UpKG: A Framework to Insert New Domains in Knowledge Graphs

Jesamin Melissa Zevallos-Quispe

1 a

, Andr

e Gomes Regino

1 b

, V

ıctor Jes

us Sotelo Chico

1 c

ıctor Hochgreb de Freitas

2 d

and Julio Cesar dos Reis

1 e

Institute of Computing, University of Campinas, Campinas, Brazil

GoBots, S

ao Paulo, Brazil

Keywords:

Knowledge Graphs, Domain Extension, e-Commerce, Compatibility, Ontology, SPARQL.

Abstract:

In recent years, the creation of Knowledge Graphs (KGs) has advanced signiﬁcantly. They have become

essential in several domains, such as e-commerce. E-commerce applications apply them in the search and

recommendation of products and virtual chatbots, among other tasks. However, e-commerce must constantly

cover new domains/categories to respond to new user needs. This process requires rigorous analysis to cover

new domains, adding novel knowledge not stored in the KG. This study proposes, builds, and evaluates a

framework we named UpKG to insert new domains in existing KGs within an e-commerce context. Our ap-

proach relies on questions and answers collected in the real-world context of the GoBots company to facilitate

the new domain insertion process. Our framework was applied in GoBots company, specialized in e-commerce

solutions in Latin America. The conducted case was on an existing KG, whose triples were dominated by the

Automobile domain. Our work covered and inserted novel triples concerning the Appliances domain. The

conducted evaluation shows that we can verify the feasibility and applicability of the UpKG framework.

1 INTRODUCTION

In recent years, the creation of Knowledge Graphs

(KGs) has made signiﬁcant progress. With the rapid

growth of data in different ﬁelds, KGs play a vital

role in transferring large amounts of data to actionable

knowledge. They are essential in various domains,

including e-commerce (Li et al., 2020) (Sant’Anna

et al., 2020) (Chen et al., 2019) because several ap-

plications related to online commerce explore them

in different areas, such as product searches, product

recommendations, and virtual chats, among others.

Several big-player companies have used KGs in e-

commerce, including Alibaba (Li et al., 2020), Wal-

mart (Xu et al., 2020), and Farfetch (Barroca et al.,

2022), among others.

Due to constant changes, e-commerce systems

must be updated to continue providing better cus-

tomer service. They must constantly cover new do-

mains to increase knowledge of KG and update KGs,

https://orcid.org/0009-0006-6419-4694

https://orcid.org/0000-0001-9814-1482

https://orcid.org/0000-0001-9245-8753

https://orcid.org/0000-0002-0529-7312

https://orcid.org/0000-0002-9545-2098

inserting knowledge through triples based on a new

domain/category to respond to the novel users’ needs.

However, keeping a KG updated involves expanding

it and inserting domains such as appliances, furniture,

and automobiles.

In addition, inserting new domains into a KG is

not trivial because the ontology expert needs to ana-

lyze and understand the components, such as ontol-

ogy structure, relations, instances, SPARQL queries,

and data insertion to the KG with the new domain,

among other aspects. For example, suppose an on-

line store has a KG of only Furniture items, and a

customer requests information on a product from the

Clothing domain. In that case, this KG cannot return

an adequate response to the customer. For this reason,

inserting the ‘Clothing’ domain in such KG leads to a

better quality of response and greater customer satis-

faction.

The literature presents research studies related to

KGs in e-commerce (Li et al., 2020) (Chen et al.,

2019) (Sant’Anna et al., 2020). Some studies ex-

plore the extension of domains in KGs (Ait-Mlouk

and Jiang, 2020; Sheng et al., 2019) and ways of pop-

ulating triples in KGs (Kertkeidkachorn and Ichise,

2018) (Sant’Anna et al., 2020). However, none of the

Zevallos-Quispe, J., Regino, A., Chico, V., Hochgreb de Freitas, V. and Reis, J.

UpKG: A Framework to Insert New Domains in Knowledge Graphs.

DOI: 10.5220/0012237600003598

In Proceedings of the 15th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2023) - Volume 2: KEOD, pages 85-95

ISBN: 978-989-758-671-2; ISSN: 2184-3228

analyzed studies explores these three critical aspects

together of inserting new domains into e-commerce

KGs, which are (i) KG in e-commerce, (ii) Populat-

ing KG, and (iii) Extending the KG.

In this study, we propose the UpKG framework,

which allows inserting new domains in an existing

KG with an e-commerce approach based on user

questions and answers. Our study presents the fol-

lowing contributions:

• Conceptually develop UpKG focusing on e-

commerce context based on user questions and

answers.

