English. We report the results of an exploratory study aimed at investigating the language of happiness in Italian tweets. Specifically, we conduct a time-wise analysis of the happiness load of tweets by leveraging a lexicon of happiness extracted from 8.6M tweets. Furthermore, we report the results of a statistical linguistic analysis aimed at extracting the most frequent concepts associated with the happy and sad words in our lexicon.
The widespread diffusion of social media has
reshaped the way we interact and communicate.
Among others, microblogging platforms as
Twitter are becoming extremely popular and people
constantly use them for sharing opinions about
facts of public interest. Furthermore, its
worldwide adoption and the fact that tweets are publicly
available, makes Twitter an extremely appealing
virtual place for researchers interested in language
analysis as a mean to investigate social
phenomena
In addition, recent research showed how
microblogging is also used for self-disclosure of
individual feelings
In this study we report the results of an
exploratory analysis of the language of happiness in
Twitter. In particular, we perform a partial
replication of the approach proposed by
2.1
Our study is based on TWITA
From the Chappy corpus, we extract a subset of words and we assign them an happiness factor (hf ) computed according to the log of the odds ratio between the probability that the word occurs in positive tweets phappy(wi) and the probability that it occurs in negative tweets psad(wi) as in Eq. 1.
hf (wi) = log phappy(wi) psad(wi) (1) We adopt additive smoothing (Laplace smoothing) for computing both phappy and psad probabilities. In our lexicon, we include and compute the happiness factor only for words that occur at least 10,000 times, for a total of 718 words. We call this list “the happiness dictionary” (Dh)3. Table 1 reports the most happy/sad words with the corresponding happiness factor (score(hf)).
We observe that some happy words (fback, ricambi, benvenuta) are due to several positive tweets that users post when they establish new connections, i.e. when they start following a 2We use :-) and :) for happy and :-( and :( for sad. 3The dictionary is available on github https:// github.com/pippokill/happyFactor new user or when they ask sombebody to follow them back (fback) as in: @usermention ciao sono nuova, fback? Grazie mille :) Sad words refer to negative emotions or evaluations, such as triste, dispiace, brutto, peccato. Interestingly, several negative words emerge from the school domain (compiti, studiare) and the word scuola has a negative score of -0.93 itself. 2.3
We are interested in understanding how happiness
words map into psycholinguistic word classes.
Hence, we check their distribution along the word
categories in the Linguistic Inquiry and Word
Count (LIWC) taxonomy
We perform a coding of the English translation of the happy/sad words into LIWC categories. When translating, we keep the information about the subject conveyed by the Italian verbs (e.g., ’penso’ is translated as ’I think’). The coding is performed manually by the authors: in a first round, one rater associates each word with the corresponding LIWC category; then, the other revises the annotation, checking for consistency and verifying also the correctness of the translation. 22 words are discarded and replaced with others from the dictionary because we could not find a mapping with any of the categories. Furthermore, we add an ad hoc category to enable modeling of words from the social media domain (retweet, follow).
Figure 1 shows how the happy and sad words distribute along the dimensions associated with the most frequent categories. Sample words for each word category are reported in Table 2. We observe that happy words in the dictionary mainly refer to positive emotions as well as to the social and social media dimensions. Conversely, sad words mainly describe negative emotions with focus on the author. Words describing cognitive mechanisms are also associated with sadness. As observed in the original study, happiness is not constant in our life and different degrees of happiness might be observed at different moments in time. As such, we analyze how happiness changes over time. In particular we take into account the days of the week and the different hours in a day. For this analysis, we exploit the whole corpus of 400M tweets and we compute the distribution (a) Happiness load by day of the week (b) Happiness load for a 24-hour day of words occurring in the happiness dictionary in each different time period. Using this strategy, in each time period the word has an happiness load obtained by multiplying its frequency in that period by its happiness factor. The happiness load of each time period is the average of all the happiness load in that period. The obtained values are mapped in the interval [-1, 1] and plotted in Figure 2a (for days) and in Figure 2b (for hours).
