Broaden Your Horizon! Play with Semantics via a Knowledge

Graph-Based Approach

Pasquale Esposito

1 a

, Crescenzo Mazzone

, Maria Angela Pellegrino

2 b

and Vittorio Scarano

2 c

Dipartimento di Studi Umanistici, Universit

a degli Studi di Salerno, Fisciano (SA), Italy

Dipartimento di Informatica, Universit

a degli Studi di Salerno, Fisciano (SA), Italy

Keywords:

Lexical Semantics, Knowledge Graph, Semantic Web, Word Senses, User Evaluation, Italian.

Abstract:

(Lexical) semantics is the study of the meaning of words by looking at either the word itself or exploring its

neighborhood. Hence, lexicons, synonyms, and analogies can be easily represented as semantic networks,

also known as knowledge graphs, to represent words and their connections. While knowledge graphs can be

perceived as a natural and intuitive representation for modeling and exploring words, directly accessing them

via standard query languages, such as SPARQL, is cumbersome, mainly for lay users. This article explores

the possibility of “playing with semantics” via a knowledge graph-based approach to let end-users explore

lexical-semantic relations without explicitly formulating SPARQL queries. We evaluated the accuracy and

the coverage of users’ expectations by inquiring about 27 Italian native speakers and compared quantitative

and qualitative results. According to the performed evaluation, knowledge graphs have the potential to fulﬁll

users’ satisfaction, but multiple source results must be merged to guarantee high coverage and accuracy.

1 INTRODUCTION

Semantics is the study of the meaning of words and

how linguistic signs mediate between concepts and

forms (Pustejovsky, 2016). Semantics studies are di-

vided into two sub-branches: logical and lexical se-

mantics. The present study mainly focuses on the lat-

ter and investigates how the relations among units of

meaning interact with their biological mental repre-

sentations. In other words, this study relies on how

native speakers perceive sense out of a combination of

units of meaning organized in sentences. Since words

are inventories of arbitrary signs, people can interpret

words or their combinations differently. Semantics

can provide a series of instruments to understand bet-

ter complex cases of words with multiple meanings

and connections in their lexical-semantic network.

Semantic networks’ linguistic-traditional and for-

mal representation can be considered similar to the in-

formation technology representation of semantic net-

works, a.k.a. Knowledge Graphs (KGs). Linguistics’

lexical semantics establishes connections and similar-

ities among words by looking at the neighborhood of

https://orcid.org/0009-0006-3464-5861

https://orcid.org/0000-0001-8927-5833

https://orcid.org/0000-0001-8437-5253

the network, i.e., words that occur within natural sen-

tences. Similarly, KGs represent semantic relations

between concepts in a directed graph consisting of

vertices, a.k.a. concepts, and edges, a.k.a. semantic

relations between concepts (Sowa, 2006).

KGs, and Semantic Web technologies in gen-

eral, are helpful in representing words and their sur-

roundings as semantic networks to let users explore

words in context, and they are crucial for knowl-

edge management and information retrieval (De Do-

nato et al., 2020). (Mitchell et al., 2008) showed

that brain imaging studies have different spatial neu-

ral activation patterns while thinking about pictures

or words. This phenomenon is widely debated in

psycho-linguistic literature and massively explored by

many experiments eager to explore neuronal seman-

tic networks, such as those conducted by (Wang et al.,

2022; Emma James and Henderson, 2023). However,

KG query languages are complex for lay users (Var-

gas et al., 2019; Bellini et al., 2014). Hence, KG

querying mechanisms should mask syntactical com-

plexities, allowing users to pose queries easily to ex-

ploit semantic network content.

This paper explores the opportunity to “play with

(lexical) semantics”, meaning that users are supported

in querying a word, retrieving its meaning, and ex-

ploring all the words connected to it as synonyms,

380

Esposito, P., Mazzone, C., Pellegrino, M. and Scarano, V.

Broaden Your Horizon! Play with Semantics via a Knowledge Graph-Based Approach.