• Evaluate the proposed UpKG framework in the

real-world context of an AI startup in Latin Amer-

ica.

• Ensure that our framework is ﬂexible enough to

add new domains focusing on e-commerce.

Our evaluation was conducted in the context of the

GoBots

company, where we inserted triples of the

Appliances domain into a KG composed of triples of

the Automobiles domain. It was possible to insert 260

individuals for both appliances and automobiles. In

addition, our results showed that it is possible to an-

swer questions from users with SPARQL queries rely-

ing on the new version obtained for the KG. We ver-

iﬁed the ontology’s consistency and coherence with

the OntoDebug

plugin of Prot

e. The main goal of

OntoDebug is to help users ﬁnd those axioms in on-

tologies, giving information if an ontology is incon-

sistent.

The remainder of this article is organized as fol-

lows. Section 2 presents a summary of related work.

Section 3 describes the UpKG framework. Section 4

presents the evaluation of our framework, where we

show the context and the results obtained from apply-

ing our framework in an existing real-world setting.

Section 5 discusses our ﬁndings. Finally, Section 6

concludes the study and describes future work.

2 RELATED WORK

There are several studies related to the use of KGs

for e-commerce. Some of them are based on speciﬁc

tasks to insert a new domain.

KG in e-commerce. Li et al. (Li et al., 2020)

proposed AlimeKG, an e-commerce KG representing

knowledge about user problems, item information,

Company specialized in artiﬁcial intelligence solutions

for e-commerce platforms in Latin America. Site available

in https://gobots.ai/

http://isbi.aau.at/ontodebug/

and item relationships. AlimeKG helps to understand

users’ needs, answer pre-sales questions, and gener-

ate explanatory texts. This KG was applied to various

business scenarios, such as buying guidance, answer-

ing property questions, and generating recommenda-

tion reasons. In addition, Li et al. (Li et al., 2020) de-

scribed how the domain KG comprises free text and

tested it with various applications in e-commerce cat-

egories.

Xu et al. (Xu et al., 2020) proposes an approach

based on a Product Knowledge Graph (PKG) to learn

the inherent relationships between products, in which

they make use of a method of learning distributed rep-

resentation enhanced with self-attention. However,

this study does not consider a customer’s information

as part of e-commerce, since it is focused on products

and how their information can be used for recommen-

dation and classiﬁcation tasks.

Populating KG. The T2KG (Kertkeidkachorn

and Ichise, 2018) creates an automatic KG from nat-

ural language texts and ﬁlls an existing KG with new

knowledge. Natural language context entities are as-

signed to the corresponding uniform resource identi-

ﬁer in the KG, which is usually the subject or object of

the triples. They combined rule-based and similarity-

based techniques to map the predicate of a triple gen-

erated from text in an existing KG. The experimental

part showed that the T2KG framework can success-

fully generate a KG and populate an existing one with

new text knowledge. However, the framework does

not work correctly when mapping predicates contain-

ing many compound words since the triplet extraction

step is not perfect due to the complexity of the text in

open domains.

Sant’Anna et al. (Sant’Anna et al., 2020) pro-

posed a method to populate a KG from a collection

of pairs of questions and answers, in which they per-

formed an extraction of entities and intents to gen-

erate triples supported by the ontology deﬁned in its

KG. The knowledge generated and obtained from the

pairs of questions and answers is stored in the KG.

However, the method requires that the attendants an-

swer the questions correctly through a manual pro-

cess. Similarly to our work, the method proposed by

Sant’Anna et al. (Sant’Anna et al., 2020) focused on

the e-commerce domain.

Integroly (Guedea-Noriega and Garc

ıa-S

anchez,

2022) is a framework that automates the KG pop-

ulating. It collects data from digital media sources

to populate a KG focusing on the political marketing

domain. They used a set of Natural Language Pro-

cessing (NLP) techniques to extract knowledge from

political marketing texts written in Spanish. In their

experiments, the authors used Twitter and political

KEOD 2023 - 15th International Conference on Knowledge Engineering and Ontology Development

Table 1: Comparison in three scopes of studies in KG connected to our present investigation in this article.

KG in

e-commerce

Extending

Populating

(Kertkeidkachorn and Ichise, 2018) X

(Sheng et al., 2019) X

(Xu et al., 2020) X

(Ait-Mlouk and Jiang, 2020) X

(Li et al., 2020) X

(Sant’Anna et al., 2020) X X

(Guedea-Noriega and Garc

ıa-S

anchez, 2022) X

Our proposal X X X

news. They found problems with synonyms, ambi-

guities, and the connections made in the KG.

