The growing interest in analysing emotions in Italian texts has led to the development of various afective resources, often independently constructed and lacking interoperability. To address this fragmentation, we adopt a Linked Open Data (LOD) approach. This paper presents three main contributions: (1) the release of Sentix 3.0, a revised and enriched polarity lexicon for Italian, together with two derivatives (MAL and WMAL) that address morphological and word frequency variation issues; (2) a new quartile-based methodology to discretize continuous polarity scores; and (3) the linking of Sentix 3. 0 and ELIta, an annotated emotion lexicon with categorical labels, to the LiITA Lemma Bank using standard ontologies (OntoLex-Lemon, MARL) and the newly introduced elita ontology for categorical emotion representation based on Plutchik's Wheel. At the heart of the linking is the word, the central node for aligning diferent lexical resources.
ian text has led to many specialized linguistic resources, such as lexicons and corpora. However, these resources are often independently built, with varying annotation methods, severely limiting their interoperability.
To address these challenges, we propose a Linked Open
Data (LOD) approach. Inspired by projects such as LiLa
[
This article presents three original contributions: • Resource linking and ontology development: we align ELIta, a lexicon annotated with categorical emotions, and the updated Sentix 3.0 with the LiITA project’s lemma bank, using the MARL ontology to formally represent afective relations. Linking is achieved through the use of
port consistent data representation and semantic
integration. To this end, the W3C has introduced foundational
In recent decades, the analysis of emotions has led to the technologies such as RDF and OWL. Building upon these
development of numerous linguistic resources, particu- standards, the MARL ontology has been developed to
larly afective lexicons and annotated datasets. Despite formally describe opinions and to associate them with
the pervasive impact of extensive end-to-end language contextual information(such as opinion topic, features
models (LLMs), afective and other annotated lexicons described in the opinion, etc.).
remain a dynamic area of research within computational This infrastructure underpins the design of
LatinAflinguistics, maintaining its vitality.They are particularly fectus [17], a polarity lexicon originated within the LiLa
efective in the development of hybrid approaches [
These lexical resources are designed to be FAIR (Find- schema, the lexicon itself is modeled as an instance of able, Accessible, Interoperable, Reusable) and transpar- class E31 Document13 from the CIDOC Conceptual Refent: as every system decision based on a lexicon can erence Model (CRM), an ontology designed to represent be traced back to specific entries, ensuring interpretabil- entities and relationships in the cultural heritage domain. ity. Moreover, unlike computationally demanding large In parallel, LatinAfectus is also declared as a lexicon-type language models, lexicons are accessible to a broader object following the LInguistic MEtadata (lime) module research community, promoting the democratization of of Ontolex. research. Lexical entries in the resource are connected to the lex
For the Italian language, several lexicons [7, 8] and icon through the lime:entry property and are instancorpora [9, 10] annotated according to the sentiment tiated as objects of the class ontolex:LexicalEntry. [11] or emotion associated with the words or texts have Each lexical entry is associated with a label, an been created and published, created by both automatic ontolex:canonicalForm (linking it to the lemma in [12, 13] and manual methods [14]. However, much like the LiLa Knowledge Base), and an ontolex:sense, the broader landscape of digital resources, there is a sig- which captures its meaning. Since LatinAfectus nificant lack of clear, standardized guidelines for afective is concerned with prior polarity, each lexical entry resources, resulting in poor interoperability. Each lexi- has only one sense, modeled as an instance of the con or corpus is typically developed independently, and class ontolex:LexicalSense. This sense is further although they often share similar objectives, such as sen- described by a label, the relation marl:hasPolarity, timent analysis or emotion detection, they difer widely and the property marl:polarityValue. The in terms of annotation categories, selected lemmas, and marl:hasPolarity relation links the sense to a annotation methodologies. category within marl:Polarity (namely positive,
In addition, the availability of these resources is highly negative, or neutral) while marl:polarityValue fragmented. They are typically hosted on disparate plat- assigns a numerical sentiment score from the predefined forms and, even when co-located within a shared repos- set: 1.0, 0.5, 0.0, -0.5, or -1.0. itory infrastructure, interoperability between them re- The integration of emotions with Linked Open Data mains limited or nonexistent. is not a new challenge. Relevant contributions in this
A notable attempt to address these issues can be found area include Iglesias et al. [20] and Sanchez-Rada et al.
in the LiLa project [15], which pioneered a new model [21]. These works introduced Onyx, that is an RDF
onof interoperability for Latin linguistic resources. This tology designed to represent emotions in textual content
vision has since been adopted for Italian by the LiITA within the framework of the Semantic Web and Linked
project [
Central to the LOD paradigm is the Semantic Web, EmotionCategory resources, which can reflect diferwhich promotes the use of interoperable and interlink- ent psychological models, and may include numerical able data schemas for online information. These schemas, values for emotion intensity and afective dimensions commonly referred to as ontologies or vocabularies, sup- such as arousal and valence. Onyx also enables the assomarl:hasPolarityValue
real number ciation of emotional annotations with external resources or entities.