DOI: 10.5220/0012620700003693

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Conference on Computer Supported Education (CSEDU 2024) - Volume 1, pages 380-387

ISBN: 978-989-758-697-2; ISSN: 2184-5026

analogies, and hierarchies. This word exploration is

enabled via a KG-based approach by implicitly pos-

ing SPARQL queries while masking underlying com-

plexity. To verify if and to what extent the pro-

posed prototype, referred to as LexicalEngine, sat-

isﬁes users’ expectations, we tested it on 27 Italian

native speakers and collected their opinions via an

online questionnaire. While LexicalEngine cov-

ers a wide range of semantic relations, such as def-

initions, synonyms, and word hierarchies, we focus

on word senses in this contribution. In particular,

we investigated if the semantics modeled via KGs

can disambiguate terms by providing end-users with

a complete and accurate pool of word senses sufﬁ-

ciently close to the one proposed by (adult) native

speakers. We compared word senses automatically

retrieved by KGs with those proposed by partici-

pants. We veriﬁed if LexicalEngine returns word

senses suggested by adult Italian speakers by measur-

ing the coverage metric. Supposing that there are

word senses foreseen by Italian speakers and not re-

turned by LexicalEngine, those words would affect

the coverage metric. The higher the coverage, the

greater the guarantee that an automatic KG-driven ap-

proach would also return any word sense suggested

by an Italian speaker. Besides the coverage, provid-

ing learners with tip support and avoiding informa-

tion load by proposing out-of-topic and inaccurate

suggestions is crucial. Hence, we reﬂected on word

senses proposed by LexicalEngine but not foreseen

by Italian speakers. For those word senses, we asked

participants to assess their accuracy by reporting if

and to what extent users considered those meanings

as valid. The higher the accuracy, the greater the in-

spirational power of the proposed prototype, able to

propose valid meanings, even not immediately fore-

seen by native speakers. The higher the assessed met-

rics, the greater the guarantee that end-users, such

as (young) learners, can be successfully supported in

self-learning lexical semantics via a KG-driven ap-

proach. By formulating those directions as research

questions (RQs), we explored the following aspects:

- To what extent do KGs cover word senses be-

longing to Italian native speakers’ mental lexicon?

- Can KGs inspire end-users by proposing un-

common but accurate word senses? To what extent?

According to the collected results,

LexicalEngine reaches a considerable cover-

age of users’ expectations in terms of word senses

envisioned by (adult) Italian native speakers by

combining multiple KG lexical-semantic features and

their data (RQ

). On the contrary, participants are

skeptical about uncommon word senses suggested

by LexicalEngine, that require to be justiﬁed to

be appreciated and understood (RQ

). This article’s

contribution goes toward acknowledging the use-

fulness of exploiting KGs in playing with lexical

semantics, envisioning applicability in self-learning

and education settings.

The rest of the article is structured as follows. Sec-

tion 2 clariﬁes the used terminology and overviews re-

lated work. Section 3 presents LexicalEngine. Sec-

tion 4 describes the performed qualitative and quan-

titative assessment, while Section 5 and Section 6 re-

port and discuss the collected results. Finally, the arti-

cle concludes with ﬁnal remarks and future directions.

2 BACKGROUND

This section deﬁnes terms that will be used in the fol-

lowing and overviews related work regarding sources

and tools to support lexical semantics via KGs.

2.1 Terminology

- In linguistics, a word sense is one of the mean-

ings of a word. For example, a dictionary may

have different senses of the word play, each hav-

ing a different meaning based on the context of the

word’s usage in a sentence. Play is generally de-

ﬁned as “be engaged in an activity for enjoyment

and recreation rather than a serious or practical

purpose”. However, it may be used as play a role

as a synonym of acting or in combination with a

sport to mean take part in (a sport).

- Hyponymy and hypernymy refer to relationships

between a general term and the more speciﬁc

terms that fall under the category of the general

term. For example, the colors red, green, blue, and

yellow are hyponyms. They fall under the general

term of color, which is the hypernym.

- In linguistics, meronymy is a semantic rela-

tion between a meronym denoting a part and a

holonym denoting a whole. In simpler terms,

a meronym is in a part-of relationship with its

holonym. For example, the ﬁnger is a meronym

for hand, which is its holonym. Holonymy is the

converse of meronymy.

- Synonymy refers to words that are pronounced

and spelled differently but share the same mean-

ing, e.g., words like car and automobile are close

synonyms that can be exchanged without compro-

mising the conceptual meaning despite having dif-

ferent stylistic functions. Similarly, one can also

have synonymic multi-word expressions as in the

case of play a role which is a synonym of acting.