Extending KG. KBot (Ait-Mlouk and Jiang,

2020) is a multilingual chatbot that understands user

queries. KBot has automated learning based on clas-

sifying intents. In addition, KBot can add a new

dataset to the existing knowledge base, allowing the

expansion of the chatbot’s capabilities and thus hav-

ing a greater understanding of queries.

KGs have many applications and are important in

the medical sector. It is challenging to build a KG for

all diseases. For this reason, DEKGB (Sheng et al.,

2019) proposed an efﬁcient and extensible framework

to build KG for speciﬁc diseases based on doctors’

knowledge. They described the process by extending

an existing health KG to include a new disease.

Table 1 shows the organization and conceptual

analysis we performed for the existing literature re-

viewed. We considered three scopes.

Our study aims to cover these three aspects: (i)

KG in e-commerce, use of KG in an e-commerce plat-

form; (ii) Populating KG, populate a KG with new

knowledge based on the ontology; and (iii) Extend-

ing KG, modify the KG by adding new instances with

their relations. The goal is to build our framework

to insert new domains in a KG with an e-commerce

context based on users’ questions and answers from

GoBots, which owns a set of objects that contain the

user’s question and the human assistant’s answer. The

following section describes our framework in detail.

3 FRAMEWORK UpKG

We propose a framework to insert new domains into

an existing KG focused on e-commerce from ques-

tions and answers. Figure 1 presents the general

pipeline of our framework. It is composed of four

modules: (i) Understand the current KG, (ii) Identify

a new domain, (iii) Restructure the ontology, (iv) Pop-

ulate the KG. These modules are further elaborated in

the following.

3.1 Understand the Current KG in

Place

Our framework works on an existing KG where the

module at this stage aims to understand the shape

and structure of the initial KG. Understanding the KG

structure is necessary as our ﬁrst step, which allows us

to identify and analyze how the existing KG is struc-

tured and constructed, which we denote as v0. This

step can be performed by a person with knowledge of

ontologies, observing the properties and attributes of

the data. For the second step, we analyze the struc-

ture of the existing ontology, which represents all the

existing relationships in the KG. The comprehension

of the ontology is as critical as that of the KG. In this

step, you can use some tools to understand the ontol-

ogy structure; one of the best-known is Prot

e. In the

third step, the relationships of the KGs are identiﬁed;

this consists of observing all the possible relationships

that the ontology allows and how it helps achieve the

objective of the KG (answer users’ questions). The

fourth step is to identify the domains currently sup-

ported by the KG v0 and thus map the domains that

are necessary to insert. In this step, the human per-

son can access the current KG information and dis-

cern which domains the KG covers.

The development of this module can be done man-

ually since all the processes described require the in-

tervention of human reasoning.

UpKG: A Framework to Insert New Domains in Knowledge Graphs

Figure 1: Overview of the UpKG Framework. (i) Understand the current KG: Module focused on understanding the current

KG and its base ontology; (ii) Identify new domain: Module focused on discovering the new domain that will be inserted;

(iii) Restructure the ontology: Critical module in which the base ontology is combined with the new properties; (iv) Populate

the KG: Final module focused on generating the triples to insert the KG.

3.2 Identify New Domain

The objective of the second module is to identify a

new domain where the KG in place cannot perform

its task, in this case, answering questions that are not

mapped in the current KG. The ﬁrst step is to iden-

tify all the compatibility questions not answered by

the existing KG. This step can be done automatically

through a script, which has as its input a collection

of user questions and answers. The saved questions

are those that have no answer from the KG. A per-

son with programming knowledge can execute this.

The second step is identifying relationships that the

ontology v0 cannot map, making it impossible to re-

spond to the KG according to its intent. The third step

refers to identifying the domain with the highest num-

ber of unanswered questions from the current KG. To

ﬁnd the domain to insert, you can create a script that

generates a histogram automatically, which represents

the domain and its number of unanswered questions.

This allows a graphical display of domains that do not

cover the current KG; we can decide which domain to

include with this histogram.

This module can be developed fully automatically,

unlike the previous module (cf. Subsection 3.1), be-

cause each step can be applied using a script.

