#andràtuttobene: Images, Texts, Emojis and Geodata in a Sentiment Analysis Pipeline Pierluigi Vitale, Serena Pelosi, Mariacristina Falco Department of Political and Communication Sciences University of Salerno [pvitale,spelosi,mfalco]@unisa.it Abstract The methodology proposed is based on a fully automatic natural language processing pipeline, This research investigates Instagram us- including the images’ analysis phase. Its output is ers’ sentiment narrated during the lock- an interactive dashboard (Figure 1) that is able to down period in Italy, caused by the explore the sentiment analysis values about every COVID-19 pandemic. The study is based single kind of text, and the synthesis of all of on the analysis of all the posts published them. Thanks to a system of interactions and fil- on Instagram under the hashtag #an- ters, the observation is leaded by the images’ fea- dràtuttobene on May 4, May 18 and June tures, such as different kind of spaces (indoor or 3, 2020. Our research carried out a view outdoor) and different kind of the photos’ subject on a national, regional and provincial (human or not human). scale. We analyzed all the different lan- The collected geographical data enabled the guages and forms (i.e. captions, hashtags, analysis of several dimensions, with an overview emojis and images) that constitute the observation based on the regional scale. Hence, it posts. The aim of this research is to pro- gave us an opportunity to focus on the deeper vide a set of procedures revealing the dif- level of the Italian provinces. This choice is moti- ferent polarity trends for each kind of ex- vated by the Italian DPCM (Decreto Presidenza pression and to propose a single compre- del Consiglio dei Ministri) published on 24 March hensive measure. 20202, in which it is stated the partial autonomy of the regions. Introduction 1. State of the Art This paper investigates the case of the Italian most used hashtag about the lockdown period for In Natural Language Processing (NLP) studies, the COVID-19 pandemic on Instagram: #an- the automatic treatment of opinionated expres- dràtuttobene1. sions and documents is known as Sentiment Anal- The research team collected 7,482 posts, the ysis. entire amount published in three specific dates: Lexical resources for sentiment analysis cre- May 4, May 18 and June 3, corresponding with ated for the Italian Language are Sentix (Basile, three different steps of the reopening phase of the Nissim 2013); SentIta (Pelosi 2015a); the lexicon country, led by the government. of the FICLIT+CS@UniBO System (Di Gennaro Instagram posts are composed by several kind 2014); the CELI Sentiment Lexicon (Bolioli of languages: captions (texts), hashtags, emojis 2013); the Distributional Polarity Lexicons (Cas- and images. The aim of this work is to design a set tellucci 2016). of procedures revealing the different polarity For the Italian language, significant contribu- trends for each one and to propose a unique meas- tions on sentiment analysis of social media come ure. This measure can show the sentiment ex- from Bosco et al. (2013, 2014), Castellucci (2014, pressed by the texts, in their semiotic broad mean- 2016) and Stranisci (2016), among others. ing. Copyright ©️2020 for this paper by its authors. Use per- billion monthly active users, according to a study by mitted under Creative Commons License Attribution Hootsuite and WeAreSocial. 2 4.0 International (CC BY 4.0). https://www.gazzettauff- ciale.it/eli/id/2020/06/17/20G00071/sg 1 The choice of Instagram is due to its success. It is in fact one of the most popular social networks, with 1 1 Hahstag processing in Sentiment Analysis is par- 2017, Tian 2017), but, it is possible to improve the ticularly challenging in terms of word segmenta- score of sentiment analysis tools by knowing the tion. Obviously, the absence of white spaces be- meaning of emojis (LeCompte 2017, Felbo 2017, tween words poses several problems that concerns Guibon 2016, Novak 2015). ambiguity. Among the most relevant contribution The content analysis of the images has been ad- in this area, we cite Zangerle (2018), Reuter dressed from several perspectives and techniques. (2016), Simeon (2016); Bansal 2015, Srinivasan