Broaden Your Horizon! Play with Semantics via a Knowledge Graph-Based Approach

381

Table 1: Comparison between lexical KGs in terms of supported semantic relations and user interface.

BabelNet Wikidata DBnary DBpedia ConceptNet

Word senses ✓ ✓

Deﬁnition ✓ ✓ ✓ ✓ ✓

Synonyms ✓ ✓ ✓ ✓ ✓

Hypernymy ✓ ✓ ✓ ✓

Hyponymy ✓ ✓

Holonymy ✓ ✓ ✓

Meronymys ✓ ✓ ✓

Translation ✓ ✓ ✓ ✓

Images ✓ ✓ ✓

Emoticons ✓ ✓ ✓

Interface Web, Mobile & APIs SPARQL endpoint Web SPARQL endpoint Web

- This study refers to analogy in its role as rhetor-

ical ﬁgure. The analogical associative process is

a conceptual-based comparison between different

unit(s) of meaning, which, as a sort of semantic

trait, transfer from one concept to another. We

acknowledge that analogy is not included in the

set of lexical semantic relations. However, in line

with the procedure followed for the study, which

is based on a language-based associative evalua-

tion as the analogical comparison, it has been in-

cluded for the purpose of furthering a line of rea-

soning or drawing an inference.

2.2 Related Work

Literature offers some other examples of digital en-

vironments and web interfaces supporting lexical se-

mantics exploration. Fabula (Martines, 2017) scaf-

folds professional writers to author stories supported

by narrative suggestions pertinent to the user’s typed

words. Similarly, Communics (Rutta et al., 2020)

supports the creation of comics with predeﬁned sen-

tences to overcome the “blank page syndrome”. Nov-

elette (Addone et al., 2021) offers a suggestion mech-

anism of analogies, synonyms, and rhymes based on a

user-deﬁned word by querying BabelNet (Navigli and

Ponzetto, 2012) and WordAssociations

Focusing on KGs modeling lexicon and seman-

tics, Wikidata (Vrande

c and Kr

otzsch, 2014) and

DBpedia (Lehmann et al., 2015) are free, open, and

general-purpose KGs acting as central storage for

Wikimedia project data. Besides general content,

they can be easily queried to retrieve word synonyms,

deﬁnitions, and taxonomies. (Rodosthenous et al.,

2020; Rodosthenous et al., 2019) retrieve lexical re-

lations via ConceptNet (Speer et al., 2018) which

is a freely available commonsense KG and natural-

language-processing toolkit that supports many prac-

tical textual-reasoning tasks over real-world docu-

WordAssociations: https://wordassociations.net

ments and does not assume that words fall into

“synsets”, sets of synonyms that are completely in-

terchangeable. Nevertheless, WordNet (Miller, 1995)

assumes a synset-based word organization where a

set of synonyms and their related senses are linked

by semantic relations. Similarly, many other differ-

ent linguistic semantic networks have been proposed

directly relying on WordNet. To name a few, Babel-

Net (Navigli and Ponzetto, 2012) extends the notion

of synset to contain multilingual lexicalizations. DB-

nary (S

erasset, 2012) models multilingual lexical data

curating word senses and translations.

Table 1 compares different semantic networks re-

garding supported semantic relations and user in-

terfaces to explore them. Most of them provide

end-users with a web interface covering a wide

range of semantic relations. Following this line,

LexicalEngine automatically retrieves semantic as-

sociations and rhetorical ﬁgures by querying multi-

ple KGs at once while masking technical challenges

in posing SPARQL queries and providing end-users

with a uniﬁed result which can be explored via a web

interface. This approach detaches our research from

the exclusive quantitative and metric-based assess-

ment of KGs and includes the qualitative evaluation

based on speakers’ linguistic expectations which are

subsequently assessed later in the process.

3 LexicalEngine - SEMANTICS

VIA A KNOWLEDGE

GRAPH-BASED APPROACH

The proposed prototype, LexicalEngine, is a web-

based interface (freely available online

for test-

ing and demonstrative purposes) and API mecha-

nism

to let end-users and developers retrieve se-

mantic relations, such as word senses, term deﬁni-

LexicalEngine demo: https://lexical-client.surge.sh

APIs: https://lexical-engine.onrender.com

CSEDU 2024 - 16th International Conference on Computer Supported Education

382

tions, synonyms, homonyms, hypernyms, hyponyms,

meronyms, translations, images, and emoticons by

implicitly querying semantic networks as BabelNet,

Wikidata, DBpedia, DBnary, ConceptNet. The rest

of this section describes the interface and the working

mechanism of LexicalEngine, provides the reader

with technical details, and outlines some scenarios in

which LexicalEngine can be used.