3.3 Restructure the Ontology

The goal of the third module is to restructure the exist-

ing ontology with the relationships not mapped to the

domain found in the previous module 3.2. The input

data to be used in this module is the new domain to be

inserted and the existing ontology denoted as v0; usu-

ally, an ontology is modiﬁed based on the new needs

of a system. Therefore, it is essential to modify the

structure of knowledge. The data from this module is

the ontology with the new domain to which we denote

it by the version v1.

The ﬁrst step in this process is to search for avail-

able ontologies in the selected domain to reference

to understand how this knowledge is structured. This

search can be carried out through a review of the lit-

erature. The second step synthesizes the ontologies

with a set of properties already used in other con-

texts, which can help us in our current KG and thus

start from a base ontology. The third step is to obtain

the new ontology properties necessary for the new do-

main and the type of questions to ask; for example, if

our current domain is Automobiles, we could not an-

swer questions that refer to Technology. In this step,

the intervention of human reasoning is necessary to

combine the existing properties with the new ones; it

is possible to ﬁnd new properties necessary to store

knowledge based on the set of questions and answers.

For example, by analyzing the structure of questions.

The fourth step manually combines the properties that

were previously mapped with our ontology v0; in this

step, we can use a tool that facilitates ontology mod-

eling (eg. Prot

) to obtain our ontology v1 as out-

put.

This module can be developed manually because

https://protege.stanford.edu/

KEOD 2023 - 15th International Conference on Knowledge Engineering and Ontology Development

in all the steps human intervention is necessary to re-

structure the ontology.

3.4 Populate the KG

The last module aims to insert or populate the KG

with new triples extracted from a collection of speciﬁc

questions and answers. These triples have to follow

the structure of the new ontology v1 to be added to

the KG.

The ﬁrst step is to obtain the information for each

question and answer from a speciﬁc domain, which

will be collected from the e-commerce system in

which this framework will be applied. A Product

refers to an item offered in an e-commerce. The sec-

ond step consists of extracting the intents and entities

from the questions and answers entered. This process

is based on the work of Sant’Anna et al. (Sant’Anna

et al., 2020), who proposed their own Natural Lan-

guage Understanding (NLU) process, where they de-

ﬁned the concepts of entity and intent.

Sant’Anna et al. (Sant’Anna et al., 2020) deﬁned

that the entity represents a term or expression with

a known meaning relevant to understanding the sen-

tence. In addition to having names and values, for

example, the “brand”, “model”, “voltage”, “volume”

present in an appliance. The intent is to identify

whether the question refers to compatibility between

entities. This step is responsible for processing the

stored set of questions answered manually by human

assistants and product information, a key source of

knowledge to update the KG and answer new simi-

lar questions. This extraction process is necessary to

structure the knowledge through RDF triples.

The third step is the creation of RDF triples, which

is a key factor in updating the current KG. This

knowledge extraction process is based on extracting

intents and entities from each pair of questions and

answers (input in our framework) to structure the ex-

tracted knowledge in RDF triples. The fourth step is

to store the collection of triplets in a database, this can

be local or remote. This step aims to access informa-

tion through SPARQL queries. In this step you can

use tools that store RDF data, as well as GraphDB

Virtuoso

, among others.

Finally, the output data of this last step is the KG

v1. This module can be developed automatically be-

cause each step can be executed through a script.

https://www.ontotext.com/products/graphdb/

https://virtuoso.openlinksw.com/

4 EVALUATION

Our framework was carried out in the context of the

GoBots and offers solutions to structure the knowl-

edge of compatibility between products in a KG based

on an ontology. The company offers a solution of

questions and answers for e-commerce platforms. We

describe the results of applying the process of each

framework module in the following.

4.1 Results Applying the upKG

Framework

4.1.1 Results in Understanding the Current KG

Figure 2 presents the application of the ﬁrst module,

which is to understand the current KG where we have

as input the KG of GoBots, which we denote as v0.

First, it is required to understand the structure of the

KG. In our case, we had access to GoBots documenta-

tion and the current KG in production to perform this

step. With the documentation, we conceptually un-

derstood how the KG was structured, the objective by

which the KG was designed, and how this was used

to answer questions in an e-commerce environment.

Figure 2: Results in understanding the current KG in GoB-

ots.

As a result of understanding the structure of KG,

we have a set of properties and attributes of the KG

v0. Figure 3 shows the initial ontology, in which we

UpKG: A Framework to Insert New Domains in Knowledge Graphs

observe the existing classes, properties, and attributes

to store the compatibility knowledge between a prod-

uct and a Car type object. In the second step, we used

the initial ontology v0 proposed in (Sant’Anna et al.,

2020). In this step, we used the tool Prot

e to visual-

ize and understand how the ontology was structured.