Several studies are moving from a fully quali- (2012) and Celebi (2018). The solution proposed tative and manual approach (Tifentale & Ma- in literature concerns mostly the use of n-grams, novich, 2015, Vitale et al., 2019, Palazzo et al., syntactic complexity, pattern length, or pos-tag- 2020, Esposito et al., 2020) to mixed methods in- ging. volving algorithms and computer vision tech- In the last years, the way to communicate niques combined with qualitative observations online involves many kinds of languages, con- (Hochman 2015, Indaco & Manovich, 2016). nected to verbal and non-verbal features. This This work, starting from a well experimented complexity makes classical textual analysis less innovative approach on previous studies (Vitale et adequate to have a real and representative per- al., 2020, Giordano et al, 2020), makes the choice spective on people’s interests and opinions. to analyze the images in their textual translation, In particular way, the conventional approach with a fully automatic analytical pipeline, de- seems to be not suitable for visual social media, signed in a semiotic point of view. Besides the se- such as Instagram, where all the languages are in- miotic interest to digital media date back to the volved and the images seem to be dominant. early 2000s and continues to the present days, The analysis of these social media tends to un- considering digital media a specific semiotic field derline the issues of textocentricity (Singhal & (Cosenza 2014, Bianchi e Cosenza 2020). Rattine-Flaherty, 2006) and textocentrism (Balo- Lastly, considering design, the visual represen- menu & Garrod, 2019), making necessary a dif- tation of the social media data is increasing wide- ferent way to approach the participant generated spread as vehicle for knowledge of several fields images (PGI) or user generated contents (UGC) in (Ciuccarelli et al., 2014). general. The research team doesn’t provide an algorithm Opinions, emotions, and contents are expressed to analyze the images but adopt the automatic in a mixed way, that is the combination of several translation from the social media algorithm, de- languages, visual and textual, and the related signed to the visual impaired users by parsing the metadata, such as: geographical position and html code of the Instagram web interface. The hashtag which they are labeled with. metadata involved is the “accessibility_caption”. The automatic treatment of emojis is faced These are lists of words, hierarchically distrib- through two main approaches: the processing of uted, that let us to define and observe subject and the textual descriptions of emojis3 (Fernández- attributes of the images, in addition to allowing Gavilanes 2018, Singh 2019) and the analysis of the analysis of the entities. the emojis (co-)occurrences (Guibon 2018, Rakhmetullina 2018, Barbieri 2016, Novak et al., 2015)4. The function of emojis is not limited to a pre- dictable labeling of the emotional content (Felbo 3 4 Among the emoji resources, we mention; Emojipedia Actually, the original meaning of emojis, specified in (emojipedia.org), iEmoji (www.iemoji.com); the anno- their descriptions, could be very different with the ones tated resources, such as The Emoji Dictionary (emo- attributed by people into specific text occurrences (Fer- jidictionary.emojifoundation.com); EmojiNet (emo- nández-Gavilanes 2018, Wood & Ruder 2016). There- jinet.knoesis.org); the ones which are specific for Sen- fore, the manual annotation of emoji dictionaries could timent Analysis and Emotion Detection purposes, such ignore important details that concerns usage dynamics as Emoji Sentiment Ranking (kt.ijs.si/data/Emoji_sen- over time (Ahanin 2020, Felbo 2017). timent_ranking), EmoTag (abushoeb.github.io/emo- Nevertheless, the representation of an emoji can vary tag) and The SentiStrength emoticon sentiment lexicon widely across different communication platforms (sentistrength.wlv.ac.uk); and the corpora, such as (Wagner, 2020) and their semantics can present culture ITAMoji (Ronazano 2018) and the Emojiworldbot cor- or language specific usage patterns (Barbieri 2016). pus (Monti 2016). Thus, the results produced by the analyses of the emoji in large corpora