Our time-wise analysis reveals a drop in
happiness on Thurdsay, with a subsequent twist towards
positive mood on Friday, before the weekend that
is the happiest moment in the week. This is
consistent with the findings of the original study
reporting mid-week blues around Wednesday and a
happiness peak on Saturday
We are interested in concepts related to words in the happiness dictionary. In the original study, the authors extract the ’ingredients’ for their recipe of happiness by ranking the most relevant 2- and 3grams from their corpus according to their happiness load. Such an approach is not easy to replicate as the number of 2- and 3-grams extracted from 400M tweets is potentially huge. Hence, starting from the words in our happiness dictionary, we extract the most 50 co-occurring words in a window of two words. Then we rank all the word pairs (dictionary word, co-occurring word)4 according to the Pointwise Mutual Information (PMI) multiplied by the happiness factor. Table 3 reports some of the most happy and sad pairs.
Starting from word pairs, we perform another
kind of analysis aiming at mapping the words
occurring in each pair with super-senses in WordNet.
A super-sense is a general semantic taxonomy
defined by the WordNet lexicographer classes as a
way for defining logical aggregation of senses in
each syntactic category. We assign a happiness
score to each super-sense by averaging the
happiness factor associated with the dictionary word
in the pair. Since each pair contains a dictionary
word and a co-occurring word, we map the
cooccurring word to its super-sense and increment
the score of the super-sense by summing the
happiness factor associated with the dictionary word.
Finally, the score of each super-sense is divided
by the number of the co-occurring words
belonging to the super-sense. For ambiguous words, we
select the super-sense associated with the most
frequent sense. In this study, we do not rely on
a Word Sense Disambiguation (WSD) algorithm
since WSD is a critical task. We need to test
the WSD performance on tweets before to use
it. Generally, WSD algorithms give performance
slightly above the most frequent sense. We plan
to test WSD in a further study. As super-senses
are defined in the English version of WordNet, we
4We do not take into account the word order in the pairs.
performed a mapping of Italian words to the
English WordNet through the use of both Morph-it!
In Table 4 we report the most happy and
sad super-senses with the most frequent words
extracted by our corpus. Consistently with
the evidence provided by the analysis of the
psycholinguistic word categories (see Section
2.3), we observe that socialness is
associated with positive feelings, with concepts
referring to people (noun.person) and communication
(verb.communication, noun.communication)
scoring high in happiness. Food (noun.food) also
seems to be a major cause of positive mood, as
well as money and gifts (noun.possession), sport
achievements (’vittoria and ’gol’ in noun.act),
and mundane locations and events (’centro’,
’piazza’, ’concerto’, ’viaggio’ in noun.location and
noun.act). This is consistent with suggestion by
Also for sadness, results confirm findings emerging from the analysis of psycholinguistic categories in LIWC. In fact, we observe that people tend to report their own individual negative feelings (rido, piango in verb.body), thoughts (verb.cognition), perceptions (e.g., ’vedo’, ’sento’), and personal needs (’bisogno’ and ’sonno’ in noun.state). We observe also stereotypical complaints about weather (piove) as well as swear words (noun.body). 5
We performed an exploratory analysis of the lexicon and concepts associated with happiness in Italian tweets. We leveraged a corpus of happy and sad tweets to extract a ”happiness dictionary’, which we use to perform a time-wise analysis of happiness on Twitter and to extract the most frequent concepts and psycholinguistic categories associated to positive and negative emotions.
This study is a partial replication of the
previous one by
Our analysis of psycholinguistic categories and the extraction of concepts and WordNet supersenses associated with them reveals interesting findings. Happiness appears related to the social aspects of life while sad tweets mainly revolves around self-centered negative feelings and thoughts. In addition, our-time wise analysis reveals a mid-week drop in happiness also observed in the original study. We also observe that happiness is high in the morning and decreases over the day. As a future work, it would be interesting to investigate if time-wise analysis based on hours produces consistent results if a weekday or the weekend is considered and if emotion-triggering concepts associated with happiness also vary over time.
We are aware of the main limitations of this
study. First of all, by relying on microblogs we
are probably able to mine emotion triggers that
do not necessarily coincide with those shared in
daily face-to-face conversations or reported in
private logs. Furthermore, we do not attempt to make
any categorization of the authors of tweets.
Indeed, different target user groups could be studied
to fulfill specific research goals and enable
perspective applications, i.e. for supporting creative
writing or for providing personalized
recommendations based on moods. Finally, we consider only
Twitter as a source of data. The same methodology
could produce different results if applied to other
social media. Indeed, recent research
Based on these observations, further replications
could focus on finer-grained emotions, also
leveraging corpora from different platforms and
including consideration of demographics and
geographical information