propositions, quantifiers. After cleaning the query, the final keyword is extracted, now this keyword
will be used to extract the final sub-knowledge graph. After extracting the sub-knowledge graph, it is then
converted into the schema so that it can be easily used by the user.
Using an example, the complete proposed system is illustrated in the following steps. Two schemas
are considered i.e., School and student schema,
Step1: Converting schema to knowledge graph
The first step converts the schema into a knowledge graph as shown in Figures 2-3, where node
name is the table name and the attributes are stored in the properties of a node. Edges of the graphs contain
the properties which also denotes the relationships between the tables. Now these knowledge graphs will be
fed to the algorithm to generate the final knowledge graph.
Step2: Executing the algorithm on knowledge graphs
In this step, previously generated knowledge graphs are fed to the algorithm that predicts the new
edges and nodes based on the similarity between two nodes (table) and if two nodes are similar then their

Automated dynamic schema generation using knowledge graph (Priyank Kumar Singh)
1264  ISSN: 2252-8938

properties (attributes) are merged to get a single node. The final knowledge is generated as shown in
Figure 4. The final knowledge graph contains much more context than the individual knowledge graphs.

Figure 2. School schema to a knowledge graph

Figure 3. Student schema to a knowledge graph

Figure 4. Merging student and school schema and result is a combined knowledge graph

Int J Artif Intell, Vol. 11, No. 4, December 2022: 1261-1269

Int J Artif Intell ISSN: 2252-8938  1265

Step3: Query and preprocessing

In this step, the user-query is processed using the natural language processing modules that split the
sentences into verbs, nouns and determiners. The irrelevant words (like: articles, preposition) are removed, to
get only the essential words. Out of these useful words the keyword is extracted out as shown in Figure 5.

Figure 5. Grade query processing

Step4: Output schema

This is the final step of the process, where the keywords extracted in step3 are used to extract the
subgraph over the final knowledge graph that is generated in step2. Depending on the level of details a user
wants, the knowledge graph is generated as shown in Figure 6. Based on the output graph, the schema is
generated with all the attributes.

Figure 6. Output grade schema with details up to level 2 when queried over the merged graph

3. RESULTS AND DISCUSSION

This section briefly introduces some illustrative queries and results representing typical schema. The
dataset that was used for experimenting has more than 10 schemas from different domains (like: school,
employee, sales, registration, and library). The graph in Figure 7 shows the final graph obtained by merging
all the nuclear graphs according to the algorithm. As we keep on adding the more schemas this graph keeps
on growing.
Input Query:
For the query “Give the schema for students” the query was first preprocessed and the keyword was
extracted i.e. “Student” shown in Figure 8. After the keyword is extracted, it is then used to extract out the
subgraph from the main knowledge graph as shown in Figure 7. From the output schema shown in Figure 9 it
can be seen that the student node has more than 10 directly connected nodes, this shows how much-enriched
details are provided by the schema.

Automated dynamic schema generation using knowledge graph (Priyank Kumar Singh)
1266  ISSN: 2252-8938

Figure 7. Knowledge graph after merging 11 schemas

Figure 8. Query for student schema

Figure 9. Student schema when queried over the merged graph

Int J Artif Intell, Vol. 11, No. 4, December 2022: 1261-1269

Int J Artif Intell ISSN: 2252-8938  1267

For the query “Give the schema for payments” as shown in Figure 10, the query is processed to
remove the unwanted words, and the keyword extracted is “payments”. Using this keyword subgraph is
extracted from the merged graph as shown in Figure 7. The subgraph generated has details up to depth 2 as
shown in Figure 11.

Figure 10. Query for payments

Figure 11. Payments schema when queried over the merged graph

4. CONCLUSION
Technological advancements in software industries have led to an increase in the usage of data
storage systems, especially, relational database systems are being extensively used. Designing a database
schema takes a considerable share of total development time in any project. Hence, there is a need for
automating the entire process of designing a database schema. In this paper, various approaches were
discussed to solve the problem and it was evident that the knowledge graph-based approach suits best for the
purpose. Here, usage of KG was investigated to dynamically generate schemas that satisfy all the constraints
discussed in the paper and also reduces the redundancy of the data in the knowledge graph. The existing KG
built from various schemas is queried to suggest schemas to the user. The proposed model was tested with
multiple schemas to generate the global knowledge graph and it was queried and verified manually. This
method helps the user in developing relationships between given schemas to draw the KG and query the
same to suggest a schema.