could present some drawbacks as well. 2 2. Methodology written Italian have been detected automatically by adopting the google translate API (Application In this work, we propose the automatic treat- Programming Interface) and removed. Moreover, ment of the sentiment expressed into 7,482 Insta- from this field all the hashtags have been ex- gram posts. tracted, to allow their standalone analysis. All the information composing the dataset (i.e. Accessibility captions have been clustered on captions, hashtags, emojis and images) are auto- two dimensions: “human or not human” and “in- matically put into relation with one another and door or outdoor”, previously defined thanks to a visualized into an interactive dashboard. The phe- list of coherent words, subsequently matched by a nomenon, can be observed through a system of fil- pattern matching phase6. Geographical coordi- ters, zooms and interdependent interactions. The nates set the images on a specific point on the result captures the topography of feelings, moods map, so it has been necessary to make a reverse and needs expressed on the Instagram platform geocoding procedure to find out region and prov- during the lockdown. ince levels.7 The NLP activities are performed in this re- Furthermore, Timestamp have been converted search through the software NooJ5, which allows in a conventional date and time format. both the formalization of linguistic resources and After these steps, images and texts became the parsing of corpora. The dictionaries and gram- ready to be analyzed through NLP procedures and mars, which have been built ad hoc for this work, mapped with geographical visualization tech- complement the open-source resources of the niques, observing them on the desired timeframe. basic Italian and English modules of NooJ (Vietri For the analysis of verbal features, we used 2014). SentIta, a semi-automatically built lexicon task All the pictures published on May 4, May 18 (Pelosi 2015a), containing more than 15,000 lem- and June 3, 2020 with the hashtag #andratut- mas, simple words and multiword units. Each en- tobene have been collected with a custom python try is annotated with polarity and intensity scores, script that simulate the human navigation. For into a scale that ranges from -3 to +3. It must be each picture, we collected the entire source code applied to texts in conjunction with a network of of the web page in a JSON (JavaScript Object No- almost 130 embedded local grammars, formalized tation) format. in the shape of Finite State Automata (Pelosi This one has been parsed to a tabular one, in 2015b), which systematically modulate the prior order to plan a format suitable for the adopted polarity of words according to their syntactic local tools. The files have been refined selecting the context8. These resources can be directly applied endpoint useful for the analysis: captions (includ- to the Instagram captions, while hashtags need to ing hashtags); images hyperlink; accessibility be initially segmented. In this phase, they are an- captions; geographical coordinates and alytically decomposed into their constituents timestamp. through 10 morpho-syntactic grammars applied Some data required a data refinement phase. simultaneously, but with different priorities. In For the captions, it has been necessary to do a this way, the selection of the most probable se- cleaning phase in which all the texts that were not quences is decided for the upstream9. 5 9 http://www.nooj-association.org/ Basically, if the system produces more than one inter- 6 For instance, in the “human” cluster we have grouped pretation, the preferred one is the one in which the con- all the accessibility caption containing words such as stituents have a longer length and the smallest number “people, man, woman, person” etc. of constituents. In other words, the system firstly com- At the same time, in the “outdoor” cluster we have pares the whole normalized string with the word forms grouped all the pictures with words such as “sea, sky- from SentIta, then continues the comparison with Eng- line, lawn, beach” and so on. lish and Italian word forms from the basic module. 