Irony and Sarcasm Detection in Twitter:The Role of Affective Content

Deteccion de Ironıa y Sarcasmo en Twitter:La Funcion del Contenido Affectivo

Delia Irazu Hernandez FarıasPRHLT Research Center, Universitat Politecnica de Valencia, Spain

Dipartimento di Informatica, University of Turin, Italy

Department of Computational SciencesNational Institute of Astrophysics, Optics and Electronics

Luis Enrique Erro 1, 72840, Santa Marıa Tonantzintla, Puebla, Mexicodirazuherfa@inaoep.mx

Resumen: Tesis doctoral en Informatica realizada por Delia Irazu Hernandez Farıasy dirigida por el Dr. Paolo Rosso (Universitat Politecnica de Valencia) y la Dra. Vi-viana Patti (University of Turin) en el marco de un convenio de cotutela entre laUniversitat Politecnica de Valencia, Espana y la Universidad de Turin, Italia. Ladefensa de la tesis fue en Valencia, Espana el 25 de septiembre de 2017 ante untribunal compuesto por: Horacio Saggion (Universitat Pompeu Fabra), ElisabettaFersini (Universita degli Studi di Milano-Bicocca) y Roberto Basili (Univerita diRoma Tor Vergata). Se obtuvo la mencion internacional tras una estancia de 12meses en la Universidad de Turin.Palabras clave: Deteccion de ironıa, deteccion de sarcasmo, procesamiento de len-guaje figurado, analisis de sentimientos

Abstract: PhD thesis in Computer Science written by Delia Irazu Hernandez Farıasunder the supervision of Dr. Paolo Rosso (Universitat Politecnica de Valencia) andDra. Viviana Patti (University of Turin). This thesis was developed under a cotu-telle between the Universitat Politecnica de Valencia, Spain and the University ofTurin, Italy. The thesis defense was done in Valencia, Spain on September 25, 2017.The doctoral committee was integrated by: Horacio Saggion (Universitat PompeuFabra), Elisabetta Fersini (Universita degli Studi di Milano-Bicocca), and RobertoBasili (Univerita di Roma Tor Vergata). The International mention was achievedafter a 12 months internship at University of Turin.Keywords: Irony detection, sarcasm detection, figurative language processing, sen-timent analysis

1 Introduction

People tend to use irony and sarcasm in so-cial media to achieve different communica-tion purposes. Dealing with such figurativelanguage devices represents a big challengefor computational linguistics. The fuzzy sepa-ration between irony and sarcasm could leadto confusion. Commonly, the term “irony” isused as an umbrella term also covering sar-casm1.

Irony is closely associated with the expres-sion of feelings, emotions, attitudes, and eva-

1In what follows we shall use “irony” in the sameperspective.

luations toward a particular target. It allowsus to convey very subjective ideas and opi-nions in an indirect way, going beyond theliteral meaning of the words. Therefore, sin-ce irony is considered as an affective mannerof communication, taking advantage of affect-related information may help to identify iro-nic content in social media.

This thesis aimed at investigating the ro-le of affect-related information in irony de-tection. We proposed to exploit affect-basedinformation to characterize ironic content inTwitter. A complex and multifaceted pheno-menon such as irony merits to be addressed

Procesamiento del Lenguaje Natural, Revista nº 62, marzo de 2019, pp. 107-110 recibido 15-11-2018 revisado 07-01-2019 aceptado 01-02-2019

ISSN 1135-5948. DOI 10.26342/2019-62-14 © 2019 Sociedad Española para el Procesamiento del Lenguaje Natural

not only considering broad aspects of affect(such as positive and negative sentiment) butalso affective information in a finer-grainedperspective (taking into account different mo-dels of emotions). This research was conduc-ted paying special attention on three mainaspects:

I. We analyzed the presence of differentaspects of affect in ironic utterances in orderto identify potential features to characterizesuch phenomenon in social media. We propo-sed a model, called emotIDM, which exploitsan extensive set of resources covering diffe-rent facets of affect ranging from sentimentto finer-grained emotions for characterizingironic utterances. To evaluate our model, wecollected a set of Twitter corpora used byscholars in previous research, to be used asbenchmarks with a two-fold purpose: to com-pare the performance of our model againstother approaches in the state of the art, andto evaluate its robustness across different as-pects related to the characteristics of the cor-pora. Results show that emotIDM has a com-petitive performance across the experimentscarried out, validating its effectiveness.

II. Aiming to investigate the differencesbetween tweets labeled with #irony, #sar-casm, and #not, we analyzed different facetsof affective information in instances contai-ning such labels. We find data-driven argu-ments suggesting that the above mentionedhashtags are used to refer different figurativelanguage devices. The results of our analysisallow us to probe distinctions and similaritiesbetween tweets labeled as #irony and #sar-casm. Our results allow us to contribute tothe assumption that there is a separation bet-ween these figurative language devices. Furt-hermore, bearing in mind the importance ofirony detection for sentiment analysis tasks,we investigated also the differences in pola-rity reversal terms of such tweets.

III. Detecting irony in user-generatedcontent could have a broad range of appli-cations. One of them is on sentiment analy-sis (SA). We analyzed the impact of ironyin the performance of systems dedicated toSA, observing a drop when they deal withsuch figurative language devices. We propo-sed an irony-aware sentiment analysis systemthat incorporates emotIDM in a first stagebefore determining the overall polarity of agiven Twitter message. The obtained resultsare competitive with the state of the art.

2 Thesis Overview

This thesis is comprised of a compendium ofresearch articles already published (two in-ternational journal papers, an internationalconference paper, and a chapter in a book)as well as unpublished content. It consist of8 chapters that are briefly introduced below.

