Inference of Co-occurring Classes: Multi-class and Multi-label Classification - Computational Intelligence Paradigms in Advanced Pattern Classification - page 187

The term Affective states represents a wide range of human emotional and so-
cial behavior, including emotions, moods, attitudes, mental and knowledge states,
beliefs, desires and more. This means that several affective states are likely to oc-
cur simultaneously, and change asynchronously over time. Nuances of affective
state and subtle dynamic changes also frequently occur.
The recording and annotation of affective data pose various challenges. Al-
though it is desirable to annotate data recorded in real settings, most settings evoke
only a limited range of affective states. The majority of affective states and affec-
tive behavior comprises nuances of expressions that relate to subtle affective states,
often threshold values of various features distinguish between the recognition of af-
fective states when gradual transitions are tracked over time. The various affective
states can occur simultaneously and change asynchronously during the course of a
sustained interaction, and even during the course of a single utterance or a sentence.
Therefore, the annotation of affective states is more complicated and less reliable
than the annotation of objects in an image for example. The annotation depends not
only on the variety of affective behavioral expressions due to personality and con-
text, but also on the perception of the annotator, which is affected by personality,
cultural background, emotional and social intelligence and so on. So the problem is
not only related to finding synonyms within an extended ontology, but rather to
getting an agreement regarding the conceptual content, this is extenuated because
of the co-occurring nature of affective states. Various methods have been devised
for annotation of affective states. These usually rely on a pre-defined taxonomy,
and then associating data samples with the given labels, or recording in advance
samples for each label and then verifying that indeed the recorded samples
represent their labels. The desire for the technology though is to represent a large
variety of affective states, their variations, transitions and nuances.
As described in the section on annotation, there are various taxonomic methods
for describing affective states and the relations between them [57]. The taxonomic
method which was chosen is the prototype method. This provides the widest range
of affective states, in a manner which is comprehensible both for humans and ma-
chines. The Mind Reading taxonomy and database comprise 514 affective states
grouped into 24 meaning groups (the results presented here are based on a beta
version which comprised of 712 affective states). Each affective state in the data-
base is represented by six sentences. This is not enough for statistical learning. On
the other hand, each meaning group comprises many affective states that share a
meaning. Using groups of affective states as the classes has several advantages
[17, 31, 59, 60]: It increases the number of samples per class. The annotation of
affective states is not accurate, gathering several affective states that share a mean-
ing and the respective samples increase the overall accuracy of the wider class.
This also means that the application can be shared between cultures that share the
broader class definitions if not the single affective state concepts, i.e. the term
thinking exists in most cultures, but various types of thinking sometimes belong
only to a specific culture, for example wool-gathering in UK English.