7 This phase has been possible in an automatic way Hence, the dictionaries receive the higher priority and adopting the python library reverse-geocoder are applied before morphological grammars. (https://github.com/thampiman/reverse-geocoder) If the system does not match any word in the lexicon, 8 The performances of our method produced satisfac- it starts the structural analysis of the string, which con- tory results in the sentence-level analysis of the textual sist of a systematic comparison of substrings with the part of the corpus: 0,85 Recall; 0,96 Precision and 0,9 all the words contained in the dictionaries, according F-score. to part of speech specific syntactic structures. Such structure, ordered here by priority assignments, can be 3 For the analysis of the non-verbal features, emojis are treated by using an electronic diction- ary, which has been semi-automatically annotated with the same information used to analyze verbal features. We created this database with recogniza- ble decimal codes in UTF-8 encoding from Emo- jipedia, then we carried out the automatic analysis of the textual descriptions of each emoji. This dictionary has been used to locate and in- terpret the emojis occurring in the posts10. After the clustering phase (human and not hu- man; indoor and outdoor), all the findings of the sentiment on all the languages can be associated to the pictures’ features, combined or not.11 3. Visualization and Results For a complete observation of the analysis’ pro- cess and of its results, we developed a data visu- alization dashboard. In the following dashboard it is possible to observe the sentiment analysis on each language processed, with the chance of in- vestigating the different trends during the days and the single hours day by day. Adopting the clusters detected in the images, a system of filters let to focus the results basing on the subjects depicted. On the left side of the dashboard, a map shows the geographical situation, merging the 4 senti- ment values in a single one (weighted average) Figure 1 The sentiment analysis values and coloring the regional shape on chromatic As a matter of fact, Novak (2015) underlined scale from the minimum value (-3) in orange, to that it is more common the use of positive emojis the maximum value (+3) in blue. The same scale with respect to the negative ones. Moreover, Boia is applied to the line chart on the right, in which et al. (2013) observed a poor correlation between each line is related to the vertical axes and colored the perceived emotional polarity of emojis and the as described before. accompanying linguistic text alone. Although it is (𝑆𝑒𝑛𝑡 𝐸𝑚𝑜𝑗𝑖𝑠 ∗ 𝑃 𝐸𝑚𝑜𝑗𝑖𝑠 + 𝑆𝑒𝑛𝑡 𝐻𝑎𝑠ℎ𝑡𝑎𝑔𝑠 ∗ 𝑃 𝐻𝑎𝑠ℎ𝑡𝑎𝑔 + 𝑆𝑒𝑛𝑡 𝑇𝑒𝑥𝑡𝑠 ∗𝑃 𝑇𝑒𝑥𝑡𝑠 ) actually challenging to predict the interaction be- 𝑃 𝐸𝑚𝑜𝑗𝑖𝑠 + 𝑃 𝐻𝑎𝑠ℎ𝑡𝑎𝑔 + 𝑃 𝑇𝑒𝑥𝑡𝑠 tween emoji and texts, there are cases in which the emojis express or reinforce the sentiment of the Each score reached by the three languages are text with which they occur and cases in which taken into account, namely texts, hashtags and they modify it or even express an opposite emo- emojis, are weighted according to the assumption tional state (Guibon 2018, Shoeb 2019). that the euphoric level of emojis’ sentiment is Hashtags are conventionally used in two ways: higher than hashtags’ one, and both are higher on one hand, to describe the contents in a list of than written texts’ one in general. According to words, and on the other hand for strategic pur- these results, we propose this weighted average poses, in order to place the images in useful the- formula, in which emojis, hashtags and texts have matic spaces. This is also the reason why we have different weights (P), respectively 33, 50, and removed from the analysis of all the Instagram- 100. multiword expressions; free nominal, prepositional, full words contained in the posts, the sentiment labelled adjectival and adverbial phrases; elementary sen- emojis cover the 19% of the total number of emojis in tences; and verbless sentences. the corpus. 