• No results found

Sentic Activation: A Two-Level Affective Reasoning Framework

Current thinking in cognitive psychology suggests that humans process information at a minimum of two distinct levels. There is extensive evidence for the existence of two (or more) processing systems within the human brain, one that involves fast, parallel, unconscious processing, and one that involves slow, serial, more conscious processing [185, 186, 187, 188]. Dual-process models of automatic and controlled social cognition have been proposed in nearly every domain of social psychology.

Evidence from neurosciences supports this separation, with identifiably different brain regions involved in each of the two systems [189]. Such systems, termed U-level

are most effective in different contexts. The former, in particular, works intuitively, effortlessly, globally, and emotionally (subsection 4.3.1). The latter, in turn, works logically, systematically, effortfully, and rationally (subsection 4.3.2).

4.3.1 Unconscious Reasoning

In recent years, neuroscience has contributed a lot to the study of emotions through the development of novel methods for studying emotional processes and their neural correlates. In particular, new methods used in affective neuroscience, e.g., fMRI, le- sion studies, genetics, electro-physiology, paved the way towards the understanding of the neural circuitry that underlies emotional experience and of the manner in which emotional states influence health and life outcomes. A key contribution in the last two decades has been to provide evidence against the notion that emotions are subcortical and limbic, whereas cognition is cortical.

This notion was reinforcing the flawed Cartesian dichotomy between thoughts and feelings [190]. There is now ample evidence that the neural substrates of cognition and emotion overlap substantially [191]. Cognitive processes, such as memory encoding and retrieval, causal reasoning, deliberation, goal appraisal, and planning, operate continu- ally throughout the experience of emotion. This evidence points to the importance of considering the affective components of any human-computer interaction [19]. Affective neuroscience, in particular, has provided evidence that elements of emotional learning can occur without awareness [192] and elements of emotional behaviour do not require explicit processing [193]. Affective information processing, in fact, mainly takes place at unconscious level (U-level) [188].

Reasoning, at this level, relies on experience and intuition, which allow considering issues intuitively and effortlessly. Hence, rather than reflecting upon various consid- erations in sequence, the U-level forms a global impression of the different issues. In addition, rather than applying logical rules or symbolic codes (e.g., words or numbers), the U-level considers vivid representations of objects or events. Such representations

are laden with the emotions, details, features, and sensations that correspond to the objects or events. Such human capability of summarising the huge amount of inputs and outputs of previous situations to find useful patterns that might work at the present time is hereby implemented by means of AffectiveSpace. By reducing the dimension- ality of the matrix representation of AffectNet, in fact, AffectiveSpace compresses the feature space of affective common sense knowledge into one that allows to better gain global insight and human-scale understanding.

In cognitive science, the term ‘compression’ refers to transforming diffuse and dis- tended conceptual structures that are less congenial to human understanding so that they become better suited to our human-scale ways of thinking. Compression is hereby achieved by balancing the number of singular values discarded when synthesising Af- fectiveSpace, in a way that the affective common sense knowledge representation is neither too concrete nor too abstract with respect to the detail granularity needed for performing a particular task. The reasoning-by-analogy capabilities of AffectiveS- pace, hence, are exploited at U-level to achieve digital intuition about the input data. In particular, the vector space representation of affective common sense knowledge is clustered according the Hourglass model using the sentic medoids technique, in a way that concepts that are semantically and affectively related to the input data can be intuitively retrieved by analogy and unconsciously crop out to the C-level.

4.3.2 Conscious Reasoning

U-level and C-level are two conceptual systems that operate by different rules of in- ference. While the former operates emotionally and intuitively, the latter relies on logic and rationality. In particular, the C-level analyses issues with effort, logic, and deliberation rather than relying on intuition. Hence, while at U-level the vector space representation of AffecNet is exploited to intuitively guess semantic and affective rela- tions between concepts, at C-level associations between concepts are made according to the actual connections between different nodes in the graph representation of affective

common sense knowledge. Memory is not a ‘thing’ that is stored somewhere in a mental warehouse and can be pulled out and brought to the fore. Rather, it is a potential for reactivation of a set of concepts that together constitute a particular meaning. Associa- tive memory involves the unconscious activation of networks of association–thoughts, feelings, wishes, fears, and perceptions that are connected, so that activation of one node in the network leads to activation of the others [194].

Sentic activation aims to implement such process through the ensemble applica- tion of multi-dimensionality reduction and graph mining techniques. Specifically, the semantically and affectively related concepts retrieved by means of AffectiveSpace at U-level are fed into AffectNet in order to crawl it according to how such seed concepts are interconnected to each other and to other concepts in the semantic network. To this end, spectral association [195] is employed. Spectral association is a technique that assigns values, or activations, to seed concepts and spreads their values across the AffectNet graph. This operation, an approximation of many steps of spreading activation, transfers the most activation to concepts that are connected to the seed concepts by short paths or many different paths in affective common sense knowledge. These related concepts are likely to have similar affective values. This can be seen as an alternate way of assigning affective values to all concepts, which simplifies the process by not relying on an outside resource such as WNA. In particular, a matrix A that relates concepts to other concepts, instead of their features, is built and the scores are added up over all relations that relate one concept to another, disregarding direction. Applying A to a vector containing a single concept spreads that concept’s value to its connected concepts. Applying A2 spreads that value to concepts connected by two links (including back to the concept itself). But the desired operation is to spread the activation through any number of links, with diminishing returns, so the operator wanted is:

1 + A +A 2 2! + A3 3! + ... = e A (4.7)

symmetric, so instead of applying Lanczos’ method [174] to AAT and getting the SVD,

it can be applied directly to A to obtain the spectral decomposition A = V ΛVT.

As before, this expression can be raised to any power and everything but the power of Λ cancelled. Therefore, eA = V eΛVT. This simple twist on the SVD allows to calculate spreading activation over the whole matrix instantly. As with the SVD, these matrices can be truncated to k axes and therefore space can be saved while generalising from similar concepts. The matrix can also be rescaled so that activation values have a maximum of 1 and do not tend to collect in highly-connected concepts such as ‘person’, by normalising the truncated rows of V eΛ/2to unit vectors, and multiplying that matrix

by its transpose to get a rescaled version of V eΛVT.

Spectral association can spread not only positive, but also negative activation val- ues. Hence, unconscious reasoning at U-level is exploited not only to retrieve concepts that are most semantically and affectively related, but also concepts that are most likely to be unrelated with the input data (lowest dot product). While the former are exploited to spread semantics and sentics across the AffectNet graph, the latter are used to contain such activation in a way that potentially unrelated concepts (and their twins) do not get triggered. Such brain-inspired ensemble application of dimensionality reduction and graph mining techniques (hereby referred as unconscious and conscious reasoning, respectively) allows sentic activation to more efficiently infer semantics and sentics from natural language text. In fact, sentic activation was tested on the bench- mark for affective common sense knowledge (BACK) built by means of CF-IOF. In particular, BACK’s concepts were compared with the classification results obtained by applying the AffectiveSpace process (U-level), spectral association (C-level) and sen- tic activation (ensemble of U-level and C-level). Results showed that sentic activation achieves +13.9% and +8.2% accuracy than the AffectiveSpace process and spectral association, respectively.

4.4

SenticNet: A Publicly Available Semantic Resource