While Onyx is conceptually rich and expressive, its structure is relatively complex, as it is designed to represent the emotional content of texts.
Opinion marl:opinion marl:aggregatedopinion m a r l : h a s P o l a r i t y
marl:polarity
marl:neutral
marl:negative
marl:positive try), modeled as individuals of the Class lila:Lemma2, which is a subclass of ontolex:Form, originally created for the LiLa project, and adopted in the LiITA Lemma Bank accordingly.
The Lemma Bank was initially populated with approximately 94, 000 lemmas derived from the online version of the Nuovo De Mauro dictionary, after excluding about 13, 000 multi-word expressions.
designed to foster interoperability among various Italian linguistic resources by leveraging the principles of Linked Open Data (LOD). ELIta ELIta [23], a recently introduced linguistic re
The core component of the LiITA KB is its Lemma source, comprises a lexicon annotated via crowdsourcing. Bank, a comprehensive collection of Italian lemmas. This annotation scheme incorporates both basic emoThese lemmas, which are conventional lexical citation tions, based on Plutchik’s model [24] with corresponding forms used across linguistic resources, serve as the cen- association degrees, and the VAD (Valence, Arousal, Domtral connection point for interlinking both lexical and inance) emotional dimensions, thus including sentiment textual data. The architecture of the LiITA KB mirrors (valence) [25]. The lexical items, primarily sourced from that of the LiLa KB for Latin1, operating under the as- the De Mauro dictionary [26], were annotated in isolasumption that all interoperable (meta)data sources within tion. To date, four distinct versions of this lexicon have the KB are word-related. been released [8]:
Following the Linked Data paradigm, LiITA achieves conceptual interoperability among its distributed re- • RAW: Full annotations with demographic data. sources by applying vocabularies commonly used within • GOLDEN: Selection of 5 consistent annotations the Linguistic Linked Open Data (LLOD). For the Lemma + majority-vote golden label. Bank specifically, this means adopting the vocabulary • INTENSITY: Aggregated intensities from defined by OntoLex-Lemon [ 18], one of the most widely GOLDEN; includes auto-generated "love" and used models for representing and publishing lexical re- "neutral". sources as Linked Data. The Lemma Bank of LiITA is a collection of canonical forms as intended by the Ontolex- • BINARY: Binary version of INTENSITY using Lemon ontology (ontolex:canonicalForm property, 0.50 threshold. the conventionally chosen representation for the entire set of inflected forms belonging to a particular lexical en1https://lila-erc.eu/data-page/ 2http://lila-erc.eu/ontologies/lila/Lemma
richment and update of the TWITA corpus to 2022, it was decided to harmonize all three resources, which were developed years apart; this involved not only updating Sentix is an afective lexicon for the Italian language cre- WMAL’s weights but also rectifying the interconnections ated in 2013 by aligning several lexical resources, namely among the three to ensure overall consistency. In particSentiWordNet [27] and MultiWordNet [28] through ular, during the transition from Sentix to MAL, i.e., from WordNet [29, 30]. lemmas to inflected forms, new identical forms with dif
The first version [ 11] was built by transferring the ferent scores and diferent senses are inevitably created. synset annotations from SentiWordNet to the respective It was decided to manage these in a coordinated manner, Italian synsets of MultiWordNet, using an automatic map- primarily by revising Sentix. ping [cf. 31] to resolve the partial alignment of Senti- A key part of this harmonization involved a deeper WordNet’s indices (based on WordNet 3.0), and those re-examination of the foundational Sentix lexicon. This of MultiWordNet (based on WordNet 1.6) [cf. 32]. The specific efort, working backwards from the original Senperformance of the lexicon was evaluated with data man- tix version, led to an expansion in the number of linkable ually annotated by independent human judges. synsets between SentiWordNet (SWN) and MultiWord
The subsequent version, Sentix 2.0, aggregated the po- Net (MWN), which will be made available3. larity scores of the diferent senses of a lemma into a The revision subsequently involved external resources single score (-1 to 1), using a weighted average with the and supervised phases to identify forms present in Sentix sense frequencies calculated on the annotated SemCor that could generate unexpected duplicate entries in MAL corpus [33, 34]. This version, which includes 41,800 dif- (i.e., entries that could be either base forms or inflected ferent lemmas, has been available on GitHub in the R forms), and to expand Sentix itself by back-linking lempackage sentixR since 2019 [35], and has been used in mas present in Morph-it! traceable to pre-existing entries various research projects over these years. (717 entries). Finally, neutral terms from SentiWord
Other lexical resources have been developed from Sen- Net not already present in Sentix were added (22, 117 tix. The first derived lexicon was MAL [ 36], created by entries). The new lexicon ultimately contains 63, 660 expanding Sentix 2.0 with inflected forms derived from lemmas [41].
