V. RESULTADOS
5.1 Resultados descriptivos
So far we have explored the more tangible and objective contextual informa-tion that machines can sense. From a computainforma-tional standpoint, the main challenge in sensing subjective context in practice is that it usually cannot be fully captured based on low dimensional sensory input. In other words, there is not a known magical mathematical model to put into the machine’s mind as a program to figure out subjectivity in context. Even for humans, inferring
subjectivity (e.g., empathy) is a challenging task in itself. For example, an emo-tional state such as hidden sadness cannot easily be understood from singular observations such as appearance. One needs to focus on a friend’s out-of- the-ordinary behaviors, such as how much she speaks or if she smiles as usual to jokes made, in order to suspect if she might be sad. In the computational realm, to make a robust prediction about the human state, the system needs to utilize any relevant information, such as history, experiences, health condition, loca-tion, time, and so forth. However, even after having all this informaloca-tion, making an inference about emotional state can be different from one person to another and might need to be learned and personalized for each individual. Neverthe-less, it has to start somewhere.
In that regard, affective computing is the field of computational science that devotedly studies human subjectivity and its computability. Drawing from the previous suggested typology, it can be said that the affective computing field directly operates in the heart of subjective context. In domain terminology, af-fect is used as an umbrella term that covers broad range of feelings that people experience, whereas emotions are deemed to be intense feelings directed toward something or someone. Additionally, moods refer to feelings less intense than emotions. Mood and emotion differentiation is important because experts be-lieve that emotions are more volatile and transient than mood, but moods can affect the characteristics of emotions. For example, during a bad mood, anger might not go away easily.
Zhang and Hui (2014) indicate that powerful sensors along with long-term usage of smart devices can enable unobtrusive collection of affective data, which in return is expected to improve traditional affective computing research [26]. As a result, researchers are seeking to explore new methodologies to infer affective states using the newly available data (e.g., touch behaviors, usage events, etc.).
We choose to demonstrate the advantages in computational detection of emo-tions by demonstrating examples of the academic research.
One of the most recent efforts in the area investigates the possibility of rec-ognizing emotions by making use of low-level acoustic features of speech. Han, You, and Tashev (2014) explored the potential of the deep-learning approach to investigate utterance-level emotion detection [27]. As opposed to the feature-modeling approach, use of low-level acoustic features makes the methodology potentially applicable in cross-language and cross-cultural settings.
Apart from emotion recognition, in recent years there has been a surge of in-terest in computational methods for opinion mining and subjectivity and senti-ment detection. Balahur et al. (2014) indicate that these methods typically focus on the identification of private states, such as opinions, emotions, sentiments,
evaluations, beliefs, and speculation, in natural language [28]. Subjectivity is usually classified as subjective or objective, whereas sentiment classification tries to add depth to analyses by classifying the text as either positive, negative, or neutral. Yet another important aspect in this body of research is the type of text, such as short messages, preferences, events, comments, and opinions in social media sites (e.g., Facebook, Twitter). Tailored automated analysis of these sources could be of great use to obtain the real-time unbiased opinions and emo-tions of the masses.
Personality is also an equally important aspect for human affect research.
Chittaranjan, Blom, and Gatica-Perez (2011) studied the relationship between behavioral characteristics derived from rich smartphone data and self-report-ed personality traits [29]. The analysis showself-report-ed that aggregatself-report-ed features from the obtained data can be indicators of Big Five personality traits, a concept well known in psychology research. These traits are claimed to capture most of the individual differences among people, and hence are fit to be used as a personal-ity measure.
In conjunction with Chittaranjan et al.’s (2011) study [29], Staiano, Lepri, Ahorony, Pienesi, Sebe, and Pentland (2012) bring another angle to the same picture [30].
Fig. 6. Nonverbal behavioral cues and social signals. With no more than these two silhouettes, it is not difficult for most people to guess that the picture portrays a couple involved in a fight. Nonverbal behavioral cues allow one to understand that the social signals being exchanged are disagreement, hostility, aggressiveness, and so forth, and that the two persons have a tight relationship (adapted from Vinciarelli et al., 2012 [31]).
They refer to usage-based analysis as an “actor-based” feature, whereas they propose a “network-based” feature study. They claim that network parameters such as number of calls made or received, their average duration, the total dura-tion of calls, the number of missed calls, and Internet usage can be predictive of personality traits. In this regard they define two different networks that can be constructed from the data collected from an individual smartphone. The first is based on distant communication, such as calls. The second is based on proximity to others utilizing the Bluetooth (BT) discovery mechanism, wherein the number of unique BT IDs discovered determines the size of the network. Consequently, they analyze the structural differences of call and BT networks to show that their relation is a predictor of Big Five personality traits.
Finally, a relatively new domain in human affect research is social signal processing (SSP). The aim of SSP is to bridge the social intelligence gap between humans and machines. Vinciarelli, Pantic, Heylen, Palacheud, Poggi, D’Errico
& Schröder (2012) define social signal as communicative or informative signal that, either directly or indirectly, provides information about social facts, namely, social interactions, social emotions, social attitudes, or relations [31] (Figure 6).
The state of the art in this area deals with problems such as social emotion rec-ognition, role recognition (e.g., dominant), analysis of (dis-)agreement, group dynamics, and negotiation outcome.