The web is becoming a hybrid space where men and machines interact and many human and social activities are now mediated through the web. In this context, detecting and managing the emotional state of a user is important to allow software and other users to adapt their reactions and also to manage the community and the content it produces. As a typical example, Wikipedia is managed by users and bots who constantly contribute, agree, disagree, debate and update the content of the encyclopedia.
To efficiently manage and interact with such a hybrid society we need to improve our means to understand and link the different dimensions of the exchanges:
- Social dimension: what are the social links and social structures in place in these societies? How do we detect these structures (e.g. overlapping communities)? What kind of metrics are relevant on these structures to understand them and manage them?
- Textual dimension: what kinds of textual content and messages are produced and exchanged in these forums, wikis, discussions pages, microblogs, etc.? How can we analyze in an automated way these heterogeneous corpora, and which insightful information can we extract from them?
- Dialogical dimension: what kind of exchanges take place in the community? Which arguments proposed by a user attack or support the arguments proposed by the other users? Which kind of connotation does an interaction have (e.g. a debate, a question-answering, a fight)?
- Emotional dimension: what are the emotional states of the human participants at different stages of collaboration and discussions? How do these states evolve? Do they spread in the community and if so, how? How are emotions reflected in the language used in textual exchanges? Which kinds of sensors can we use to detect emotions? How can we modify the influence of negative emotions in order to improve global discussions and decisions?
Beyond the challenges individually raised by each dimension, a key problem is to link these dimensions and their analysis together to detect, for instance, a debate turning into a flame war, a content reaching an agreement, a good or bad emotion spreading in a community.
In the SEEMPAD joint team we decided to focus on a very precise goal, i.e. generating, annotating and analyzing a dataset that documents a debate. We aim at synchronizing several dimensions:
- Social links (intensity, alliances, etc.)
- Interactions happening (who talks to whom)
- Textual content of the exchanged messages
- Social-based semantic relations among the arguments
- Emotions, polarity, opinions detected from the text
- Emotions, physical state detected from sensors