Morph-it!, a morphological lexicon for the Italian lan- The resources used for the update were: SentiWordNet guage [37]. This was intended to address the inherent and MultiWordNet, using the Open Multilingual Wordnet dificulties of lemmatization in Italian sentiment analysis (https://omwn.org/) [42, 43], the TreeTagger library [44], - stemming from the language’s morphological complex- and, of course, Morph-it! itself. ity and the limitations of available NLP tools - which are particularly exacerbated when analyzing user-generated 4. Methodology content from social networks (often containing spelling errors, jargon, irregularities, and non-standard syntactic Following the methodology established for the afective structures). MAL - which inherits Sentix 2.0’s scores - has lexicon of Latin within the LiLa project, LatinAfectus, the shown to achieve an improvement in overall sentiment same approach was adopted for the ELIta and Sentix 3.0 analysis performance. lexicons by using the MARL ontology [19] to represent
A second derived resource is WMAL [34], a dictionary polarity properties.
of inflected forms like MAL, where MAL’s scores were In this context, neutrality thresholds for polarity
larecalculated by weighting the original scores inversely bels were defined within the range between the first and
with respect to their words frequencies in the TWITA third quartiles of the lexicon. The weighting
methodolcorpus [38] by using the inversed version of the Zipf scale ogy and polarity calculation based on the new updated
measure [39] that consists in a logarithmic scale based on version of WMAL are key to Sentix 3.0’s polarity
categorthe well-known Zipf law of word frequency distribution ical classification. In this respect, the first and the third
[40]. The two main WMAL lexical and methodological quartile-based interval [1; 3] was calculated on the
speculations were respectively to give more weight to WMAL scores to better individualize the neutral
threshlow frequent terms and to reduce the polarity imbalance olds and consequently the positive and negative
polarwhen using parametric threshold values to assign po- ity values outside. The quartile-based strategy to detect
larity classes: even small variations in these values in neutral scores has already provided promising results
fact showed to have an opposite impact on the ability
to correctly predict negative versus positive polarity [
On the occasion of WMAL’s update, driven by the en
Italian synsets, due to changes between subsequent versions of Wordnet [cfr. 32], often prevents their retrieval when attempting to map from SWN to MWN through standard tools, such as the OMW
Python package
elita:Aspettativa
elita:Fiducia
elita:Disgusto
elita:Sorpresa
elita:word elita:HasEmotion
elita:Emotion
elita:Amore Joy elita:Gioia
elita:Tristezza
elita:Paura
elita:Rabbia
across annotated corpora[
The other relationship used to describe Sentix 3.0 data owl:ObjectProperty links a subject (e.g., a word or is marl:hasPolaritValue in which values are contin- expression) to an instance of elita:Emotion, thereby uous from -1 to 1. expressing the emotional content attributed to that ele
For example, the lemma in ELIta "abbandonare" ment (Fig. 2). was annotated as marl:hasPolarity "Negative", and Despite its simplicity, the ontology maintains conmarl:hasPolarityValue "− 0.833". ceptual continuity with Onyx through the use of the
However, since the MARL ontology does not provide elita:HasEmotion property, which functionally corspecific properties for representing categorical emotions, responds to onyx:hasEmotionCategory. Moreover, and the structure of Onyx is relatively complex, the elita the emotional categories defined in elita reflect the anontology is introduced to fill this gap with a simpler and notations present in the ELIta lexical resource, ensuring more transparent model, specifically designed for anno- alignment with existing linguistic data. tating individual words with emotion categories. This design enables the annotation and querying of
The elita ontology has been developed to represent resources using fine-grained emotional categories, efeccategorical emotions in a structured and interoperable tively complementing polarity-based approaches in the manner, with a particular focus on applications in lin- representation of lexical entries. At the same time, it guistic and sentiment analysis. In contrast to the MARL ensures interoperability with existing emotion ontoloontology, which primarily addresses sentiment polarity gies while providing a lightweight, application-oriented (positive, neutral, negative), and Onyx, which incorpo- model specifically tailored to lexical annotation.
representation of the lexical resources, the lexicons were converted into RDF format. The linking procedure involved mapping lexical entries, each associated with a unique URI, to their corresponding lemmas within the LiITA’s Lemma Bank. The results of this linking process are presented in the following sections (an example of a linked entry is shown in Figure 3). 5.1. Linking ELIta processing.
