1. Introduction and Background

1In recent years we had great advancements in the analysis of communication, in face to face meetings as well as in Online Social Networks (OSN) (Boyd and Ellison, 2007). For example, resources for computational psycholinguistics like the Linguistic Enquiry Word Count (LIWC) (Tausczik and Pennebaker, 2010), have been applied to OSN like Facebook and Twitter for personality recognition tasks (Golbeck et al., 2011) (Schwartz et al., 2013) (Celli and Polonio, 2013) (Quercia et al., 2011). Interesting psychological research analyzed the motivations behind OSN usage (Gosling et al., 2011) (Seidman, 2013) and whether user profiles in OSN reflect acual personaliy or a selfidealization (Back et al., 2010).

2Also Conversation Analysis (CA) of face to face meetings, that has a long history dating back to the ’70s (Sacks et al., 1974), has taken advantage of computational techniques, addressing detection of consensus in business meetings (Pianesi et al., 2007), multimodal personality recognition (Pianesi et al., 2008) and dectection of conflicts from speech (Kim et al., 2012).

3In this paper we make a comparison of the linguistic behaviour of OSN users both online and in face to face meetings. To do so, we collected Facebook data from 11 volunteer users, who participated to an experimental setting where we recorded face to face multiparty conversations of their meetings. Our goal is to discover relationships between a rich set of psycholinguistic dimensions (Tausczik and Pennebaker, 2010) extracted from Facebook metadata and meeting transcriptions. Our contributions to the research in the fields on Conversation Analysis and Social Network Analysis are: the release of a corpus of speech transcriptions aligned to Facebook data in Italian and the analysis of correlations between psycholinguistic dimensions in the two settings.

4The paper is structured as follows: in section 2 we describe the corpora and the data collection, in section 3 we explain the method adopted and report the results, in section 4 we draw some conclusions.

2. Data and Method

5We collected 11 volunteer Italian native speakers, who provided the consent to use their Facebook metadata, and organized meeting sessions with them to collect spoken linguistic data. The meetings consist in sessions of one hour, where participants, 6 in the first session and 5 in the second one, performed free multi-party conversations. Groups were balanced by gender and aged between 18 and 50 years. There were no restrictions, predefined task or topic to elicitate speech. In order to prevent biases in the interactions we put in the groups persons who do not know each other.

6We recorded and manually transcribed a corpus of spoken conversations from the meeting sessions, splitting utterances by turns where a speaker ends its speech or is interrupted by another speaker. Then we annotated each utterance with dialogue act (DA) labels. To select DA labels we referred to Novielli & Strapparava (Novielli and Strapparava, 2010), who performed a dialogue act annotation on meetings transcriptions in Italian. We just added the label ”laugh” to their label set. The final dialogue act label set we used is reported in Table 1.

7The agreement on the annotation of dialogue act labels between 2 non-expert labelers is k = 0.595 (Fleiss et al., 1981). This moderate agreement score, and the feedback from the annotators, indicate that the task is hard due to the presence of long and complex utterances.

8We aligned the data from spoken conversations with public data from the participants’ Facebook profiles. Using Facebook APIs, we collected data from 6 months before the meeting session to 1 year later. We collected public status updates, including text messages, links, pictures, and multimedia files posted and received on the participants’ walls. We distinguished between statuses posted by the users and statuses posted on the users’ wall by others. Eventually we computed the numerical metadata reported in table 2 and we analyzed the textual pots.

9We anonymized both the transcription and the Facebook data. The final corpus contains 2 audio files (one hour each) with transcriptions (about 21000 tokens and about 1600 utterances in total; 1750 words and 133 utterances on average per participant), and Facebook data of the participants (about 80000 tokens, about 5800 posts including multimedia status updates). We automatically annotated the textual data in the corpus with the Italian version of LIWC (Alparone et al., 2004). Doing so, we annotated words with 85 psychological dimensions, such as linguistic categories (verbs, prepositions, future tense, past tense, swears, etc.), psychological processes (anxiety, anger, feeling, cognitive mechanisms, etc.), and personal concerns (money, religion, leisure, TV, achievement, home, sleep. etc.). In the next section we report the results of the analysis of the data collected.

3. Experiments and Results

10Scope From a communication analysis perspective, face to face meetings and Facebook are two very different settings: in Facebook the communication is written, asynchronous, mediated and with an audience that is a mix of friends and unknown people. On the contrary in face to face meetings the communication is oral, synchronous, not mediated, and the audience is unknown people. In a theory of communication (Shannon and Weaver, 1949), illustrated in Figure 1, all those levels are variables related to the sender, receiver and medium. Here we restrict the scope of this work to the analysis of message level, leaving to future work the possibility to extend this analysis to the characteristics of the media or the participants.