Chapter 2 (Hernandez Farıas and Rosso,2016) presents an overview of some state-of-the-art approaches to deal with irony andsarcasm detection in social media. Further-more, an analysis of the performance of sen-timent analysis systems on the presence ofironic content is presented. In Chapter 3(Hernandez Farıas, Benedı, and Rosso, 2015),we described an irony detection model exploi-ting surface patterns as well as some featurescoming from sentiment analysis. The latterbeing the most relevant ones, thereby givinginsights into the importance of such informa-tion for detecting irony in Twitter.

Chapter 4 (Hernandez Farıas, Patti, andRosso, 2016) presents emotIDM which con-siders several aspects of affect ranging fromsentiment to fine-grained models of emotions.The obtained results outperform the perfor-mance of the state-of-the-art approaches va-lidating the importance of affect-related in-formation for irony detection in Twitter. InChapter 5 (Sulis et al., 2016), we analy-zed different facets of affective informationin tweets labeled with #irony, #sarcasm and#not with the aim of distinguishing betweenthem. We also considered the role in termsof polarity reversal of tweets containing suchhashtags.

Chapter 6 (Hernandez Farıas et al., 2017)describes an approach for performing senti-ment analysis in tweets with figurative lan-guage. We proposed a pipeline that comprisestwo phases: first, we exploited emotIDM foridentifying irony; then, by taking advantageof several affective resources we determine apolarity degree also considering the presenceof ironic content. In Chapter 7, we presen-ted the obtained results of some further ex-periments and analysis carried out with theaim to enhance the research work. Finally, inChapter 8 we draw the main conclusions ofthe thesis, as well as the contributions andfuture work.

3 Contributions

Irony is a complex mode of communicationclosely related to the expression of feelings.

Delia Irazú Hernández Farías

108

In this thesis, we introduced the problem ofdetecting irony and sarcasm in social mediaby considering mainly the subjective intrin-sic value enclosed in ironic expressions. Inthis line, the following contributions were ma-de within the development of the present re-search:

Overview of irony detection and its impacton SA. We presented a brief description ofthe proposed approaches in the literaturetogether with an analysis of shared tasksregarding SA where the participating sys-tems were evaluated on the presence of fi-gurative language devices. Overall, thereis a drop in performance of the systems. Itallows validating the assumption concer-ning to the importance of irony detectionfor determining the sentiment in a text.

emoIDM: an irony detection model. Weproposed a novel model for identifyingirony in social media by taking advantageof different facets of affective informationto capture ironic intention in Twitter. Un-like other research works, we did not buildour own dataset, instead we collected aset of Twitter corpora previously used inthe literature. The obtained results overalloutperform those from the related workvalidating the importance of such kind ofcontent for irony detection.

Analysis on the use of different hashtagsrelated to ironic phenomena. We investiga-ted the use of different hashtags (#irony,#sarcasm, and #not), concluding thatthese labels are indeed used to tag dif-ferent phenomena. Moreover, we exploredthe controversial subject to separate ironyfrom sarcasm outperforming the state ofthe art. With respect to #not, it seemsthat it is used to represent a different figu-rative language device, although closer to#sarcasm than #irony. Furthermore, weinvestigated the behaviour of tweets labe-led with such hashtags in terms of polarityreversal. It seems that in tweets labeledwith #sarcasm often there is a full rever-sal (varying from a polarity to its opposi-te, almost always from positive to negati-ve polarity), whereas in the case of thosetagged with #irony there is an attenua-tion of the polarity (mostly from negativeto neutral).

Development of an irony-aware sentimentanalysis system. Irony is a phenomenon

having an impact on the performance ofsystems dedicated to calculate the overallsentiment expressed in a given piece oftext. We incorporated our irony detectionmodel in a pipeline of sentiment analy-sis that relies mainly on sentiment andemotional-related resources. The obtainedresults were competitive and serve to va-lidate the relevance of exploiting affectiverelated features as well as the presence ofirony for determining the sentiment in agiven tweet.

Acknowledgments

This work has been funded by the Na-tional Council for Science and Technology(CONACyT - Mexico) with the Grant No.218109/313683. Part of the research was ca-rried out in the framework of the SomEM-BED TIN2015-71147-C2-1-P MINECO pro-ject.

References

Hernandez Farıas, D. I., C. Bosco, V. Patti,and P. Rosso. 2017. Sentiment polarityclassification of figurative language: Ex-ploring the role of irony-aware and multi-faceted affect features. In ComputationalLinguistics and Intelligent Text Processing- 18th International Conference, CICLing2017, Revised Selected Papers, Part II, pa-ges 46–57.

Hernandez Farıas, D. I., V. Patti, and P. Ros-so. 2016. Irony Detection in Twitter: TheRole of Affective Content. ACM Trans.Internet Technol., 16(3):19:1–19:24.

Hernandez Farıas, D. I. and P. Rosso. 2016.Irony, Sarcasm, and Sentiment Analysis.Chapter 7. In F. A. Pozzi, E. Fersini,E. Messina, and B. Liu, editors, Senti-ment Analysis in Social Networks. MorganKaufmann, pages 113–127.

Hernandez Farıas, I., J.-M. Benedı, andP. Rosso. 2015. Applying Basic Featu-res from Sentiment Analysis for Automa-tic Irony Detection. In Pattern Recogni-tion and Image Analysis, volume 9117 ofLNCS, pages 337–344. Springer Interna-tional Publishing.

Sulis, E., D. I. Hernandez Farıas, P. Rosso,V. Patti, and G. Ruffo. 2016. Figurativemessages and affect in Twitter: Differen-ces between #irony, #sarcasm and #not.Knowledge-Based Systems, 108:132–143.

Irony and sarcasm detection in twitter: The role of affective content

109