The lexical entries that presented a one-to-one match were associated with the lemmas in the Lemma Bank using the Ontolex ontology and the relation ontolex:CanonicalForm.
60
Linking Results − ELIta and the LiITA Lemma Bank
70% ) % (40 e g a t cneeP r 20
compassing lemmas (as defined by LiITA), emojis, and multi-word expressions like "a malincuore" (reluctantly).
Notably, ELIta’s lexical entries include in some cases 0 2.8% both masculine and feminine forms of adjectives and one to one one to many none nouns. This inclusion aimed to facilitate the assessment of gender-based perceptual diferences, particularly when Figure 4: Percentages of linking between ELIta and LiITA morphological gender is the sole distinguishing factor. Lemma Bank.
To align ELIta with LiITA’s Lemma Bank, emojis were initially removed. The remaining lexical entries were The high percentage of matches between ELIta and then compared and linked to LiITA’s lemma URI where LiITA (shown in Fig. 4) is mainly due to the use of the feasible. This process yielded 4, 705 ELIta words that ex- same lexical source as the backbone of both resources. hibited a one-to-one match with lemmas or hypolemmas In particular, both rely heavily on the Nuovo De Mauro in LiITA (as shown in Table 2). The remaining approxi- dictionary: about 70 percent of ELIta’s entries come from mately 2, 000 entries, however, matched multiple lemmas the Nuovo Vocabolario di Base [26], while LiITA’s Lemma within the lemma bank, or none, necessitating further Bank is based on the lexical base of an online version of 27.1%
the dictionary Nuovo De Mauro4. The 190 lexical entries in ELIta that do not have a match in LiITA are mainly multi-word expressions or idiomatic phrases, which were not included in the construction of the LiITA lemma bank.
In addition, some entries without correspondence correspond to feminine forms that were explicitly annotated in ELIta. These forms often carry diferent afective annotations than their male counterparts, which are usually used as the canonical form of the lemma in LiITA. Consequently, these female variants were not included in the linking process.
One-to-many correspondences, accounting for 27.1 percent of the total, are largely attributable to words that correspond to multiple parts of speech (PoS) in LiITA, and thus correspond to multiple lemmas. In ELIta, the annotation process, following the methodology described in [14], did not specify the part of speech for each entry. As a result, PoS-based disambiguation is not possible in the current version of the resource. 4https://dizionario.internazionale.it/
5.2. Linking Sentix 3.0 mentioned above is also the use of standards such as RDF and SPARQL to provide useful information on what is The same procedure was applied to Sentix 3.0, with each identified by a URI, for the purpose of (meta)data repentry linked to a lemma from the lemma bank whenever resentation and retrieval. If RDF (Resource Description possible. The results revealed that most entries were new Framework) [45] is the data model underlying the Seto the lemma bank and therefore had no matching lemma. mantic Web, SPARQL5 is a query language for (meta)data This is primarily because the lexicon contains a large represented in RDF. number of multi-word expressions, such as "oggetti per Integrating afective resources into LiITA significantly la casa" (household items), "vedova nera" (black widow), enhances its query capabilities, allowing for advanced "dificoltà di apprendimento" (learning dificulties). As in SPARQL interrogations across LiITA’s own data and its the case of ELIta, the one-to-many links are due to the linked lexical and textual resources. presence of lemmas in Sentix that may belong to diferent For instance, it is possible to6: parts of speech in LiITA Lemma Bank. To address the one-to-many mappings, a possible solution would be to Query 1. Retrieve the Distribution of Emotions in disambiguate entries based on part of speech. However, ELIta It’s possible to query the distribution of emotions the current version of Sentix consists of isolated lexical as defined within the ELIta lexicon. the SPARQL query entries, and PoS tagging typically relies on contextual counts the number of lemmas associated with each emoinformation. Since such context is not available in the tion label in the ELIta resource. By linking lemmas from lexicon, it is not currently possible to assign PoS labels the LiITA Lemma Bank to their corresponding emotion reliably. While previous versions of Sentix included some annotations in ELIta, the query retrieves the textual label PoS information, this is not present in the current release. of each emotion and aggregates the lemmas accordingly. As a result, any attempt at disambiguating part of speech The results are grouped by emotion label and sorted in would be arbitrary. descending order based on lemma count.
Nevertheless, successful linking was achieved in 36% (as shown in Fig. 5) of cases, with 22,946 lemmas matched on a one-to-one basis (see Table 3). 5https://www.w3.org/TR/rdf-sparql-query/ 6The SPARQL queries used to generate these examples are available in the appendix and can be executed via the LiITA endpoint.