Interface and Working Mechanism. A general

workﬂow is graphically reported in Figure 1.

Figure 1: Workﬂow of LiteralEngine to retrieve word-

related semantic relations by querying multiple KG and vi-

sually rendering results to end-users.

LexicalEngine requires a user-deﬁned word as in-

put. Supposing that end-users want to explore cuore,

the Italian translation of heart. In the learning mode,

as visible in Figure 2, LexicalEngine also requires

the semantic relation of interest. Let us hypothesize

that end-users are interested in looking for synonyms.

Each user-deﬁned request is transformed into a set of

SPARQL queries or API requests over all the KGs

supporting the conﬁgured semantic relation, accord-

ing to Table 1. In the case of synonyms, all the

supported KGs are queried in parallel. Results are

asynchronously collected by keeping the correspon-

dence between the source and results. Results are

parsed by a dedicated module, referred to as results

formatter in Figure 1 that takes care of visually ren-

dering results to end-users via unsorted lists or word

clouds, as widely explored in the literature to engage

learners (Shu et al., 2020). Back to our example,

Figure 2 reports the interface looking for Italian

synonyms of heart, showing that it can be interpreted

as a center, a metaphor for love, or the heart as an

organ.

Not all the semantic relations are covered by all

the KGs, as reported in Table 1. In that case, only

KGs supporting the user-requested semantic relation

are queried. Moreover, even if a service is supported

by a given source, the reply should not be given for

granted. For instance, it may happen that the user-

deﬁned word is not covered by all the sources, as hap-

pens in the reported use case. Looking at Figure 2,

only two KGs returned results, even if synonyms is a

service covered by all the queried KGs.

Technical Details. As a web-interface,

LexicalEngine logic is written in JavaScript.

The web interface relies on the LexicalEngine

API mechanism. Hence, any user request per-

formed by the visual interface is converted in an

LexicalEngine API call of the corresponding ser-

vice. For each semantic relation relying on multiple

KGs, KG queries are run in parallel. At the moment,

English, Italian, French, German, and Holland are

supported, but introducing other languages only re-

quires dealing with a few translations of the interface

components, such as the covered semantic relations

and setting options. However, it is crucial to check the

language coverage of the queried sources. Besides

the ones that are currently supported, sources can be

customized by verifying semantic relations covered

by the introduced KG and performing a dedicated

call accordingly. LexicalEngine distinguishes

between learning and inspiring mode as a result of

the performed evaluation. While in learning mode,

end-users are aware of the semantic relation that

links suggestions to the user-deﬁned request, in the

inspiring mode, suggestions derived from different

semantic relations and sources are just merged

together. No sorting is applied. The source code of

LexicalEngine is freely available on GitHub

Use Cases. The synonym lookup scenario is an ex-

ample to improve writing skills and avoid repeti-

tions. Moreover, in the evaluation stage, we also

asked participants about the application context of

LexicalEngine they foresee. Surprisingly, partici-

pants envisioned exploiting the proposed prototype to

edit text, learn semantics and novel word senses, and

argue. Moreover, such a system has been proposed

by teachers to overcome the blank page syndrome in

storytelling, where suggestions are not merely consid-

ered as an opportunity to avoid repetitions but should

inspire storytellers by letting them play on words and

explore semantic relations. Supposing that the ed-

ucators assign a topic to learners that can be sum-

marized in a single word, e.g., ﬁeld, and learners

can investigate suggestions automatically generated

by LexicalEngine, such as landing ﬁeld, battleﬁeld,

lines of business, or even ﬁeld of operation.

Interface: https://github.com/Crex99/lexicalClient.git,

API: https://github.com/Crex99/enchancingLexical.git

Broaden Your Horizon! Play with Semantics via a Knowledge Graph-Based Approach

383

Figure 2: LiteralEngine interface while looking for synonyms of heart (cuore in Italian).

