S33100107

556 views

Published on

International Journal of Engineering Research and Applications (IJERA) is an open access online peer reviewed international journal that publishes research and review articles in the fields of Computer Science, Neural Networks, Electrical Engineering, Software Engineering, Information Technology, Mechanical Engineering, Chemical Engineering, Plastic Engineering, Food Technology, Textile Engineering, Nano Technology & science, Power Electronics, Electronics & Communication Engineering, Computational mathematics, Image processing, Civil Engineering, Structural Engineering, Environmental Engineering, VLSI Testing & Low Power VLSI Design etc.

Published in: Technology, Education
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
556
On SlideShare
0
From Embeds
0
Number of Embeds
3
Actions
Shares
0
Downloads
0
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

S33100107

  1. 1. Amira F. El Gohary, Torky I. Sultan, Maha A. Hana, Mohamed M. El Dosoky / InternationalJournal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.100-107100 | P a g eA Computational Approach for Analyzing and DetectingEmotions in Arabic TextAmira F. El Gohary1, Torky I. Sultan1, Maha A. Hana1, Mohamed M. ElDosoky21( Information Systems Department, Faculty of Computers and Information, Helwan University)2(Biomedical Engineering Department, Faculty of Engineering, Helwan University)ABSTRACTThe field of Affective Computing (AC) expects tonarrow the communicative gap between thehighly emotional human and the emotionallychallenged computer by developing computationalsystems that recognize and respond to theaffective states of the user. Affect-sensitiveinterfaces are being developed in number ofdomains, including gaming, mental health, andlearning technologies. Emotions are part ofhuman life. Recently, interest has been growingamong researchers to find ways of detectingsubjective information used in blogs and otheronline social media. This paper concerned withthe automatic detection of emotions in Arabictext. This construction is based on a moderatesized Arabic emotion lexicon used to annotateArabic children stories for the six basic emotions:Joy, Fear, Sadness, Anger, Disgust, and Surprise.Our approach achieves 65% accuracy for emotiondetection in Arabic text.Keywords - Emotion recognition, affectivecomputing, human- machine interaction, emotionalArabic lexicon, text analysis.1. INTRODUCTIONHuman Computer Interaction (HCI) isdefined as “a discipline concerned with thedesign, evaluation and implementation ofinteractive computing systems for human use andwith the study of major phenomena surroundingthem”. It is an interdisciplinary field which arisesfrom the overlapped roots in computer science,software engineering, image processing, humanfactors, cognitive science, psychology. One of theprimary aims in human- computer interactionresearch is to develop an ability to recognizeaffective state of a user, thereby making a computerinterface more usable, enjoyable, and effective [1].Effective design of Affective computing systemsrelies on cross-disciplinary theories of emotion withthe practical engineering goal of developing affect-sensitive interfaces. Affect detection is a verychallenging problem because emotions areconstructs with fuzzy boundaries and withsubstantial individual difference variations inexpression and experience [2].Effective design of Affective computing systemsrelies on cross-disciplinary theories of emotion withthe practical engineering goal of developing affect-sensitive interfaces.Emotion detection researches have investigatedemotion detection in prosody, changes inphysiological state, facial expressions and text.There is a little of researches in emotion detectionfrom text in comparison to the other areas ofemotion detection [4].Text based emotion detection can be used inbusiness, education, psychology and any other fieldwhere there is paramount need to understand andinterpret emotions.Emotion detection in text aims to infer theunderlying emotions influencing the author/writer bystudying their input texts. This is based on the factthat if a person is happy, it influences them to usepositive words. And, if a person is sad, frustrated orangry, the kind of words they use can signify theirunderlying negative emotion. Emotion detection intext has a number of important applications. In thearea of business development, emotion detection canhelp marketers to develop strategies for customersatisfaction, new product development and servicedelivery. Psychologists can benefit from being ableto infer people‟s emotions based on the text that theywrite which they can use to predict their state ofmind. In the field of education, the ability ofcomputers to automatically track attitudes andfeelings with a degree of human intuition hascontributed to the development of Text-to-Speechsystems and Intelligent Tutoring Systems (ITS).Web communication can be facilitated throughvideo, voice recordings, images and text [4], [5], [6].In the absence of face-to-face contact to detect facialexpressions and intonations in voice, the alternativeoption is to infer emotions from text in onlineforums. Work in emotion detection can be classifiedinto: that which looks for specific emotion denotingwords, that which determines tendency of terms toco-occur with seed words whose emotions areknown, that which uses hand coded rules, and thatwhich uses machine learning techniques. Anemotion lexicon is a list of emotions and the wordsthat are indicative of each word, and it is prerequisiteto identify emotions in the text. As of now, high-