10 While the oriented words located into captions and hashtags respectively cover the 6% and the 9% of the 4 specific hashtags, such as: #likeforlike, #fol- lowoforfollow etc., that are not suitable or even could be misleading or biased for our investiga- tion. At the same time, the hashtags are also used as part of the messages, in substitution of words, so they deserve to be included in the final meas- ure, but not with the same relevance of the cap- tions. The performances of emoji, hashtag and texts as indicators for sentiment analysis purposes, alone and combined with one another, have been tested on our corpus. We verified a significant im- provement in terms of document-level precision when the indicators are considered together (0.98), if compared with the precision of texts (0.91), hashtag (0.81) and emoji (0.65) considered alone. The different precisions reached by the three languages considered alone empirically con- firm the diversification of weights we proposed in our formula. This weighted measure has, then, been compared by three different judges12 with the arithmetic mean on a sample of 100 Instagram posts from our corpus and performed better in the 92% of the cases. Nevertheless, the geographical dimension is very important to observe the different kind of languages in the online community (Arnaboldi et al., 2017). Through an overlay function, moving Figure 2 Overlay function: provincial scale the cursor on the map (figure 2), we show the ge- The research brought together linguistic analysis ographical data in the deeper level of the single and design into a more general semiotic frame- province, focusing on each region. The result rep- work. The aim was, in fact, to put in shape the resents the possible different polarity value be- pandemic phenomenon through a selection of lin- tween different provinces. For instance, on May 4 guistic relevance. in the provinces of Oristano (Sardegna), Genova The virus caused a series of unpredictable changes (Liguria) and Viterbo (Lazio), the sentiment value narrated on Instagram through the hashtag #an- is negative, despite the positive average value of dràtuttobene. A mantra for the Igers and an isot- the region. However, the average sentiment value opy for the analysts (Greimas & Courtès 1979). over the three days analyzed is found always pos- Working on multiple levels, the research has of- itive, with different evidences on regional and fered a general and a local view of the emotions provincial scale. Lastly, users can explore the re- told during the lockdown period. Starting from a sults focusing on one or more region though a fil- lexical base, made up by a list of words, and using ter function (by clicking or selecting). All the fil- electronic dictionaries also for the images, the ters are interdependent, so it is possible to select analysis organized a large amount of data, devel- all the functions available investigating the phe- oping a real map of emotions and needs expressed nomenon from all the possible perspectives. during the first wave of pandemic. The map can be visualized trough a dashboard letting users ob- Conclusion serve general and local reactions, down to the sin- gle province. The emotional effects of sense have Throughout the quantitative and qualitative anal- been evaluated thanks to a polar and unique meas- ysis of the different expressive forms used on In- ure. stagram, this work proposes a general view of COVID-19 in Italy. 12 For the evaluation of the three judges, we have cal- dorff’s Alpha formula (kalpha). The three coders se- culated the intercoder reliability adopting the Krippen- lected have a kalpha of 0.9. 5 In the end, did everything really go well for In- Bosco, C., Patti, V., & Bolioli, A. (2013). Developing stagram's Italy? In general, it seems so. The aver- corpora for sentiment analysis: The case of irony age sentiment value over the three days analyzed and senti-tut. IEEE intelligent systems, 28(2), 55- is always positive, with variations on regional and 63. provincial scale. Going down the single province, Boia M., Faltings B., Musat C. C., & Pu P. (2013). A:) we can find differences, as the Sardinia, Lazio and is worth a thousand words: How people attach sen- Liguria cases. timent to emoticons and words in tweets. In “2013 International Conference on Social Computing”, pp. 345-350. IEEE. References Castellucci, G., Croce, D., & Basili, R. (2016). A lan- guage independent method for generating large Agerri, R. and