Table 4 shows the result of such distribution, with "Aspettativa" (Anticipation) being the most frequent emotion. Query 2. Retrieve Polarity Distribution in Sentix 3.0: The SPARQL query counts the number of lemmas associated with each polarity label (e.g., "Positive", "Negative", "Neutral") in the Sentix 3.0 resource. By linking lemmas from the LiITA Lemma Bank to their corresponding polarity annotations in Sentix 3.0, the query retrieves the textual label of each polarity and aggregates the lemmas accordingly. The results are grouped by polarity and sorted in descending order based on lemma count.
This query type helps determine the count of Italian words marked as positive, negative, or neutral based on lemmas shared between Sentix 3.0 and LiITA. Table 5 provides the result, indicating a higher number of neutral lemmas. Query 3. Return the average Sentix polarity score for each emotion annotated in ELIta: Another possible query can be used to identify the most negative emotion in ELIta based on Sentix 3.0 Polarity Value.
The query retrieves the average Sentix polarity value for each emotion label found in the ELIta resource. It does so by: 1. Linking lemmas from LiITA (via lila:lemma) to
their associated emotions in ELIta. 2. Retrieving corresponding polarity values from
Sentix.
Amore (Love) Gioia (Joy)
Fiducia (Trust) Aspet ativa (Anticipation) n o ito Sorpresa (Surprise) m E
Paura (Fear) Tristezza (Sadness)
Rabbia (Anger) Disgusto (Disgust)
Polarity 0.2 0.0 −0.2
consistently emerges as the most negative emotion when analyzing average polarity scores for emotions based on Sentix 3.0 annotations. This is visualized in Figure 6, where the color gradient reflects polarity intensity.
Average Polarity per Emotion
Negative
Neutral
Positive −0.4 −0.2 Average Polarity (Sentix) 0.0 0.2 Query 4. Determine the polarity of lemmas annotated with contrasting emotions: Since words in ELIta can be associated with multiple emotions, we explored instances where Joy and Sadness, two emotions of opposing polarities, co-occurred in the annotation of the same lemma.
As a first step, we queried the number of words in ELIta that are simultaneously associated with both Joy and Sadness, grouped by their Sentix polarity label.
More specifically, the query: 1. Selects lemmas from the LiITA Lemma Bank that are linked to ELIta entries. 2. Filters for those lemmas tagged simultaneously with the emotions elita:Gioia (Joy) and elita:Tristezza (Sadness). 3. Matches these lemmas to their corresponding Sentix 3.0 polarity labels (Positive, Negative, or Neutral). 4. Counts how many lemmas fall into each polarity category.
5. Sorts the output by descending frequency.
and "umore" (mood), identified through a dedicated query7, consistently exhibited a negative polarity according to Sentix 3.0. Their respective polarity values are shown in Table 7.
The examples showcased a range of queries that extract information not only from individual resources but also by integrating data from both Sentix 3.0 and ELIta, highlighting how interoperability enables more comprehensive analysis of afective lexical information.
afective lexical resources, Sentix 3.0 and ELIta, published according to the Linked Open Data (LOD) paradigm. It also presents the new version of the Sentix 3.0 resource and its derivatives, MAL and WMAL, now available on GitHub. Additionally, the ontology developed for rendering the ELIta emotional lexicon within the Linguistic Linked Open Data (LLOD) framework is introduced.
Both resources have been linked to the LiITA Lemma Bank, thus contributing to and enriching the possibilities of investigation and promoting interoperability among LLOD resources.
Through the LOD paradigm, these resources also support interdisciplinary applications, particularly within the digital humanities (e.g., cultural heritage, social sciences), where linguistic knowledge graphs find practical applications (e.g., through frameworks like CIDOC CRM or LiLa).
Nonetheless, this work represents only the initial phase of fully aligning these afective resources with LiITA. Future eforts will focus on resolving one-to-many mappings and incorporating new lemmas into the LiITA Lemma Bank where applicable. 7The corresponding query is provided in the appendix.
erary texts or interlingual evaluations between regional variants of Italian, can be addressed through interoperability in an ecosystem where the word is the basis of knowledge.