4 EVALUATION

This section reports the methodology, participants,

and metrics of the performed evaluation aiming to

verify if KGs can automatically propose complete

(RQ

) and accurate (RQ

) word senses if compared

with those foreseen by (adult) Italian native speakers.

4.1 Participants

The evaluation has been performed asynchronously

by collecting participants’ replies via an anonymous

questionnaire administered via Google Forms

. The

link to the questionnaire has been sent to the authors’

contacts by encouraging them to share it. 27 Italian

native speakers joined the evaluation, voluntarily and

for free, 52% females. The sample is heterogeneous

in terms of regions of provenance: 70% from the

Campania region, 18% from the Lombardia region,

and 11% from the Lazio region. The sample is hetero-

geneous also concerning the age range ranging from

18 to over 50, mainly between 18-25 (41%) and over

50% (22%), while the remaining is equally distributed

over the remaining age range. The educational quali-

ﬁcation of the participants corresponds at least to the

diploma, and they are students or workers in a wide

https://doi.org/10.5281/zenodo.10687619

range of companies or institutions. Hence, the sam-

ple is heterogeneous regarding age, profession, edu-

cational title, and provenance to reduce as many bi-

ases as possible in the evaluation results. All the par-

ticipants assessed their interest in using sophisticated

language, grading it at least 3 out of 5. It emphasizes

that participants are interested in curating their vocab-

ulary even if they are not linguists by profession.

4.2 Methodology

The authors agreed on 10 Italian words for the eval-

uation, choosing among those frequently occurring

in written and spoken language provided with multi-

ple word senses. Literal translations of the evaluated

words are tree, field, organ, time/weather (both

tempo in Italian), star, heart, sun, fishing/peach

(both pesca in Italian), belief/cupboard (both

credenza in Italian), and wing. For each word, word

senses have been retrieved by querying BabelNet and

ConceptNet, while synonyms have been retrieved by

querying DBpedia, Wikidata, DBnary, BabelNet, and

ConceptNet. Participants are ﬁrst invited to think

about word senses for each proposed word without

any support. Subsequently, they are asked to evaluate

senses returned by LiteralSearch. The assessment

phase involves qualitatively evaluating the word units

from the speaker’s perspective. The consequent grad-

CSEDU 2024 - 16th International Conference on Computer Supported Education

384

ing of the proposed ones relies on the participants’

perspective as native speakers. The grading is based

on a three-value scale: 0 value if unrelated, 0.5 if par-

tially related, and 1 if wholly related. Finally, par-

ticipants are invited to quantitatively summarize their

opinions at an overall level regarding coverage of their

expectations, gaining inspiration, originality, and the

perceived completeness for the proposed word senses

on a 5-point Likert scale, and free to leave a qualita-

tive opinion via open-ended questions.

4.3 Metrics

As a result of the conducted protocol, for each of the

10 tested words, we have two pools of word senses:

those proposed by participants (WS

) and those re-

trieved via LexicalSearch (W S

). The two pools

are not disjoint. Hence, we have the following sets:

a set of word senses matching participants’ mean-

ing expectations and covered by LexicalSearch,

i.e., W S

∩ W S

;

a set of word senses matching participants’

meaning expectations but uncovered by

LexicalSearch, i.e., W S

\W S

;

a set of word senses proposed by LexicalSearch

but absent in participants’ lexical perspective, i.e.,

W S

\W S

According to the RQs, we measured both the cov-

erage of users’ expectations and the accuracy of un-

foreseen word senses through a qualitative analysis

of the data provided by the LexicalSearch. The

metrics chosen to evaluate the results obtained with

the assessment are based on cognitive sciences and

mental lexicon principles. Citing (Read, 1993), na-

tive speakers have remarkably stable patterns of word

association, which can be taken to reﬂect the sophis-

ticated lexical semantic networks they have devel-

oped through language acquisition. While the cover-

age has been computed by comparing S

with W S

the accuracy is computed by looking at grades as-

sessed by participants concerning S

. For each word,

we checked the overlap between users’ deﬁned word

senses and those retrieved by LexicalSearch. If

the ones returned by LexicalSearch do not cover

a user-deﬁned word sense (S

), we checked if syn-

onyms cover it. For each word sense proposed by

LexicalSearch and not foreseen by any evalua-

tor (S

), we collected users’ opinions on its accu-

racy and checked ofﬁcial Italian dictionaries (Devoto,

2016; Sabatini, 1997; De Mauro, 1999)

to see if the

word sense could be justiﬁed.