As a result of understanding current KG, we have a

set of properties and attributes in the ontology v0.

In the third step, like the previous step, we used

the Prot

e tool to understand the relationships pre-

sented in the initial ontology, which contains the fol-

lowing classes: Compatibility(FullCompatibility and

NoCompatibility), ConsumerItem, Product and Car.

Where the Product class represents a product in an e-

commerce system, ConsumerItem is an abstract class

representing an item that is related to the product, this

relationship is given by the Compatibility class, either

of the FullCompatibility and NoCompatibility types.

As a result of understanding relationships in ontology,

we have a set of relations of ontology v0.

In the fourth step, we deﬁned the domain or do-

mains that cover the current KG. In this case, the cur-

rent domain is Automobile, which was represented

by the Car class, keeping the following attributes:

hasYear, hasBrand and hasModel, which represent

the year, brand, and model of an object Car.

4.1.2 Result in Identifying New Domain

GoBots collected a set of pairs of questions and an-

swers related to a product that human attendants an-

swered. This collection is stored in a MongoDB

database. Figure 4 shows the process of identifying

the new domain to be inserted in the new version of

the KG. For the ﬁrst step, we identiﬁed all the ques-

tions that the current KG cannot answer, so we do not

consider the questions related to the domain of Au-

tomobiles. This task was carried out with a Python

script that connects directly to the GoBots dataset to

extract the question collection within a delimited date

range. In this experiment, the date range was two

months. As a result of the ﬁrst step is a collection

of questions that the KG did not answer.

In the second step, we identiﬁed the relations that

are not mapped in the ontology and, therefore, are im-

possible to answer by the KG. In this step, we an-

alyzed the structure of the questions and compared

them with the properties contained in the GoBots KG

v0. Among them, we have questions like ”Is it com-

patible with LSE09 220v?”, where LSE09 refers to

a Washing Machine (Appliance domain) that works

with a voltage of 220V. This example shows that the

KG cannot store this knowledge using the Car class

deﬁned in the ontology v0. As a result of the sec-

ond step, we have a set of relationships that were not

mapped by the current ontology.

In the third step, we created a script in Python

to identify the domains (categories) with the largest

number of questions sorted descending; the three do-

mains with the most questions and answers were Ap-

pliances, Home and Furniture, and Tools. In the next

steps, we present the insertion of the domain Appli-

ances to the new version of the ontology and later to

KG.

4.1.3 Result in Restructuring the Ontology

The ontology v0 is limited to the Automobiles do-

main, which does not allow receiving queries from

other domains. That is why when applying our UpKG

framework, the need to insert the “Appliances” do-

main was identiﬁed to cover a large part of the ques-

tions not answered by the current KG. In addition, we

decided to maintain the Compatibility intent between

a product and an object from the Automobile domain.

The ontology obtained from applying our framework

was denoted as v1, which can cover two domains of

questions users ask in an e-commerce system.

Figure 5 shows the process of restructuring the

GoBots ontology to accept the Appliance domain.

The ﬁrst step is to perform an ontology search; in our

case, they are appliances. A search for publicly avail-

able product ontologies was carried out because it is

intended to expand the KG to several e-commerce cat-

egories. We explored Schema.org

, which provides a

list of properties in Product and allows us to identify

the properties needed for the new domain. For this

reason, in the second step, we selected Schema.org as

the base for our ontology.

The third step is to identify the new properties nec-

essary for our ontology. In this case, we used standard

properties present in Schema.org regarding Product,

such as Model, Brand, and Material. We also created

new properties that were not found in Schema.org.

These are created from human reasoning to map the

relationships presented in the questions about Appli-

ances: Voltage and Volume.

The fourth step was to combine the properties; in

this step, we used the Prot

e tool to combine our

property list with the existing GoBots ontology v0.

The ﬁnal result of this step is an ontology that accepts

two domains: Automobiles and Appliances. We de-

noted this ontology as v1.

We added an Appliance class that is a subclass of

ConsumerItem just like the Car class to our initial on-

tology, in which we added all the attributes mapped

as: hasVolume and hasVolts.

Figure 9 shows in Prot

e tool visualization all

https://schema.org/Product

KEOD 2023 - 15th International Conference on Knowledge Engineering and Ontology Development

Figure 3: Ontology v0 that represents the compatibility knowledge between a product (a car seatbelt, for instance) and a

consumer item (an Audi TT, for instance) (Sant’Anna et al., 2020).