REFERENCES
[1] E. W. Schneider, “Course modularization applied: the interface system and its implications for sequence control and data
analysis,” Association for the Development of Instructional Systems (ADIS), pp. 10–73, 1973, doi: 10.1037/e436252004-001.
[2] L. Ehrlinger and W. Wöß, “Towards a definition of knowledge graphs,” in Conference: Joint Proceedings of the Posters and
Demos Track of 12th International Conference on Semantic Systems - SEMANTiCS2016 and 1st International Workshop on
Semantic Change & Evolving Semantics (SuCCESS16), 2016, vol. 1695, pp. 1–4.
[3] B. L. Douglas, “CYC: A Large-Scale Investment in Knowledge Infrastructure,” Communications of the ACM, vol. 38, no. 11,
pp. 33–38, 1995, doi: 10.1145/219717.219745.
[4] K. Bollacker, C. Evans, P. Paritosh, T. Sturge, and J. Taylor, “Freebase: A collaboratively created graph database for structuring
human knowledge,” Proceedings of the ACM SIGMOD International Conference on Management of Data, pp. 1247–1249, 2008,
doi: 10.1145/1376616.1376746.
[5] X. Dong et al., “Knowledge vault: A web-scale approach to probabilistic knowledge fusion,” Proceedings of the ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, pp. 601–610, 2014, doi: 10.1145/2623330.2623623.
[6] J. Lehmann et al., “DBpedia - A large-scale, multilingual knowledge base extracted from Wikipedia,” Semantic Web, vol. 6,

Automated dynamic schema generation using knowledge graph (Priyank Kumar Singh)
1268  ISSN: 2252-8938

no. 2, pp. 167–195, 2015, doi: 10.3233/SW-140134.

[7] F. M. Suchanek, G. Kasneci, and G. Weikum, “Yago: a core of semantic knowledge,” 16th International World Wide Web
Conference, WWW2007, pp. 697–706, 2007, doi: 10.1145/1242572.1242667.
[8] A. Carlson, J. Betteridge, B. Kisiel, B. Settles, E. R. Hruschka, and T. M. Mitchell, “Toward an architecture for never-ending
language learning,” Proceedings of the National Conference on Artificial Intelligence, vol. 3, pp. 1306–1313, 2010.
[9] N. Nakashole, M. Theobald, and G. Weikum, “Scalable knowledge harvesting with high precision and high recall,” Proceedings
of the 4th ACM International Conference on Web Search and Data Mining, WSDM 2011, pp. 227–236, 2011,
doi: 10.1145/1935826.1935869.
[10] W. Wu, H. Li, H. Wang, and K. Q. Zhu, “Probase: a probabilistic taxonomy for text understanding,” Proceedings of the ACM
SIGMOD International Conference on Management of Data, pp. 481–492, 2012, doi: 10.1145/2213836.2213891.
[11] G. Sudeepthi, G. Anuradha, and M. S. P. Babu, “A survey on semantic web search engine,” International Journal of Computer
Science, vol. 9, no. 2, pp. 241–245, 2012. Accessed: 20-Aug-2021. [Online]. Available:
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.403.482&rep=rep1&type=pdf.
[12] D. Damljanovic and K. Bontcheva, “Named entity disambiguation using linked data,” Proceedings of the 9th Extended Semantic
Web Conference (ESWC 2012), Poster session, pp. 231–240, 2012. Accessed: 20-Aug-2021. [Online]. Available:
http://2012.eswc-conferences.org/sites/default/files/eswc2012_submission_334.pdf.
[13] X. Han and J. Zhao, “Named entity disambiguation by leveraging wikipedia semantic knowledge,” International Conference on
Information and Knowledge Management, Proceedings, pp. 215–224, 2009, doi: 10.1145/1645953.1645983.
[14] R. Hoffmann, C. Zhang, X. Ling, L. Zettlemoyer, and D. S. Weld, “Knowledge-based weak supervision for information
extraction of overlapping relations,” ACL-HLT 2011 - Proceedings of the 49th Annual Meeting of the Association for
Computational Linguistics: Human Language Technologies, vol. 1, pp. 541–550, 2011.
[15] J. Daiber, M. Jakob, C. Hokamp, and P. N. Mendes, “Improving efficiency and accuracy in multilingual entity extraction,” ACM
International Conference Proceeding Series, pp. 121–124, 2013, doi: 10.1145/2506182.2506198.
[16] H. Wang, M. Zhao, X. Xie, W. Li, and M. Guo, “Knowledge graph convolutional networks for recommender systems,” The Web
Conference 2019-Proceedings of the World Wide Web Conference, WWW 2019, pp. 3307–3313, 2019,
doi: 10.1145/3308558.3313417.
[17] Apple, “Apple Siri,” 2016. [Online]. Available: https://www.apple.com/ios/siri/. (Accessed Aug. 24, 2021).
[18] “IBM Watson,” 2017. [Online]. Available: https://www.ibm.com/watson/. (Accessed Aug. 24, 2021).
[19] L. Li et al., “Real-world data medical knowledge graph: construction and applications,” Artificial Intelligence in Medicine,
vol. 103, 2020, doi: 10.1016/j.artmed.2020.101817.
[20] J. L. K. Xu, W. Hu and S. Jegelka, “How powerful are graph neural networks,” Proc. ICLR, pp. 1–17, 2019.
[21] Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu, “A comprehensive survey on graph neural networks,” IEEE
Transactions on Neural Networks and Learning Systems, vol. 32, no. 1, pp. 4–24, 2021, doi: 10.1109/TNNLS.2020.2978386.
[22] Z. Zhao, S.-K. Han, and I.-M. So, “Architecture of knowledge graph construction techniques,” Mathematics, International
Journal of Pure and Applied, vol. 118, no. 19, pp. 1869–1883, 2018.
[23] M. Strube and S. P. Ponzetto, “WikiRelate! Computing semantic relatedness using Wikipedia,” AAAI, vol. 6, pp. 1419–1424,
2006.
[24] O. Alotaibi and E. Pardede, “Transformation of Schema from Relational Database (RDB) to NoSQL Databases,” Data, vol. 4,
no. 4, p. 148, Nov. 2019, doi: 10.3390/data4040148.
[25] P. Merlo, J. Henderson, G. Schneider, and E. Wehrli, “Learning document similarity using natural language processing,”
Linguistik Online, vol. 17, no. 5, Dec. 2003, doi: 10.13092/lo.17.788.
[26] M. A. Rodriguez, “The gremlin graph traversal machine and language (Invited Talk),” DBPL 2015 - Proceedings of the 15th
Symposium on Database Programming Languages, pp. 1–10, 2015, doi: 10.1145/2815072.2815073.