Grande dizionario italiano dell’uso accessible online

at https://dizionario.internazionale.it

5 RESULTS

Coverage (Connected to Rq

). All the participants

proposed at least a word sense for each word, with a

maximum of 5 proposals. By aggregating the word

senses proposed by all the participants, at least two

word senses are collected for each word. Figure 3 re-

ports the coverage of users’ expectations. It is worth

noting that the evaluated words are on the x-axis on

the chart represented in Figure 3, and its column mod-

els the percentage of users’ expectations satisﬁed by

word senses and those covered by synonyms and the

percentage of not satisﬁed expectations. Each user-

deﬁned word sense is weighted in terms of people

suggesting the same word sense.

Figure 3: Users’ expectations coverage.

Not all user-deﬁned proposals can be correctly

considered word senses as they also cover synonyms

and analogies. To name a few examples, some par-

ticipants attached brands to some words, such as All

stars to star or the name of a laundry detergent or

oranges to sun, or proper names of song authored by

the Italian singer Al Bano to sun or the one of the ro-

mances authored by De Amicis to Heart. After nam-

ing some word senses, evaluators usually moved to

metaphors or analogies referring to famous sayings.

For instance, some of them attached handsome to sun

referring to the saying You are as beautiful as the sun,

or they attached money to time referring to the saying

Time is money. Some proposals went in the direction

of analyzing the anatomy of the words. For instance,

some evaluators attached palindrome to wing as its

Italian translation, ala, is a palindrome. Those cases

negatively affected the coverage.

Accuracy (Connected to Rq

). Only in two cases

LiteralSearch returned a proposal not covered by

any evaluator. In both cases, evaluators considered

them not related. However, by exploring the reason

at the basis of the proposals, we discovered that the

suggestion attached to ala is hat as it is the name of

Broaden Your Horizon! Play with Semantics via a Knowledge Graph-Based Approach

385

the hat brim and the suggestion attached to pesca is

a speciﬁc ﬁshing net traditionally used in an Italian

region. Hence, both proposals are considered valid.

Overall Assessment. Participants graded inspira-

tion, satisfaction, originality, and coverage using a 5-

point Likert scale. The mean value is close to 4 and

the conﬁdence interval reaches a minimum close to

3.5 in all the cases, as reported in Table 2.

Table 2: Statistics of the overall grades auto-assessed by

participants reported in terms of minimum (min) and maxi-

mum (max) score, mean value, standard deviation (st.dev.),

and conﬁdence interval (C.I.) with α = 0.05.

Inspiration Satisfaction Originality Coverage

Range 1-5 1-5 1-5 1-5

Min 2.00 3.00 2.00 2.00

Mean 3.77 4.15 3.81 3.96

St. dev. 0.91 0.78 0.98 1.00

Max 5.00 5.00 5.00 5.00

C.I. [3.42, 4.12] [3.85, 4.45] [3.43, 4.19] [3.58, 4.34]

6 DISCUSSION

At Least Half of the Users’ Expectations Are Sat-

isﬁed. RQ

veriﬁes if word senses automatically re-

trieved by querying KGs cover users’ expectations be-

longing to Italian native speakers’ mental lexicon. As

visible in Figure 3, at least half of the users’ expec-

tations are satisﬁed in all cases, and the percentage

of coverage is at least equal to 75% if we consider

both word senses and synonyms. It lets us assessing

that word senses automatically retrieved by KGs guar-

antee sufﬁcient/high coverage of adult Italian native

speakers’ expectations. It becomes crucial in the case

of at-a-distance learning to ensure that (young) schol-

ars can be effectively supported by automatic tools

without requiring the support of adults.

Suggestions Require Rationales. LiteralSearch

proposed word senses not foreseen by native speak-

ers in just a few cases. However, those sugges-

tions have been considered inaccurate by participants

(RQ

), even if they can be justiﬁed by looking at Ital-

ian dictionaries. It suggested that it is not always ob-

vious to understand why any out-of-context word is

proposed as a word sense by LiteralSearch.Hence,

suggestions should be attached to exemplary sen-

tences or deﬁnitions to better understand the link be-

tween the user-selected word and the returned sug-

gestion. This insight will drive future enhancements

of the LiteralSearch prototype.