Figure 4: Diagram showing the result of identify new do-

main in GoBots.

the Data Properties that connect to the “Automobiles”

and “Appliances domains”, some of which present

independent attributes for each domain. Regarding

the Object Properties, our ontology maintains the two

existing relationships between the Product and the

ConsumerItem through a Compatibility type proposed

in (Sant’Anna et al., 2020). These are represented

Figure 5: Results in restructuring the GoBots ontology with

Appliance Domain.

by compatibleWith and hasCompatibility, where com-

patibleWith ranges to a ConsumerItem and domain to

a Compatibility type. However, hasCompatibility has

a Compatibility type as range and a Product domain.

Figure 6 shows the metrics of the ontology struc-

UpKG: A Framework to Insert New Domains in Knowledge Graphs

ture in terms of classes, properties, and axioms. The

ontology v0 presents 7 classes while our ontology v1.

Figure 6 presents a total of 8 classes. This shows that

inserting a new domain in ontology was simple, con-

sidering that ontology v0 was well-designed to accept

other domains.

Figure 6: Metrics of the ontology v1.

Figure 7 graphically represents how the v1 ontol-

ogy was structured.

Figure 8 and Figure 9 show the Object Property

and Data Property of the ontology v1, respectively.

The structure and relationships of the ontology v1

were validated using Prot

e built-in debugging tool

OntoDebug

. OntoDebug (Schekotihin et al., 2018)

is a Prot

e plugin that implements an interactive ap-

proach to ontology debugging, since having a faulty

ontology the tool ﬁnds a set of faulty axioms that ex-

plain the problem. The main goal of OntoDebug is to

help ﬁnd those axioms in ontologies, giving informa-

tion if an ontology is inconsistent. Figure 10 presents

our ﬁnal ontology v1 validated with OntoDebug. We

obtained as a result of the execution, that our ontology

is coherent and consistent.

4.1.4 Result in Populating the KG

To populate our KG, we collected a set of questions

and answers from the GoBots database between April

01, 2023 and June 01, 2023, obtaining a total of 47142

questions. We ﬁltered out those with Compatibility

intent, obtaining 137 questions.

Figure 11 shows the automatic process of popu-

lating the KG. The ﬁrst step is to obtain the product

information related to the question. This extraction of

information was carried out by a POST request, which

returns the necessary attributes (SKU, GTIN or Pro-

ductID) to create a product in our ontology v1 (cf.

Figure 7).

The second step was the extraction of the intents

and entities of the questions and answers collected

from the GoBots database. This collection of ques-

tions and answers is related to the domain of Appli-

ances. In this step, we deﬁned and created an original

http://isbi.aau.at/ontodebug/

Python Script to identify the questions with Compati-

bility intent.

Once we have ﬁltered the questions with the com-

patibility intent, we proceed with the automatic ex-

traction of the entities involved. As an example, the

question ”Is it compatible with LSE09 220v?”, from

which we extract the entities present, such as the

Model and Voltage. To perform this extraction, we

use a script, which receives a question as input and

returns all the entities deﬁned (domain, model, brand,

volume, volts, design, and material).

As a result of this extraction, we have an in-

tents/entities collection. The third step consists of the

creation of RDF triplets; it is necessary to emphasize

that this step is performed automatically, like the pre-

vious two; this step is responsible for generating all

the relationships allowed by our ontology v1 for the

different domains, among them generating instances

of Automobiles and Appliances by triplets. These are

possible to make thanks to the collection of extracted

entities.

The fourth step was to store the data in the RDF

Stores. At this stage, we explored GraphDB

, which

is a database that stores graphs. This tool was used

to store the collected triples and execute SPARQL

queries, thus populating the new KG v1 with ques-

tions and answers from the Appliances and Automo-

biles domain.

Of these 137 questions, the automatic insertion

was done directly with a script to the OWL ﬁle gener-

ated by Prot

e. However, not all the questions were

inserted, as some did not have all the necessary enti-

ties to generate knowledge. A total of 260 individuals

were automatically inserted: 53 individuals in the Au-

tomobiles domain, 34 individuals in the Appliances

domain, 84 individuals in Product, 85 in FullCompat-

ibility, and ﬁnally, 4 in NoCompatibility, this being

our KG v1.