García-Serrano, A. (2010). Q-WordNet: scale polarity lexicons. In Proceedings of the Tenth Extracting polarity fromWordNet senses. In Pro- International Conference on Language Resources ceedings of the International Conference on Lan- and Evaluation (LREC'16) (pp. 38-45). guage Resources and Evaluation, pages 2300–2305. Celebi, A., & Özgür, A. (2016). Segmenting hashtags Ahanin, Z., & Ismail, M. A. (2020). Feature extraction using automatically created training data. In Pro- based on fuzzy clustering and emoji embeddings for ceedings of the Tenth International Conference on emotion classification. International Journal of Language Resources and Evaluation (LREC'16) Technology Management and Information System, (pp. 2981-2985). 2(1), 102-112. Celebi, A., & Özgür, A. (2018). Segmenting hashtags Arnaboldi, M., Brambilla, M., Cassottana, B., Ciuc- and analyzing their grammatical structure. Journal carelli, P., & Vantini, S. (2017). Urbanscope: A lens of the Association for Information Science and to observe language mix in cities. American Behav- Technology, 69(5), 675-686. ioral Scientist, 61(7), 774-793. Cosenza, G., (2014). Introduzione alla semiotica dei Balomenou, N., & Garrod, B. (2019). Photographs in nuovi media, Laterza, Milano. tourism research: Prejudice, power, performance and participant-generated images. Tourism Man- Di Gennaro, P., Rossi, A., & Tamburini, F. (2014). The agement, 70, 201-217. FICLIT+ CS@ UniBO System at the EVALITA 2014 Sentiment Polarity Classification Task. In Pro- Bansal, P., Bansal, R., & Varma, V. (2015). Towards ceedings of the Fourth International Workshop deep semantic analysis of hashtags. In European EVALITA 2014. conference on information retrieval (pp. 453-464). Springer, Cham. Eisner, B., Rocktäschel, T., Augenstein, I., Bošnjak, M., & Riedel, S. (2016). emoji2vec: Learning emoji Barbieri, F., Ronzano, F., & Saggion, H. (2016). What representations from their description. arXiv pre- does this emoji mean? a vector space skip-gram print arXiv:1609.08359. model for twitter emojis. In Proceedings of the Tenth International Conference on Language Re- Esposito, F., Falco, M., & Vitale, P. (2020). Experienc- sources and Evaluation (LREC 2016); 2016 May ing Museums A Qualitative and Quantitative De- 23-28; Portorož, Slovenia. scription About Igers’ Narration of an Exhibit Space. In Workshops of the International Confer- Basile, V., & Nissim, M. (2013). Sentiment analysis on ence on Advanced Information Networking and Ap- Italian tweets. In Proceedings of the 4th Workshop plications (pp. 1011-1018). Springer, Cham. on Computational Approaches to Subjectivity, Sen- timent and Social Media Analysis (pp. 100-107). Felbo, B., Mislove, A., Søgaard, A., Rahwan, I., & Lehmann, S. (2017). Using millions of emoji occur- Bianchi, C., & Cosenza, G. (2020). LexiaRivista di rences to learn any-domain representations for de- semiotica, 33-34. Semiotica e Digital marketing, tecting sentiment, emotion and sarcasm. arXiv pre- Roma, Aracne. print arXiv:1708.00524. Bolioli, A., Salamino, F., & Porzionato, V. (2013). So- Fernández-Gavilanes, M., Juncal-Martínez, J., García- cial Media Monitoring in Real Life with Blogmeter Méndez, S., Costa-Montenegro, E., & González- Platform. ESSEM@ AI* IA, 1096, 156-163. Castaño, F. J. (2018). Creating emoji lexica from Bosco, C., Allisio, L., Mussa, V., Patti, V., Ruffo, G. unsupervised sentiment analysis of their descrip- F., Sanguinetti, M., & Sulis, E. (2014). Detecting tions. Expert Systems with Applications, 103, 74- happiness in Italian tweets: Towards an evaluation 91. dataset for sentiment analysis in Felicitta. In 5th In- Giordano, G., Primerano, I., & Vitale, P. (2020). A ternational Workshop on Emotion, Social Signals, Network-Based Indicator of Travelers Performa- Sentiment & Linked Open Data, Es³Lod 2014 (pp. tivity on Instagram. Social Indicators Research, 1- 56-63). European Language Resources Association. 19. 