  2. 2. Amira F. El Gohary, Torky I. Sultan, Maha A. Hana, Mohamed M. El Dosoky / InternationalJournal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.100-107101 | P a g equality, high-coverage emotion lexicons do not existfor any language. Few lexicons for some languageare available such as the WordNet Affect lexicon(WAL) for six basic emotions, for English wordsand the General Inquirer (GI) which categorizeswords into a number of categories, includingpositive and negative semantic orientation [7].In this paper, a computational approach foranalyzing and tracking of emotions from Arabicchildren stories is proposed. Along with aconstruction of a moderate-sized Arabic lexicon toidentify the emotional expressions at word, sentenceand document level. The VSM model is used for theclassification of stories according to the six basicemotion categories. The proposed architecture istested by Arabic textual dataset to find the resultsand discuss the preliminary findings [8].The paper is organized as follows: Section 2 reviewsrelated work that covers emotions models used tocapture the affective states of a text, informationretrieval as a tool for emotion detection, andcomputational approaches for emotion detection.Section 3 describes the emotion architecture that hasbeen proposed and validated. Section 4 explains theexperimental setup that is used and discusses theresults. The paper is concluded by discussing thefindings, and proposing areas to be considered infuture research in Section 5.2. RELATED WORK2.1Computational approaches for emotiondetectionEmotions can be represented by twodifferent models. One popular approach involves theuse of a categorical representation, in whichemotions consists of labels such as boredom,frustration, and anger. An alternative approachemphasizes the importance of the fundamentaldimensions of valence and arousal in understandingemotional experience. Dimensional approaches havelong been studied by emotion theorists and theysuggest the existence of at least two fundamentaldimensions: valence (pleasure/displeasure) andarousal (activation/deactivation). Other researchershave found „dominance‟ a third dimension importantto represent emotional phenomena. A categoricalmodel is adopted for most emotion recognition intext and video, while dimensional representation issuitable in the case of speech. [9]Three approaches currently dominate the emotiondetection task are: keyword based, learning basedand hybrid based approach. These make use offeatures selected from syntactic (n-grams, pos tags,phrase patterns) and semantic (synonym sets) data todetect emotions.2.1.1Keyword Spotting approachThis approach depends on the presence ofkeywords and may involve pre-processing with aparser and emotion dictionary. It is easy toimplement, intuitive and straight forward since itinvolves identifying words to search for in text.2.1.2 Learning based approachThis approach uses a trained classifier tocategorize input text into emotion classes by usingkeywords as features. It is easier and faster to adaptto domain changes since it can quickly learn newfeatures from corpora by supplying a large trainingset to a machine learning algorithm for building aclassification model. However, acquiring largecorpora may not always be feasible. The majordrawback of this approach is that it leads to inexactboundaries between emotion classes and a lack ofcontext analysis.2.1.3 Hybrid based approachThis approach consists of a combination ofthe keyword based implementation and learningbased implementation. The main advantage of thisapproach is that it can yield higher accuracy resultsfrom training a combination of classifiers and addingknowledge-rich linguistic information fromdictionaries and thesauri. The advantage of this isthat it will offset the high cost involved in usinghuman indexers for information retrieval tasks andminimize complexities encountered whileintegrating different lexical resources. [10]2.2 Emotions in TextThe research on detecting emotionalcontent in text refers to written language andtranscriptions of oral communication. Early researchto link text and emotions aimed at understandinghow people express emotions through text, or howtext triggers different emotions, was conducted byOsgood.Osgood [11] used multidimensional scaling(MDS) to create visualizations of affective wordsbased on similarity ratings of the words provided tosubjects from different cultures. The words can bethought of as points in a multidimensional space,and the similarity ratings represent the distancesbetween these words. The emergent dimensionsfound by Osgood were “evaluation,” “potency,” and“activity.” Evaluation quantifies how a word refersto an event that is pleasant or unpleasant, potencyquantifies how a word is associated to an intensitylevel, and activity refers to whether a word is activeor passive.Lutz [12] has found similar dimensions butdifferences in the similarity matrices produced bypeople of different cultures. Samsonovich andAscoli [13] used English and French dictionaries togenerate “conceptual value maps,” a type of“cognitive map” similar to Osgood‟s, and found thesame set of underlying dimensions. Other researches[14], [15], [16] involve a lexical analysis of the textin order to identify words that are predictive of theaffective states of writers or speakers.