4.2 Results in Using the Outcome KG

With our KG v1 populated, we access the data

through SPARQL queries. These queries were de-

ﬁned for different relevant user questions, taking into

account the Compatibility intent.

To present the generated KG’s value, quality, and

effectiveness, we show questions for the “Automo-

biles” and “Appliances” domains with their respective

SPARQL queries.

• ”Is it compatible with LSE09 220v?”. This

question refers to the product with GTIN

763293865672, where they are asking if the prod-

https://www.ontotext.com/products/graphdb/

KEOD 2023 - 15th International Conference on Knowledge Engineering and Ontology Development

Figure 7: Ontology structure v1. Represents the compatibility between products and an object of Car or Appliance type.

Figure 8: Object Property of the ontology v1.

Figure 9: Data Property of the ontology v1.

uct is compatible with a LSE09 model wash-

ing machine (Appliance) and that it is compati-

ble with a 220V voltage. The query generated in

SPARQL would be the following:

select ?product ?comp_type ?appliance

where {

?product rdf:type UPkg:Product;

UPkg:hasCompatibility ?comp.

{?product UPkg:hasGTIN "763293865672"}

?comp

UPkg:compatibleWith ?appliance;

rdf:type ?comp_type.

?comp_type

rdfs:subClassOf UPkg:Compatibility.

?appliance

UPkg:hasApplianceModel "LSE09";

UPkg:hasVolts "220V".

}

• Is this pump compatible with Ford ﬁesta Zetec

year 2003?. This question refers to the product

with GTIN 827365281647, if it is compatible with

a Ford (Car), model Fiesta and year 2003. The

query generated in SPARQL would be the follow-

ing:

select ?product ?comp_type ?car

where {

?product rdf:type UPkg:Product;

UPkg:hasCompatibility ?comp.

{?product UPkg:hasGTIN "827365281647"}

?comp

UPkg:compatibleWith ?car;

rdf:type ?comp_type.

?comp_type

rdfs:subClassOf UPkg:Compatibility.

?car

UPkg:hasCarBrand "Ford";

UPkg:hasCarModel "Fiesta";

UPkg:hasYear "2003".

}

• Good morning, my washing machine is a

UpKG: A Framework to Insert New Domains in Knowledge Graphs

Figure 10: The debug output of the UpKG ontology shows that it is consistent and coherent.

Figure 11: Results in populating the KG with Appliance

Domain.

Brastemp 15 KG and model BWS15ABANA20,

Is this valve model compatible?. This question

refers to the product with GTIN 1234567890123,

they are asking if the product is compatible with

Brastemp brand, BWS15ABANA20 model and

15KG of volume washing machine (Appliance).

The query generated in SPARQL would be the

following:

select ?product ?comp_type ?appliance

where {

?product rdf:type UPkg:Product;

UPkg:hasCompatibility ?comp.

{?product UPkg:hasGTIN "1234567890123"}

?comp

UPkg:compatibleWith ?appliance;

rdf:type ?comp_type.

?comp_type

rdfs:subClassOf UPkg:Compatibility.

?appliance

UPkg:hasApplianceBrand "Brastemp";

UPkg:hasApplianceModel "BWS15ABANA20";

UPkg:hasVolume "15KG".

}

5 DISCUSSION

This article addressed the problem of addressing

new domains in extending an existing KG in an e-

commerce setting. Inserting a new domain in a KG is

key to user questions and answers about products on

an e-commerce platform.

Our framework was applied under the existing

GoBots KG. This solution aimed to facilitate the in-

sertion of new domains according to the needs of e-

commerce. Our solution has the potential to improve

customer service on an e-commerce platform.

Our framework allowed us to successfully insert a

new domain, which can be applied to other domains

following the steps deﬁned by the UpKG framework.

The existing KG must have the following require-

ments to apply the UpKG framework: it must focus

on e-commerce, have a collection of user questions

and answers, and have questions answered by human

attendees.

As we observed results in Section 4.2, the appli-

cation of our framework enabled the insertion of a

new domain to an existing KG in a real-world en-

vironment. We reached a new version of an ontol-

ogy considering classes that represent knowledge in

both domains. We addressed the objects that link our

classes and, ﬁnally, the properties of the classes, re-

spectively. In the same way, it is important to have

many relationships between a product and an object

of type ConsumerItem to be able to answer several

questions from consumers. From another perspective,

using all our stored knowledge for different tasks, in-

cluding product recommendations, is possible.