6 Greimas, A. J., Courtés, J. (1979). Sémiotique: diction- cessing and Speech Tools for Italian. Final Work- naire raisonné de la théorie du langage, Paris, shop, EVALITA 2018 (Vol. 2263, pp. 1-9). CEUR- Hachette. WS. Guibon, G., Ochs, M., & Bellot, P. (2018). From emoji Simeon, C., Hamilton, H. J., & Hilderman, R. J. usage to categorical emoji prediction. In 19th Inter- (2016). Word segmentation algorithms with lexical national Conference on Computational Linguistics resources for hashtag classification. In 2016 IEEE and Intelligent Text Processing (CICLING 2018). International Conference on Data Science and Ad- Springer Lecture Notes in Computer Science, Swit- vanced Analytics (DSAA) (pp. 743-751). IEEE. zerland. Singh, A., Blanco, E., & Jin, W. (2019). Incorporating Hochman, N. (2015). The social media image: Modes emoji descriptions improves tweet classification. In of visual ordering on social media (Doctoral disser- Proceedings of the 2019 Conference of the North tation, University of Pittsburgh). American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies, Indaco, A., & Manovich, L. (2016). Urban social me- Volume 1 (Long and Short Papers) (pp. 2096-2101). dia inequality: definition, measurements, and appli- cation. arXiv preprint arXiv:1607.01845. Singhal, A., & Rattine-Flaherty, E. (2006). Pencils and photos as tools of communicative research and LeCompte, T., & Chen, J. (2017). Sentiment analysis praxis: Analyzing Minga Perú’s quest for social jus- of tweets including emoji data. In 2017 International tice in the Amazon. International Communication Conference on Computational Science and Compu- Gazette, 68(4), 313-330. tational Intelligence (CSCI) (pp. 793-798). IEEE. Srinivasan, S., Bhattacharya, S., & Chakraborty, R. Monti J., Sangati F., Chiusaroli F., Benjamin M., and (2012). Segmenting web-domains and hashtags us- Mansour S.. (2016). Emojitalianobot and emo- ing length specific models. In Proceedings of the jiworldbot - new online tools and digital environ- 21st ACM international conference on Information ments for translation into emoji. In Proc. CLiC-it and knowledge management (pp. 1113-1122). 2016, volume 1749 of CEUR Workshop Proceed- ings. Stranisci, M., Bosco, C., Patti, V., & HERNANDEZ FARIAS, D. I. (2015). Analyzing and annotating for Novak, P. K., Smailović, J., Sluban, B., & Mozetič, I. sentiment analysis the socio-political debate on# la- (2015). Sentiment of emojis. PloS one, 10(12). buonascuola. In second Italian Conference on Com- Palazzo, M., Vollero, A., Vitale, P., & Siano, A. Urban putational Linguistics (pp. 274-279). aAcademia and rural destinations on Instagram: Exploring the University Press. influencers’ role in# sustainabletourism. Land Use Tian, Y., Galery, T., Dulcinati, G., Molimpakis, E., & Policy, 100, 104915. Sun, C. (2017). Facebook sentiment: Reactions and Pelosi S., (2015a): SentIta and Doxa: Italian Databases emojis. In Proceedings of the Fifth International and Tools for Sentiment Analysis Purposes. The Tifentale, A., & Manovich, L. (2015). Selfiecity: Ex- second Italian Conference on Computational Lin- ploring photography and self-fashioning in social guistics (CLiC-it 2015). Trento, December 3-4 media. In Postdigital aesthetics (pp. 109-122). 2015. Book of Proceedings. Accademia University Palgrave Macmillan, London. Press srl, Torino. Vitale, P., Mancuso, A., & Falco, M. (2019). Muse- Pelosi, S. (2015b). A Lexicon-based Approach to Sen- ums’ tales: visualizing instagram users’ experience. timent Analysis: The Italian Module for Nooj. For- In International Conference on P2P, Parallel, Grid, malising Natural Languages with Nooj 2014, 37. Cloud and Internet Computing (pp. 234-245). Rathnayake, C., & Ntalla, I. (2020). “Visual Afflu- Springer, Cham. ence” in Social Photography: Applicability of Image Vitale, P., Palazzo, M., Vollero, A., Siano, A., & Segmentation as a Visually Oriented Approach to Foroudi, P. (2020). The Role of Igers in the Territo- Study Instagram Hashtags. Social Media+ Society, rial Dynamics of Sustainable Tourism-Oriented 6(2), 2056305120924758. Destinations. In International Symposium: New Reuter, J., Pereira-Martins, J., & Kalita, J. (2016). Seg- Metropolitan Perspectives (pp. 759-767). Springer, menting twitter hashtags. Intl. J. on Natural Lang. Cham. Computing, 5(4). Ronzano, F., Barbieri, F., Wahyu Pamungkas, E., Patti, V., & Chiusaroli, F. (2018). Overview of the evalita 2018 italian emoji prediction (itamoji) task. In 6th Evaluation Campaign of Natural Language Pro- 7