Not a Single Winner! Looking at the suggestions

returned by each KG, there is no single resource able

to satisfy users’ expectations fully. Hence, it is neces-

sary to combine results returned by multiple KGs to

reach the current coverage of users’ expectations. It

is worth noting that according to how the evaluation

has been performed, LiteralSearch compared with

each queried source. Since suggestions from multiple

KGs outperforms each resource queried in isolation,

LiteralSearch outperforms related works.

Inspiration Requires Relaxing Semantic Relations

Classiﬁcation. By both considering suggestions re-

turned as word senses and synonyms, we reached a

wider coverage of users’ expectations. In fact, na-

tive speakers also often jump among word senses,

analogies, and synonyms without separating seman-

tic relations. While in the learning setting, scholars

must have a clear deﬁnition and unambiguous exam-

ples of semantic relations, when we hypothesize to

use LiteralSearch to overcome the blank page syn-

drome, suggestions returned by different semantic re-

lations might be merged into a uniﬁed set to inspire

users without any hard classiﬁcations. Consequently,

learning and inspiration modalities have been sepa-

rately proposed in LiteralSearch.

7 CONCLUSIONS

This article explores the coverage and accuracy per-

ceived by Italian native speakers in playing with lex-

ical semantics via a KG-based approach. The per-

formed evaluation relies on LiteralSearch, a web-

based interface to retrieve semantic relations starting

from a user-deﬁned word, which reaches a good level

of coverage of users’ expectations (RQ

). Hence,

KGs are promising sources to learn and inspire orig-

inal meanings while letting them learn unforeseen

word senses (RQ

). Being an interface for searching

lexical semantic repositories in a very user friendly

fashion, our approach represent a starting point ap-

proach to word-networks exploration. As seen in

(Emma James and Henderson, 2023), it is well-

established that prior knowledge affects new learn-

ing and semantic neighbors also affect the acquisi-

tion of new words. Hence, the proposed approach

can broaden learners’ horizons by suggesting novel

meanings and word interpretations by exploiting word

senses and synonyms. Learning lexical semantics via

KGs has the potential to be useful also for young

learners or foreigners to explore and learn about word

meanings and to get inspiration to overcome the blank

page syndrome in storytelling. To conclude, this

CSEDU 2024 - 16th International Conference on Computer Supported Education

386

study envision a bidirectional human-to-data effort to

both enhance speakers linguistic competence and KG

data access. Moreover in detail, the former is justi-

ﬁed by a high-quality support to explore, retrieve, and

learn about word senses and lexical relations. On the

other side, synsets do not distinguish formalism, sug-

gesting as synonyms words that cannot really be used

as alternatives. Further human efforts are required to

explicitly tag words in synsets by their formalism.

Limitations and Future Directions. KGs are usu-

ally well-curated for English speakers. Even if the

queried KGs aim to provide end-users with multilin-

gual semantic networks, further effort should be in-

vested in wider coverage of other languages, such as

Italian. As already stated while discussing results,

suggestions require to be contextualized and justiﬁed.

It will result in enhancing LiteralSearch sugges-

tions by providing users with examples, descriptions,

or justiﬁcations in terms of retrieved semantic relation

in the direction of explaining proposed suggestions.

REFERENCES

Addone, A., Donato, R. D., Palmieri, G., Pellegrino, M. A.,

Petta, A., Scarano, V., and Serra, L. (2021). Novel-

ette, a usable visual storytelling digital learning envi-

ronment. IEEE Access, 9:168850–168868.

Bellini, P., Nesi, P., and Venturi, A. (2014). Linked open

graph: Browsing multiple SPARQL entry points to

build your own LOD views. Journal of Visual Lan-

guages & Computing, 25(6):703–716.

De Donato, R., Garofalo, M., Malandrino, D., Pellegrino,

M. A., and Petta, A. (2020). Education meets knowl-

edge graphs for the knowledge management. In

Methodologies and Intelligent Systems for Technology

Enhanced Learning. Workshops, pages 272–280.

De Mauro, T. (1999). Grande dizionario italiano dell’uso.

UTET, Torino.

