In a recent trend in web communication, news outlets and blog platforms allow readers to express opinions about what they have read by choosing an associated feeling, or sentiment expression. This emerging trend, fitting between liking and full text comments, has not still found the popularity it should. The thesis of this paper is that this is also due to the way the sentiment choice is presented to the user. In order to test this hypothesis we have devised a pilot experiment; results confirm that a simpler way of choice increases sentiment expression and yields result that are more aligned with ground truth.
A recent trend in web communication, aimed at improving
engagement of newspapers and blogs readers and favouring
the expression and exchange of feelings between them, is
emerging: a number of news outlets and blog platforms now
allow readers to express opinions about what they have just
read by indicating what kind of feeling – and in some
instances how intensely – they felt during perusal of the article
or news piece (examples of this can be seen e.g. at the
bottom of every article published by Corriere Della Sera
website, the most read Italian newspaper). Expressing sentiments
about a web text sits in the middle between liking
In liking users can express their support for content posted
by other users – be it an image, post, short text, video or
any other multimedia artefact – by pressing a specific
button, usually placed under the aforementioned content. The
naming is due to Facebook using the term like to label the
button, but under different nomenclatures (e.g. +1 for Google
Plus) the same function exists in almost any currently used
SN. Liking enjoys an immense popularity, and can certainly
be pinpointed as a fundamental feature of a truly social
internet context
As such, sentiment expression on the web fits a nice niche exactly where it is needed: simple enough to be really appreciated by users’ ephemeral attention, but better adept at describing and categorizing feelings toward multimedia material than the simple liking. Despite this, the emerging trend has not still found the popularity it should.
The thesis of this paper is that sentiment expression is not yet widely adopted in lieu of liking (also) because the typical GUI means employed to deploy it are not simple enough on the user; to test this hypothesis we have devised a simple pilot experiment, based on the idea of simplifying user choice to just one emotion at a click distance (the others being still available at a two-clicks distance) with the help of automatic classification techniques.
In this section we give a concise, general description of the basic elements used to build our experiment. The following section will describe in more detail the implementation techniques of such elements as used in the experiment.
Sentiment analysis is a set of techniques through which it is
possible to find, in an automatic fashion, the general
sentiment (or the opinion) expressed within a text.
Sentiment analysis is employable in various application fields. In marketing, for example when a new product is launched in commerce; the producers, analyzing the opinion of the users, can understand its popularity, if it is appreciated by the users or not, and why. Moreover, sentiment analysis can be important during elections; is possible to find if voters prefer one candidate over another through the opinions that they express, and so on.
The term opinion mining and sentiment analysis are often
used interchangeably, and with the same meaning intended;
but, as explained in
Emojis are iconic representations enjoying a vast popularity on the web, especially since the rise of mobile platforms and instant messaging tools.
Emoji became available in 1999 on Japanese mobile
phones, and, according to the Unicode Consortium,
“represent things such as faces, weather, vehicles and buildings,
food and drink, animals and plants, or icons that represent
emotions, feelings, or activities”. In 2007 Google completed
the conversion of “enhanced emoticons” to Unicode
privateuse codes, and in 2009 a set of 722 Unicode characters was
defined as the union of Japanese emoji characters. More
pictographs were added in 2010, 2012 and 2014.
As an example of the deployment of pictographs in
sentiment expression, in January 2013, the popular social
network Facebook allowed users to post what they were
feeling, watching, reading, drinking, eating, and more. Using a
drop-down menu, Facebook allowed users to express related
feelings through a sentiment list containing many elements,
such as happy, excited, loved, sad, crazy, blessed, perplexed,
amused, hopeful, and so on.
As already discussed in the introduction, what is gained in
depth and information retrieval is lost in terms of simplicity,
and the latter systems have not gained the universal
acceptance that liking has. Choosing among different pictograms
is no more a strictly atomic process, and requires direct
action and a modicum of thinking, which makes it transforming
into an habit a more difficult endeavour. While liking has
become a sort of automatic reflex, sentiment expression still has
to really catch on due to users’ general laziness and lack of
concentration or interest
Emotions play a central role in everyday interactions.
In
The experimental subjects panel was constituted by 48 degree students from CS faculty, enrolled by presence. Data were collected during March 2015. Participants were recruited on a voluntary basis and they were treated in accordance with the “Ethical Principles of Psychologists and Code of Conduct” (American Psychological Association, 1992). Gender distribution and mean age are shown in Table 3.
In order to have a base to which compare the results of our
test, we employed the Italian Sentiment Emotion Expression
database (I-SEEdb)
jIsj A å is i0 = is2Is0
jIs0j where Sfanger, disgust, fear, happiness, sadness, surpriseg is the set of emotions and is 2 Is are the intensities gathered (1) (2) for the sentiments. s0 and i0 are the emotion extracted and his associated intensity.
In order to perform our experiment we selected the twelve most polarized (i.e. highest scored) snippets, two from each category. A sample snippet is shown in Table 2.
It was a complex restoration operation the one done on ”La Fornarina” by Raphael Sanzio. A number of scientific and imagining experiments have been carried out, using technical methodologies such as x-rays, false-colour and reflectography snapshots.