BIOGRAPHIES OF AUTHORS

Priyank Kumar Singh is currently pursuing his Bachelor's degree in Computer

Science and Engineering at R.V College of Engineering. He completed his schooling and Pre-
University degree at St Mary's Academy, Saharanpur, India. His areas of interest include Natural
Language Processing, Big Data analysis, Web development, and Python Development. He can
be contacted at email: [email protected]

Sami Ur Rehman is currently pursuing his Bachelor's degree in Computer Science

and Engineering at R.V College of Engineering. He completed his schooling at Iqra High School
and his Pre-University degree at M.E.S Chaitanya P.U. college. His areas of interest include
Artificial Intelligence, Natural Language Processing, Cybersecurity, and Web development. He
can be contacted at email: [email protected]

Int J Artif Intell, Vol. 11, No. 4, December 2022: 1261-1269

Int J Artif Intell ISSN: 2252-8938  1269

Darshan J. is currently pursuing his Bachelor's degree in Computer science at R.V

College of Engineering. He has completed his schooling at Evershine English School and his PU
degree at R.V PU college, Bengaluru in the year 2019. His areas of interest include IoT, Web
development, and Machine learning. He can be contacted at email: [email protected]

Shobha G. Professor, Computer Science, and Engineering Department, R.V

College of Engineering, Bengaluru, India has teaching experience of 26 years, her specialization
includes Data mining, Machine Learning, and Image processing. She has published more than
150 papers in reputed journals/conferences. She has also executed sponsored projects worth INR
200 lakhs funded by various agencies nationally and internationally. She is a recipient of various
awards such as the Career Award for young teachers 2007-08 constituted by All India Council of
Technical Education, Best Researcher award from Cognizant 2017, GHC Faculty Scholar for
Women in Computing in 2018, IBM Shared University Research Award in 2019, HPCC
Systems community recognition award 2020. She is also an advisory committee member for IET
India Scholarship Award 2021. She can be contacted at email: [email protected]

Deepamala N. Associate Professor, Computer Science, and Engineering

Department, R.V College of Engineering, Bengaluru, India has Industry experience of 5 years
and teaching experience of 11 years. She has worked in MNC like D-Link, Nortel Networks, and
Radware before pursuing a teaching career. Her specialization is in Natural Language Processing
and Networking. She has one patent in her name and has 20 publications in various reputed
journals. She has executed consultancy and R&D projects as PI and Co-PI. She has published a
chapter in Elsevier’s book. She can be contacted at email: [email protected]

Automated dynamic schema generation using knowledge graph (Priyank Kumar Singh)