Devoto, G. (2016). Il Devoto-Oli : vocabolario della lingua

italiana / Giacomo Devoto, Gian Carlo Oli ; a cura di

Luca Serianni e Maurizio Trifone. Le Monnier, Mi-

lano, 3. ed edition.

Emma James, M. Gareth Gaskell, G. M. J. T. and Hender-

son, L. M. (2023). Word learning in the context of

semantic prior knowledge: evidence of interference

from feature-based neighbours in children and adults.

Language, Cognition and Neuroscience, 38(2):157–

174.

Lehmann, J., Isele, R., Jakob, M., Jentzsch, A., Kon-

tokostas, D., Mendes, P. N., Hellmann, S., Morsey,

M., van Kleef, P., Auer, S., and Bizer, C. (2015). Db-

pedia - A large-scale, multilingual knowledge base ex-

tracted from wikipedia. Semantic Web, 6(2):167–195.

Martines, F. (2017). Fabula. https://www.scribis.it/fabula/

index.html. [Online, Last access January 2024].

Miller, G. A. (1995). Wordnet: A lexical database for en-

glish. Communications ACM, 38(11):39–41.

Mitchell, T. M., Shinkareva, S. V., Carlson, A., Chang,

K. M., Malave, V. L., Mason, R. A., and Just,

M. A. (2008). Predicting human brain activity as-

sociated with the meanings of nouns. Science,

320(5880):1191–1195.

Navigli, R. and Ponzetto, S. P. (2012). BabelNet: The au-

tomatic construction, evaluation and application of a

wide-coverage multilingual semantic network. Artiﬁ-

cial Intelligence, 193:217–250.

Pustejovsky, J. (2016). Lexical semantics, page 33–64.

Cambridge University Press.

Read, J. (1993). The development of a new measure of l2

vocabulary knowledge. Language Testing, 10(3):355–

371.

Rodosthenous, C., Lyding, V., K

onig, A., Horbacauskiene,

J., Katinskaia, A., ul Hassan, U., Isaak, N., Sangati,

F., and Nicolas, L. (2019). Designing a prototype ar-

chitecture for crowdsourcing language resources.

Rodosthenous, C., Lyding, V., Sangati, F., K

onig, A.,

ul Hassan, U., Nicolas, L., Horbacauskiene, J., Katin-

skaia, A., and Aparaschivei, L. (2020). Using crowd-

sourced exercises for vocabulary training to expand

ConceptNet. In Proceedings of the Twelfth Language

Resources and Evaluation Conference, pages 307–

316. European Language Resources Association.

Rutta, C. B., Schiavo, G., Zancanaro, M., and Rubegni, E.

(2020). Collaborative comic-based digital storytelling

with primary school children. In Proceedings of the

19th International Conference on Interaction Design

and Children, IDC, pages 426–437. ACM.

Sabatini, F. (1997). DISC : Dizionario italiano Sabatini

Coletti. Giunti, Firenze, 1 edition.

erasset, G. (2012). Dbnary: Wiktionary as a lmf based

multilingual rdf network. In Proc. of the 8th Interna-

tional Conference on Language Resources and Evalu-

ation (LREC), pages 2466–2472.

Shu, X., Wu, J., Wu, X., Liang, H., Cui, W., Wu, Y., and

Qu, H. (2020). Dancingwords: exploring animated

word clouds to tell stories. Journal of Visualization,

24:85–100.

Sowa, J. F. (2006). Semantic Networks. American Cancer

Society.

Speer, R., Chin, J., and Havasi, C. (2018). Conceptnet 5.5:

An open multilingual graph of general knowledge.

Vargas, H., Aranda, C. B., Hogan, A., and L

opez, C. (2019).

RDF explorer: a visual SPARQL query builder. In

International Semantic Web Conference (ISWC), vol-

ume 11778, pages 647–663. Springer.

Vrande

c, D. and Kr

otzsch, M. (2014). Wikidata: a free

collaborative knowledgebase. Communications of the

ACM, 57(10):78–85.

Wang, A., De Deyne, S., McKague, M., and Perfors, A.

(2022). Core words in semantic representation. In

Proceedings of the Annual Meeting of the Cognitive

Science Society, volume 44.

Broaden Your Horizon! Play with Semantics via a Knowledge Graph-Based Approach

387