Figure 1: Schema of communication as transmission. We limit the scope of this work to the message level

11Experiments First of all we analyzed the topics in Facebook and meeting transcriptions. We removed the stopwords and we generated two word clouds with the 70 most frequent words in each dataset with 5 as minimum term frequency. We report the word clouds in Figure 2. The comparison of the two clouds reveal that participants to the experiments in Facebook discussed and planned actions (“dormire”, “andare”) places (“rimini”, “copenhagen”) and times (“sera”, stasera”, “domani”) while in meetings they told and discussed mainly about places (“bologna”, “rimini”) and people (“tipo”, “gente”).

Figure 2: Word clouds of the 70 most frequent words in meeting transcriptions and Facebook data

12In order to discover relationships between psycholinguistic dimensions in Facebook and face to face meetings, we labelled the texts with LIWC, and we computed how much the psycholinguistic dimensions correlate in the two settings. We observed few, but strong, significative correlations (for significative we mean correlations with pvalue smaller than 0.05 and correlation greater than 0.5), reported in table 3.

13The dimensions with strong correlation are related to powerful emotions, difficult to control, like anxiety and anger, but also to the tendency to express feelings and emotions with words. Swears, that is the dimension with the highest combination of correlation coefficient and significance, is related as well to a dimension difficult to control. Maybe less interesting for our purposes are other dimensions with high correlations related to the content of discourse, like “home”, “TV”, “future” and “sleep”. We ran automatic topic modeling with a Hierarchical Latent Dirichlet Allocation (Teh et al., 2006) (Blei et al., 2003) to reveal that participants spoke about “TV” and “sleep” in both settings, but about “home” and “future” only in Facebook and not in face to face meetings. This is why these values are negative.

14We also compared behavioral data from Facebook and meetings. In particular we computed the correlations between Facebook metadata and dialogue acts annotated in meeting transcriptions, plus metadata from face to face meetings, namely the average length of utterances in words and characters. Results, reported in Table 4, show that there are few, but very interesting, significative correlations. The number of likes received by the participants on Facebook correlate positively with a tendency to ask questions in meetings. This is quite surprising and perhaps reveals a will to engage the audience asking questions. Crucially, other significative correlations are related to shares generated in Facebook by the participants. In particular this is correlated with long statements in face to face meetings. In practice, people posting contents that are reshared online, in face to face meetings tend to produce long statements and talk more than the others.

Table 4: Pearson’s correlations between metadata from Facebook and dialogue act labels from face to face meetings. *=p-value smaller than 0.05.

4. Discussion and Conclusions

15In this paper, we attempted to analyse the correlations between psycholinguistic dimensions observed in Facebook and face to face meetings. We found that the type of words significantly correlated to both settings are related to strong emotions (anger and anxiety), We suggest that these are linguistic dimensions difficult to control and tend to be constant in different settings. Crucially, we also found that likes received on Facebook are correlated to the tendency to ask questions in meetings. Literature on impression formation/management report that people with high self-esteem in meetings will elicit self-esteem enhancing reactions from others (Hass, 1981). This could explain the link between the tendency to ask questions in meetings with unknown people and the tendency to post contents that elicit likes in Facebook. Moreover, the tendency to ask questions in spoken conversations is correlated to observed emotional stability (Mairesse et al., 2007) and that emotionally stable users in Twitter tend to have more replies in conversations than neurotic users (Celli and Rossi, 2012). We suggest that the correlation we found can be partially explained by these two privious findings.

16Another very interesting finding is that the tendency to be reshared on Facebook correlates to the tendency to speak a lot in face to face meetings. Again, literature about impression formation/management can explain this, because people with high self-esteem tend to engage people and to speak a lot, while people adopting defensive strategies tend to be assertive less argumentative. In linguistics it is an open debate whether virality depends from the influence of the source (Zaman et al., 2010) or the content of message being shared (Guerini et al., 2011) (Suh et al., 2010). In particular, the content that evokes higharousal positive (amusement) or negative (anger or anxiety) emotions is more viral, while content that evokes low arousal emotions (sadness) is less viral (Berger and Milkman, 2012). Given that the tendency to express both positive and negative feelings and emotions in spoken conversations is a feature of extraversion (Mairesse et al., 2007), and that literature in psychology links the tendency to speak a lot to extraversion (Gill and Oberlander, 2002), observed neuroticism (Mairesse et al., 2007) and dominance (Bee et al., 2010). we suggest that the correlation between long turns in meetings and highly shared contents in Facebook may be due to extraversion, dominance and high self-esteem.

17We are going to release the dataset we collected on demand.