Research articleApplication of multi-dimensional scaling and artificial neural networks for biologically inspired opinion mining
Introduction
Emotions are intrinsically part of our mental activity and play a key role in decision-making and cognitive communication processes. They are special states of the mind, shaped by natural selection, for adjusting various aspects of human organism in a way that it can better face particular situations, e.g., anger evolved for reaction, fear evolved for protection, and affection evolved for reproduction. Therefore, emotions cannot be shelved in the development of intelligent systems: in order for a machine to be really intelligent, it has to possess the ability to recognise, understand, and express emotions. To this end, a great number of emotion categorisation models and emotion-sensitive systems has been developed in recent years for performing tasks such as affect recognition and polarity detection.
In the context of sentic computing1 (Cambria and Hussain, 2012), in particular, graph mining techniques and multi-dimensionality reduction techniques (Cambria et al., 2012) have been employed on a knowledge base obtained by blending ConceptNet (Speer and Havasi, 2012), a directed graph representation of common-sense knowledge, with WordNet-Affect (WNA) (Strapparava and Valitutti, 2004), a linguistic resource for the lexical representation of affect. In this work, a novel cognitive model based on the combined use of principal component analysis (PCA) and artificial neural networks (ANNs) is exploited on the same knowledge base to further improve the way multi-word expressions are organised in a brain-like universe of natural language concepts. Results demonstrate noticeable enhancements in emotion recognition from natural language text with respect to previously adopted strategies and pave the way for future development of more biologically inspired approaches to the emulation of affective common-sense reasoning.
The rest of this paper is organised as follows: the next section introduces related works in the field of opinion mining; the following one illustrates how the affective common-sense knowledge base is constructed; next, a section describes the multi-dimensional scaling techniques adopted to perform reasoning on such a knowledge base; the following section presents the emotion categorisation model used for clustering affective knowledge; then, a section describes in detail the proposed cognitive architecture and how this can be exploited for brain-inspired opinion mining; finally, the last section offers some concluding remarks and future work recommendations.
Section snippets
Related work
Existing approaches to opinion mining can be grouped into three main categories, with few exceptions: keyword spotting, lexical affinity, and statistical methods. Keyword spotting is the most naı¨ve approach and probably also the most popular because of its accessibility and economy. Text is classified into affect categories based on the presence of fairly unambiguous affect words like ‘happy’, ‘sad’, ‘afraid’, and ‘bored’. Elliott’s Affective Reasoner (Elliott, 1992), for example, watches for
Building the affective common-sense knowledge base
The affective common-sense knowledge base developed within this research work is built upon ConceptNet, the graph representation of the Open Mind corpus, which structurally similar to WordNet (Fellbaum, 1998), but whose scope of contents is general world knowledge, in the same vein as Cyc (Lenat and Guha, 1989). Instead of insisting on formalising common-sense reasoning using mathematical logic (Mueller, 2006), ConceptNet uses a new approach: it represents data in the form of a semantic network
Multi-dimensional scaling for affect recognition
The best way to solve a problem is to already know a solution for it. But, if we have to face a problem we have never met before, we need to use our intuition. Intuition can be explained as the process of making analogies between the current problem and the ones solved in the past to find a suitable solution. Marvin Minsky attributes this property to the so called ‘difference-engines’ (Minsky, 1986). This particular kind of agents operates by recognising differences between the current state
Emotion categorisation model
In order to accordingly organise and interpret AffectiveSpace, an affective categorisation model is needed. The Hourglass of Emotions (Cambria et al., 2012), a model inspired by Plutchik’s studies on human emotions (Plutchik, 2001), was selected. It reinterprets Plutchik’s model by organising primary emotions around four independent but concomitant dimensions, whose different levels of activation make up the total emotional state of the mind. Such a reinterpretation is inspired by Minsky’s
Bio-inspired opinion mining engine
The proposed architecture extends a framework previously proposed by the authors (Mazzocco et al., 2012) and investigates if an emulation of the biological neural system, represented by two ANNs, could outperform the state-of-the-art k-medoids clustering approach (Cambria et al., 2011). Similarly to previous works (Cambria et al., 2012, Havasi et al., 2009, Cambria et al., 2010), the proposed architecture uses PCA to organise the space where concepts lie but, rather than using standard
Conclusions and future work
With the advent of Web 2.0, the extraction of opinions and sentiments from the huge amount of available unstructured information derived from blog, wikis, and social networks is a very arduous task. While existing approaches to opinion mining mainly work at a syntactic-level, computational techniques and tools were hereby employed to analyze text natural language at a semantic-level. In particular, we developed a bio-inspired opinion mining engine that, first, deconstructs natural language text
References (51)
- et al.
A novel demonstration of enhanced memory associated with emotional arousal
Consciousness and Cognition
(1995) - et al.
Circular-ELM for the reduced-reference assessment of perceived image quality
Neurocomputing
(2013) On the momentum term in gradient descent learning algorithms
Neural Networks
(1999)- et al.
Sentiment analysis in multiple languages: Feature selection for opinion classification in web forums
ACM Transactions on Information Systems
(2008) Solving the emotion paradox: Categorization and the experience of emotion
Personality and Social Psychology Review
(2006)Bodily changes in pain, hunger, fear and rage: An account of recent researches into the function of emotional excitement
(1915)- Cambria, E., Hussain, A., Durrani, T., Havasi, C., Eckl, C., & Munro, J. (2010). Sentic computing for patient centered...
- Cambria, E., Havasi, C., & Hussain, A. (2012). SenticNet 2: A semantic and affective resource for opinion mining and...
- Cambria, E., Olsher, D., & Kwok, K. (2012). Sentic activation: A two-level affective common sense reasoning framework....
- et al.
Sentic computing: Techniques, tools, and applications
(2012)