The application of the UpKG framework was de-

KEOD 2023 - 15th International Conference on Knowledge Engineering and Ontology Development

veloped using real GoBots data relying on a collection

of questions and answers. The stage of populating the

KG appeared to be the most complicated because you

need to automate various processes, including Ex-

tracting question intents/entities, creating triples, and

storing the data in an RDF Store. However, automat-

ing this popular KG process allows us to process more

data in less time than a manual process. Furthermore,

it allowed us to evaluate our UpKG framework com-

pletely.

6 CONCLUSION

Inserting new domains in an existing KG focused on

e-commerce is an absolute necessity to enable cover-

ing a large amount of knowledge because it is con-

stantly under evolution. This study proposed a frame-

work suited to inserting new domains in a KG. Four

main modules were proposed: Understanding the cur-

rent KG, Identifying new domains, Restructuring the

ontology, and Populating the KG. Our framework was

applied in the context of a real-world company using

a collection of questions and answers about Appli-

ances. We showed that the framework application was

developed considering the addition of new individu-

als, obtaining a coherent and consistent KG within

the new domain, evidencing the viability and appli-

cability of the framework. In future work, we plan to

bring the KG v1 to production to have a further con-

sistent evaluation assessing questions asked by users

in real time. Finally, we aim to insert new domains

detected in stage two of our framework to have more

knowledge in our KG.

ACKNOWLEDGEMENTS

This study was ﬁnanced in part by the Coordenac¸

de Aperfeic¸oamento de Pessoal de N

ıvel Superior

– Brasil (CAPES) - Finance Code 001 and by the

ao Paulo Research Foundation (FAPESP) (Grant

#2022/13694-0)

. We also would like to thank GoB-

ots for collecting data and sharing their environment.

REFERENCES

Ait-Mlouk, A. and Jiang, L. (2020). Kbot: a knowledge

graph based chatbot for natural language understand-

ing over linked data. IEEE Access, 8:149220–149230.

The opinions expressed in this work do not necessarily

reﬂect those of the funding agencies.

Barroca, J., Shivkumar, A., Ferreira, B. Q., Sherkhonov,

E., and Faria, J. (2022). Enriching a fashion knowl-

edge graph from product textual descriptions. arXiv

preprint arXiv:2206.01087.

Chen, S., Li, C., Ji, F., Zhou, W., and Chen, H. (2019).

Driven answer generation for product-related ques-

tions in e-commerce. In Proceedings of the Twelfth

ACM International Conference on Web Search and

Data Mining, pages 411–419.

Guedea-Noriega, H. H. and Garc

ıa-S

anchez, F. (2022). In-

tegroly: Automatic knowledge graph population from

social big data in the political marketing domain. Ap-

plied Sciences, 12(16):8116.

Kertkeidkachorn, N. and Ichise, R. (2018). An automatic

knowledge graph creation framework from natural

language text. IEICE TRANSACTIONS on Informa-

tion and Systems, 101(1):90–98.

Li, F.-L., Chen, H., Xu, G., Qiu, T., Ji, F., Zhang, J.,

and Chen, H. (2020). Alimekg: Domain knowledge

graph construction and application in e-commerce.

In Proceedings of the 29th ACM International Con-

ference on Information & Knowledge Management,

pages 2581–2588.

Sant’Anna, D. T., Caus, R. O., dos Santos Ramos, L.,

Hochgreb, V., and dos Reis, J. C. (2020). Generat-

ing knowledge graphs from unstructured texts: Expe-

riences in the e-commerce ﬁeld for question answer-

ing. volume 2722 of CEUR Workshop Proceedings,

pages 56–71. CEUR-WS.org.

Schekotihin, K., Rodler, P., and Schmid, W. (2018). On-

todebug: Interactive ontology debugging plug-in for

prot

e. In Foundations of Information and Knowl-

edge Systems: 10th International Symposium, FoIKS

2018, Budapest, Hungary, May 14–18, 2018, Pro-

ceedings 10, pages 340–359. Springer.

Sheng, M., Shao, Y., Zhang, Y., Li, C., Xing, C., Zhang,

H., Wang, J., and Gao, F. (2019). Dekgb: an extensi-

ble framework for health knowledge graph. In Inter-

national Conference on Smart Health, pages 27–38.

Springer.

Xu, D., Ruan, C., Korpeoglu, E., Kumar, S., and Achan, K.

(2020). Product knowledge graph embedding for e-

commerce. In Proceedings of the 13th international

conference on web search and data mining, pages

672–680.

UpKG: A Framework to Insert New Domains in Knowledge Graphs