— Next Generation EU, Mission 4 Component 1 CUP J53D2301727OOO1 — PRIN 2022 PNRR project LiITA: Interlinking Linguistic Resources for Italian via Linked Data. features and machine learning for sentiment anal- in the lila knowledge base of linguistic resources for ysis, IEEE Access 12 (2024) 81793–81812. doi:10. latin, in: Proceedings of the 2020 Globalex Work1109/ACCESS.2024.3382836. shop on Linked Lexicography, 2020, pp. 20–28. [6] F. Koto, T. Beck, Z. Talat, I. Gurevych, T. Baldwin, [18] J. P. McCrae, J. Bosque-Gil, J. Gracia, P. Buitelaar, Zero-shot sentiment analysis in low-resource lan- P. Cimiano, The ontolex-lemon model: developguages using a multilingual sentiment lexicon, in: ment and applications, in: Proceedings of eLex Y. Graham, M. Purver (Eds.), Proceedings of the 2017 conference, 2017, pp. 19–21. 18th Conference of the European Chapter of the [19] A. Westerski, C. A. Iglesias, F. T. Rico, Linked Association for Computational Linguistics (Volume opinions: Describing sentiments on the structured 1: Long Papers), Association for Computational Lin- web of data, in: SDoW@ISWC, 2011. URL: https: guistics, St. Julian’s, Malta, 2024, pp. 298–320. URL: //api.semanticscholar.org/CorpusID:10397646. https://aclanthology.org/2024.eacl-long.18/. doi:10. [20] C. Iglesias, J. Sánchez-Rada, G. Vulcu, P. Buite18653/v1/2024.eacl-long.18. laar, Linked data models for sentiment and [7] F. Tamburini, Neural Models for the Automatic Pro- emotion analysis in social networks, in: F. A.
cessing of Italian, Pàtron, Bologna, 2022. Pozzi, E. Fersini, E. Messina, B. Liu (Eds.), [8] E. Di Palma, ELIta (emotion lexicon for italian), 2024. Sentiment Analysis in Social Networks, MorILC-CNR for CLARIN-IT repository hosted at Insti- gan Kaufmann, Boston, 2017, pp. 49–69. URL: tute for Computational Linguistics "A. Zampolli", https://www.sciencedirect.com/science/article/pii/ National Research Council, in Pisa. B9780128044124000048. doi:https://doi.org/ [9] R. Sprugnoli, Multiemotions-it: A new dataset for 10.1016/B978-0-12-804412-4.00004-8. opinion polarity and emotion analysis for italian, [21] J. F. Sánchez-Rada, C. A. Iglesias, Onyx: A in: Proceedings of the Seventh Italian Conference linked data approach to emotion representation, on Computational Linguistics (CLiC-it 2020), Ac- Information Processing Management 52 (2016) cademia University Press, Torino, 2020, pp. 402–408. 99–114. URL: https://www.sciencedirect.com/ URL: http://hdl.handle.net/10807/165687. science/article/pii/S030645731500045X. doi:https: [10] F. Bianchi, D. Nozza, D. Hovy, FEEL-IT: Emotion //doi.org/10.1016/j.ipm.2015.03.007, and sentiment classification for the Italian language, emotion and Sentiment in Social and Expressive in: Proceedings of the Eleventh Workshop on Com- Media. putational Approaches to Subjectivity, Sentiment [22] F. Burkhardt, C. Pelachaud, B. W. Schuller, E. Zoand Social Media Analysis, Association for Compu- vato, EmotionML, Springer International Publishtational Linguistics, Online, 2021, pp. 76–83. ing, Cham, 2017, pp. 65–80. URL: https://doi. [11] V. Basile, M. Nissim, Sentiment analysis on Italian org/10.1007/978-3-319-42816-1_4. doi:10.1007/ tweets, in: Proceedings of the 4th Workshop on 978-3-319-42816-1_4.
Computational Approaches to Subjectivity, Senti- [23] E. Di Palma, ELIta: A new Italian language resource ment and Social Media Analysis, 2013, pp. 100–107. for emotion analysis, in: F. Dell’Orletta, A. Lenci, [12] O. Araque, L. Gatti, J. Staiano, M. Guerini, De- S. Montemagni, R. Sprugnoli (Eds.), Proceedings pechemood++: A bilingual emotion lexicon built of the 10th Italian Conference on Computational through simple yet powerful techniques, IEEE Linguistics (CLiC-it 2024), CEUR Workshop ProTransactions on Afective Computing 13 (2022) 496– ceedings, Pisa, Italy, 2024, pp. 297–307. 507. doi:10.1109/TAFFC.2019.2934444. [24] R. Plutchik, A general psychoevolutionary theory [13] L. Passaro, L. Pollacci, A. Lenci, Item: A vector space of emotion, in: Theories of emotion, Elsevier, 1980, model to bootstrap an italian emotive lexicon, in: pp. 3–33.
Second Italian Conference on Computational Lin- [25] J. A. Russell, A circumplex model of afect., Journal guistics CLiC-it 2015, Academia University Press, of Personality and Social Psychology 39 (1980) 1161– 2015, pp. 215–220. 1178. [14] M. Montefinese, E. Ambrosini, B. Fairfield, N. Mam- [26] T. De Mauro, Il nuovo vocabolario di marella, The adaptation of the afective norms for base della lingua italiana, Internazionale english words (anew) for italian, Behavior Research (2016). URL: https://www.internazionale.