Category: Tech
Score: fhappinessg, 4.0
The experiment is fully computer assisted. All the software is developed in house. Experimental subjects are introduced into a secluded room where the experiment takes place. They are seated in front of a computer screen and, after statistic data gathering (age, gender) the mechanics of the experiment is explained to them, aided by written instructions whose screenshot is shown in Figure 1.
Each subject is randomly assigned to one of the two possible experimental conditions, with the software taking care of the right balance between genders and age:
H0: after each snippet, six emojis representing the six basic Ekman’s emotions are shown, in random order, plus one button labelled ”Prosegui” (Next question). The subject can press on one of the emojis or on the button.
H1: after each snippet, one emoji representing the emotion associated with the high I-SEEdb score is shown, plus a small pop-up button and the ”Prosegui” button. The subject can click on the emoji, on the pop-up (in which case a list comprising the other five emotions, again espressed as emoji, plus a sixth emoj representing indifference will be shown) or on the button.
Screenshots from the two conditions are given in Figure 24. In Figure 4 the extended dropdown menu (invoked when the subject presses the triangle pop-up) is shown.
As soon as the subject chooses an emotion (or skips the choice by pressing the ”Prosegui” button, an option he is explicitly given in the instructions), another snippet is presented on screen. Each user is presented with the twelve different snippets.
Results from the experiment are shown in Table 3
Looking at the results of the two text we can see a
considerable difference, in condition H0 the user tend to mix
similar emotion, occasionally because of misunderstandings
about the correspondence of a specific emoji to the associated
emotion (as verified by post-experiment interviews), but more
often because sentiment could not be mutual exclusive and
could ”overlap” each other.
In the condition H1 the vast majority of subjects choose the suggested emoji despite having the possibility to select another one using the drop-down menu (at a one-click distance). In our opinion this is because, unlike in condition H0, showing only one isolated emoji helps the subject to more accurately recognize the feeling aroused by the reading of the article, and such emotion in most of the cases corresponds to that expressed by the suggested emoji.
Moreover, if we consider the number of neutral classifications of the snippets, H0 presents more neutral classifications WRT H1. This fact may be a consequence of what was already said before about emoji’s misunderstanding plus the non-mutual exclusivity of sentiments. As a partial confirmation, subjects that are not able to clearly make a choice prefer to remain neutral, and choose the neutral emoji. This phenomenon is less pronounced in condition H1, where the showing of just one emoji may be of help in order to recognize the appropriate feeling and express it.
In synthesis, results shown that presenting only one preclassified emoji instead of the whole gamut (which is always an option, albeit it requires a bit more effort to be exercised) produces results that are more in line with ground truth, and, what is more important for the correct deployment of such a system in the real world, increases significantly the number of times users effectively express the emotion, leading to a better engagement of the user and to more data gathering for the system to work on.
As presented in this pilot experiment, the system is not directly implementable in a real world application: preclassification of the articles is obviously impossible in social networks or news outlets where big data are produced and published. In a real application this problem could be solved in at least two different ways. One approach would be to use automatic classification algorithms to derive a ground truth to use in lieu of the human pre-classification. Even sub-optimal solutions, as the ones readily available today, would give a good headstart to initialise the algorithms, and the missing information could be gathered by the sentiment expression choices of the first users. Another solution would be to adopt an hybrid system, where the first classifications would be done using a condition similar to H0, and as soon as sufficient data is gathered to form some sort of ground truth the system would switch its interface to a single-emoji-plus-dropdown mode. In the context of SN we also need to consider the previous interactions between users, as strong feelings may slightly alter the meaning of the expressed emotion. Furthermore the readers that habitually does not contemplate liking can be encouraged by the system to express themselves, and as such an eventual risk of uniformity given by the use of a restricted set of feelings would be balanced by the increase in feeling expression. Further improvement of this methodology could be obtained by employing fuzzy techniques that treat degrees of emotion as Fuzzy Sets. (Tabacchi & Termini, 2014; Termini & Tabacchi, 2014; Seising & Tabacchi, 2013; Tabacchi & Termini, 2013, 2012)
Further research work will be devoted to increase the cardinality of the experiment and to a better integration of automatic and manual pre-classification of texts, in order to further improve the quality of ground truth and to build toward a real-world implementation of the system.
This work has been partially supported by P.O. F.E.S.R Sicilia 2007-2013 Progetto POSING - Investiamo nel vostro futuro – l’Europa cresce in Sicilia. rial. In D. Ciucci, J. Montero, & G. Pasi (Eds.), Proceedings of the 8th conference of the european society for fuzzy logic and technology. Atlantis Press. doi: doi:10.2991/eusflat.2013.30 Tabacchi, M. E., & Termini, S. (2014). Some reflections on fuzzy set theory as an experimental science. In A. Laurent, O. Strauss, B. Bouchon-Meunier, & R. R. Yager (Eds.), Information processing and management of uncertainty in knowledge-based systems (Vol. 442, p. 546-555). Springer International Publishing.
Termini, S., & Tabacchi, M. E. (2014). Fuzzy set theory as a methodological bridge between hard science and humanities. International Journal of Intelligent Systems, 29(1), 104-117. doi: 10.1002/int.21632 Wallbott, H. G., & Scherer, K. R. (1986). Cues and channels in emotion recognition. Journal of personality and social psychology, 51(4), 690.