An interval type-2 fuzzy logic system to translate between emotion-related vocabularies

Abstract This paper describes a novel experiment that demonstratesthe feasiblity of a fuzzy logic (FL) representation of emotion-related words used to translate between different emotional vo-cabularies. Type-2 fuzzy sets were encoded using input fromweb-based surveys that prompted users with emotional wordsand asked them to enter an interval using a double slider. Thesimilarity of the encoded fuzzy sets was computed and it wasshown that a reliable mapping can be made between a largevocabulary of emotional words and a smaller vocabulary ofwords naming seven emotion categories. Though the map-ping results are comparable to Euclidian distance in the va-lence/activation/dominance space, the FL representation hasseveral benefits that are discussed.Index Terms: emotion, fuzzy logic, knowledge engineering. 1. Introduction As a matter of practice, many emotion studies begin with anenumeration of emotion categories that will be the object of in-quiry. For example, some common sets of emotions used in-clude: { negative, non-negative }, { angry, happy, sad, neu-tral }, and { angry, disgusted, fearful, happy, neutral, sad, sur-prised}. While enumeration of categories may work in somedomains, in many other cases imposing a categorization cancause over-simplification and lead to lack of interoperabilitywith other systems of categorization. Categorizing emotionclasses is especially exemplary of this problem due to the sub-jectivenatureofemotions. Also,thisproblemmanifestsitselfinmany practical issues of emotion research. It is difficult to gen-eralize the results of experiments that use different emotionalcategories. Forced choice in human labeling tasks may bias re-sults according to the categories presented. If an ”other” choiceis provided, it is not clear what to do with open-ended user in-put. It may be possible to have a precisely defined set of emo-tions in certain cases, such as extreme emotions associated withevolutionary behavior like fight-or-flight responses, but consid-ering human language as the focal point in this argument, thesubtlety of all emotions expressed by any person in any contextwarrants a more complete solution.One popular way to circumvent the problems of a fixedset of emotions is the dimensional approach, which representsemotions in terms of coordinates in an emotional space. Thetwo most common dimensions used are valence, representingan axis of pleasure/displeasure of the emotion, and activation,representinganaxisoftheenergyorstrengthoftheemotion[1].Sometimes a third dimension is also used. In this study we usedominance as a third dimension, which represents the scale be-tween aggressive and submissive. Another third dimension intheliteratureiscontrol, whichisascaleofhowmuchconsciouscontrol a person has over a given emotion. Representing emo-tions in a dimensional space is an important insight which isused in this paper’s research.Common ways of using the dimensional approach fallshort. Usually, an emotion’s location in such a dimensionalspaceisestimatedusingLikertscalesorsliders. Inbothofthesecases,usersmustpickasinglepointonagivenscale. Anyintra-user uncertainty is lost, though inter-subject uncertainty may beestimated from the ratings accumulated across users. Intra-useruncertainty may be estimated from repeated presentation of agiven stimulus, but this can be infeasible when broad coverageis needed.The approaches used in this paper, Interval Type-2 FuzzyLogicandtheComputingWithWordsparadigm,provideasolu-tion to the problems raised above. The methodology we presentuses double sliders to collect dimensional emotion ratings fromsubjects, as described in [2]. The span of the two sliders ona scale represents the region that the user wishes to select andis used to encode the user’s lexical knowledge into an intervaltype-2 fuzzy set. A similarity metric for interval type-2 fuzzysets [3] is used to map from one vocabulary to another, makingit possible to compare different sets of emotion categorizationsand deal with a wider range of user input, for example in thecase when users choose “other” on a forced choice task andthen provide open-ended input.The motivation for undertaking this work comes from try-ing to deal with an open-ended set of emotionally descriptivewords used as labels of the blog posts of the website livejour-nal.com. On this site, users can label their blog posts with a”mood” label. This makes the corpus a natural choice to studyemotions in the blogs. However, users may enter any word, sothere is a lot of noise in the labels and the set of labels has a“longtailed”Zipfdistribution. Insteadofdiscardingallthedataexceptthoselabeledwithafewbasicemotionaltags,wewantedto map a larger vocabulary, the words used in the blog labels,toasmaller, controlledvocabulary. Also, thesurvey, whichcol-lecteddatafrominternetusers,couldbeseenasawaytoextendthe idea of social computing from the blogs and mood tags to amore uniform semantic representation.Besides analyzing the blog data, some possible applica-tions of this research include: corpus annotation; informa-tion retrieval; translation and paraphrasing; social computing;emotion-aware applications where a small, fixed set of emo-tion labels is not practical; and fusion of different classifierswhere each classifier might use a different set of classes, as inthe case of multimodal analysis, where each modality is goodat recognizing a certain subset of emotions. In spoken languageprocessing, the speech itself can be thought of having differentmodalities, such as the acoustics and the symbolic string repre-sented as text. Also, the fact that fuzzy logic is an extension ofclassical logic allows it to provide for semantic representations