Methods 46 (2014) 887–903. it/opinione/tullio-de-mauro/2016/12/23/ [15] M. Passarotti, The project of the index thomisticus il-nuovo-vocabolario-di-base-della-lingua-italiana. treebank, in: Digital Classical Philology, De Gruyter [27] S. Baccianella, A. Esuli, F. Sebastiani, Sentiwordnet Saur, 2019, pp. 299–320. 3.0: An enhanced lexical resource for sentiment [16] T. Berners-Lee, J. Hendler, O. Lassila, The semantic analysis and opinion mining., in: Lrec, volume 10, web, Scientific american 284 (2001) 34–43. Valletta, 2010, pp. 2200–2204. [17] F. Mambrini, M. Passarotti, Representing etymology [28] E. Pianta, L. Bentivogli, C. Girardi, MultiWordNet: Developing an aligned multilingual database, in: [41] V. Basile, M. Nissim, C. Bosco, M. Vassallo, First International Conference on Global WordNet, G. Gabrieli, A. Vardanega, Sentix, 2025. doi:10. 2002, pp. 293–302. 5281/zenodo.15609215. [29] C. Fellbaum, Towards a representation of idioms [42] F. Bond, M. W. Goodman, E. Rudnicka, L. M. in WordNet, in: Usage of WordNet in Natural Lan- da Costa, A. Rademaker, J. P. McCrae, Documentguage Processing Systems, 1998. ing the Open Multilingual Wordnet, in: G. Rigau, [30] C. Fellbaum, WordNet: An Electronic Lexical F. Bond, A. Rademaker (Eds.), Proceedings of the
Database, MIT press, Boston, 1998. 12th Global Wordnet Conference, Global Word[31] J. Daude, L. Padro, G. Rigau, Mapping WordNets net Association, University of the Basque Country, Using Structural Information, 2000. URL: https: Donostia - San Sebastian, Basque Country, 2023, pp. //arxiv.org/abs/cs/0007035. doi:10.48550/arXiv. 150–157.
cs/0007035. [43] F. Bond, K. Paik, A survey of wordnets and their [32] E. Kafe, How Stable are WordNet Synsets?, in: LDK licenses, in: Proceedings of the 6th Global WordNet
Workshops, 2017, pp. 113–124. Conference (GWC 2012), Matsue, 2012, pp. 64–71. [33] H. Langone, B. R. Haskell, G. A. Miller, Annotating [44] H. Schmid, M. Baroni, E. Zanchetta, A. Stein, Il WordNet, in: Proceedings of the Workshop Fron- sistema ‘tree-tagger arricchito’–The enriched Treetiers in Corpus Annotation at HLT-NAACL 2004, Tagger system, IA Contributi Scientifici 4 (2007) Association for Computational Linguistics, Boston, 22–23.
Massachusetts, USA, ????, pp. 63–69. [45] E. J. Miller, An introduction to the resource descrip[34] M. Vassallo, G. Gabrieli, V. Basile, C. Bosco, Polarity tion framework, Journal of library administration Imbalance in Lexicon-based Sentiment Analysis, 34 (2001) 245–255. in: F. Dell’Orletta, J. Monti, F. Tamburini (Eds.), Proceedings of the Seventh Italian Conference on Computational Linguistics CLiC-it 2020 : Bologna, Appendix Italy, March 1-3, 2021, Collana Dell’Associazione Italiana Di Linguistica Computazionale, Accademia This appendix reports the SPARQL Queries illustrated in University Press, Torino, 2020, pp. 457–463. doi:10. 6.
4000/books.aaccademia.8964. [35] V. Basile, Valeriobasile/sentixR, 2019-2024. Query 1 [36] M. Vassallo, G. Gabrieli, V. Basile, C. Bosco, The tenuousness of lemmatization in lexicon-based sen- Retrieve the distribution of emotions: timent analysis, in: Proceedings of the Sixth Italian Conference on Computational Linguistics, volume 2481, Ceur, 2019, pp. 1–6. PREFIX l i l a : < h t t p : / / l i l a − e r c . eu / [37] E. Zanchetta, M. Baroni, Morph-it! A free corpus- o n t o l o g i e s / l i l a / > based morphological resource for the Italian lan- PREFIX e l i t a : < h t t p : / / w3id . o r g / e l i t a guage, in: Proceedings of Corpus Linguistics Con- / > ference Series 2005 (ISSN 1747-9398), volume 1, Uni- PREFIX o n t o l e x : < h t t p : / / www. w3 . o r g / ns versity of Birmingham, 2005, pp. 1–12. / lemon / o n t o l e x #> [38] V. Basile, M. Lai, M. Sanguinetti, Long-term social PREFIX r d f s : < h t t p : / / www. w3 . o r g media data collection at the university of turin, in: / 2 0 0 0 / 0 1 / r d f − schema #> Proceedings of the Fifth Italian Conference on Computational Linguistics (CLiC-it 2018), Torino, Italy, SELECT ? e m o t i o n L a b e l (COUNT ( ∗ ) as ? December 10-12, 2018., 2018. URL: http://ceur-ws. count ) org/Vol-2253/paper48.pdf. WHERE { [39] W. J. B. van Heuven, P. Mandera, E. Keuleers, ? lemma a l i l a : Lemma .
M. Brysbaert, Subtlex-uk: A new and im- ? e l i t a L e m m a o n t o l e x : c a n o n i c a l F o r m proved word frequency database for british en- ? lemma . glish, Quarterly Journal of Experimental Psy- ? e l i t a L e m m a e l i t a : HasEmotion ? chology 67 (2014) 1176–1190. URL: http://dx.doi. e m o t i o n . org/10.1080/17470218.2013.850521. doi:10.1080/ ? e m o t i o n r d f s : l a b e l ? e m o t i o n L a b e l 17470218.2013.850521. } [40] G. K. Zipf, Human Behaviour and the Principle of GROUP BY ? e m o t i o n L a b e l
Least Efort: an Introduction to Human Ecology, ORDER BY DESC ( ? count )
Addison-Wesley, 1949.
Query 2
PREFIX l i l a : < h t t p : / / l i l a − e r c . eu / o n t o l o g i e s / l i l a / > PREFIX marl : < h t t p : / / www. g s i . upm . e s / o n t o l o g i e s / marl / ns #> PREFIX o n t o l e x : < h t t p : / / www. w3 . org / ns / lemon / o n t o l e x #> PREFIX r d f s : < h t t p : / / www. w3 . org / 2 0 0 0 / 0 1 / r d f −schema #> SELECT ? p o l a r i t y L a b e l (COUNT ( ∗ ) as ? count ) WHERE { ? lemma a l i l a : Lemma . ? sentixLemma o n t o l e x :
c a n o n i c a l F o r m ? lemma . ? sentixLemma marl : h a s P o l a r i t y ? p o l a r i t y . ? p o l a r i t y r d f s : l a b e l ?
p o l a r i t y L a b e l } GROUP BY ? p o l a r i t y L a b e l ORDER BY DESC ( ? count ) Query 3
PREFIX e l i t a : < h t t p : / / w3id . org / e l i t a / > PREFIX o n t o l e x : < h t t p : / / www. w3 . org / ns / lemon / o n t o l e x #> PREFIX r d f s : < h t t p : / / www. w3 . org / 2 0 0 0 / 0 1 / r d f −schema #> PREFIX marl : < h t t p : / / www. g s i . upm . e s / o n t o l o g i e s / marl / ns #> PREFIX xsd : < h t t p : / / www. w3 . org / 2 0 0 1 /
XMLSchema#>
WHERE { ? lemma a l i l a : Lemma . ? elitaLemma o n t o l e x : c a n o n i c a l F o r m ? lemma . ? elitaLemma e l i t a : HasEmotion ? emotion . ? sentixLemma o n t o l e x : c a n o n i c a l F o r m ? lemma . ? sentixLemma marl : h a s P o l a r i t y V a l u e ? p o l a r i t y V a l u e .
? emotion r d f s : l a b e l ? e m o t i o n L a b e l . } GROUP BY ? e m o t i o n L a b e l ORDER BY ASC ( ? a v g P o l a r i t y ) Query 4
SELECT ? l a b e l ? value WHERE { ? lemma a l i l a : Lemma . ? elitaLemma o n t o l e x : c a n o n i c a l F o r m
? lemma . ? elitaLemma e l i t a : HasEmotion
e l i t a : G i o i a . ? elitaLemma e l i t a : HasEmotion
e l i t a : T r i s t e z z a . ? sentixLemma o n t o l e x :
c a n o n i c a l F o r m ? lemma . ? sentixLemma marl : h a s P o l a r i t y ?
p o l a r i t y . ? sentixLemma marl :
h a s P o l a r i t y V a l u e ? value . ? elitaLemma r d f s : l a b e l ? l a b e l . ? p o l a r i t y r d f s : l a b e l " N e g a t i v e "
@en . Declaration on Generative AI During the preparation of this work, the author(s) used ChatGPT (OpenAI), Grammarly, and DeepL Write / DeepL Translate in order to: Text translation, Paraphrase and reword, and Grammar and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.