  3. 3. Amira F. El Gohary, Torky I. Sultan, Maha A. Hana, Mohamed M. El Dosoky / InternationalJournal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.100-107102 | P a g eSeveral of these approaches rely on the LinguisticInquiry and Word Count (LIWC) [17], a validatedcomputer tool that analyzes bodies of text usingdictionary-based categorization. LIWC-based affectdetection methods attempt to identify particularwords that are expected to reveal the affectivecontent in the text [18]. Other researchers have usedcorpora-based approaches that assume that peopleusing the same language would have similarconceptions for different discrete emotions and usedthesauri of emotional terms. For example, Wordnetis a lexical database of English terms and is widelyused in computational linguistics research [19].Strapparava and Valituti [20] extended Wordnetwith information on affective terms.The Affective Norm for English Words(ANEWs) (Bradley and Lang, 1999, Stevenson,Mikels and James, 2007) [21], [22] is one of severalprojects to develop sets of normative emotionalratings for collections of emotion elicitation objects,in this case English words. International AffectivePicture System (IAPS) (Lang, Greenwald, Bradleyand Hamm, 1993), a collection of photographs.These collections provide values for valence,arousal, and dominance for each item, averaged overa large number of subjects who rated the items usingthe Self-Assessment Manikin (SAM) introduced byLang and colleagues. Finally, affective norms forEnglish Text (ANET) [23] provide a set ofnormative emotional ratings for a large number ofbrief texts in the English language.There are also some text-based affectdetection systems that rely on a semantic analysis ofthe text. For example, Gill et al. (2008) [24]analyzed 200 blogs and reported that texts judged byhumans as expressing fear and joy were semanticallysimilar to emotional concept words (phobia, terrorfor fear and delight, and bliss for joy). They usedLatent Semantic Analysis (LSA) (Landauer,McNamara, Dennis and Kintsch, 2007) [25] and theHyperspace Analogue to Language (HAL) (Lundand Burgess, 1996) [26] to automatically computethe semantic similarity between the texts andemotion keywords. Although this method ofsemantically aligning text to emotional conceptwords showed some promise for fear and joy texts, itfailed for texts conveying six other emotions, suchas anger, disgust, and sadness. So, it is an openquestion whether semantic alignment of texts toemotional concept terms is a useful method foremotion detection.The most complex approach to textualaffect sensing involves systems that constructaffective models from large corpora of worldknowledge and applying these models to identify theaffective words in texts (Akkaya, Wiebe andMihalcea, 2009, Breck, Choi and Cardie, 2007, Liu,Lieberman and Selker, 2003, Pang and lee, 2008,Shaikh, Prendinger and Ishizuka, 2008) [27], [28],[29]. This approach is called sentiment analysis,opinion extraction, or subjectivity analysis because itfocuses on the valence of a textual sample ( positiveor negative, bad or good) rather than assigning thetext to a particular emotion category (angry and sad)[30]. Supervised and unsupervised machine learningtechniques have been used to automaticallyrecognize emotion in text. Supervised techniqueshave the disadvantage that large annotated datasetsare required for training. The emotionalinterpretations of a text can be highly subjective, somore than one annotator is needed, and this makesthe process of the annotation very time consumingand expensive. Thus, unsupervised methods arepreferred in the field of Natural LanguageProcessing (NLP) and emotions. Strapparava andMihalcea compared a supervised (Naïve Bayes) andfour unsupervised techniques (combinations of LSAwith Wordnet Affect) for recognizing six basicemotions (Strapparava and Mihalcea, 2008).D‟Mello and colleagues (DMello, Craig,Witherspoon, McDaniel and Graesser, 2010) [31]used LSA but for detecting utterance types andaffect in students‟ dialogue within an IntelligentTutoring System. As it is required in a categoricalmodel of emotions, D‟Mello proposed a set ofcategories for describing the affect states in studentsystem dialogue. Kort (2001) combined the twoemotion models, placing categories in a valencearousal plane. To date, most affective computingresearchers have utilized and evaluated supervisedmethods based on the categorical emotion model.Most of the works mentioned above carried out forEnglish.2.3 Information retrieval as a tool for emotiondetectionInformation retrieval is the acquiring ofspecific information about a topic and is usuallybased on a query [32]. Text classification consists ofcategorizing text at the document, sentence or tokenlevel (document classification, opinion recognition,word sense disambiguation) [33]. Classification is aninformation retrieval task used for analyzing thecontent of unstructured data expressed in naturallanguage on the web (email, scientific documentsand government reports) [34].Text classification is currently the mainmethod for emotion detection in text [35].Nevertheless, the highly subjective nature ofemotion, this method faces many challenges. Themajor challenge identified is the inclusion ofsubjectivity detection mechanisms. Kao et al. [36]have specifically identified three issues associatedwith keyword spotting techniques. They are:“ambiguity in keyword definitions”, “incapability ofrecognizing sentences without keywords” and “lackof linguistic information” [37].
  4. 4. Amira F. El Gohary, Torky I. Sultan, Maha A. Hana, Mohamed M. El Dosoky / InternationalJournal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.100-107103 | P a g e3. ARCHITECTURE FOR EMOTIONDETECTIONThe focus of this study is on sentenceemotion based on emotion words using a datasetcomposed of children‟s stories. The dataset consistsof 100 documents containing 2,514 sentences.Sentences are the basic units for emotionalexpression, so they are used as basic level foremotion annotation. Each document is decomposedinto non-overlapping and sequential sentences.Stories consist of words that are written by with theintension to evoke emotions and to attract thechildren‟s attention are suitable for this experimentbecause stories tended to evoke emotions that attractthe reader‟s attention.The problem of classifying the emotionalorientation of sentences in the domain of children‟sstories can be decomposed into two subproblems: Annotating a datasetA dataset of 65 documents is annotated onsentence level with eight emotion categories, thebasic six emotions ( ‫سعبدة‬,‫حسن‬,‫خوف‬,‫غضب‬,‫اشمئساز‬,‫مفبجأة‬ ), in addition to the (neural category) incase of the sentence does not contain any emotionwords, and the (mixed category) when the sentencecontains emotional words belongs to more than oneemotion category. The documents have beenannotated by two annotators independently tomeasure agreement on the annotation of this dataset.The annotators were asked to select the appropriateemotions for each sentence based on the presence ofwords with emotional content, and also the overallfeeling invoked by the sentence. The agreementbetween two annotators is calculated by findingobserved and expected agreement. Observedagreement (Ao) calculates how much two annotatorsagreed on the individual annotations that eachannotator made. Expected agreement (Ae) calculateshow much the annotators are expected to agree ifthey each randomly assigned emotions to thesentences. Kappa value is calculated with thefollowing equation: K = ( Ao - Ae ) / (1 - Ae)A Kappa value of 1.0 is total agreement and 0.0 iscomplete random labeling.For our dataset the Observed agreement(Ao) is (0.16), and the Expected agreement (Ae) is(0.55). The Kappa value is (0.46) that suggestingmoderate agreement. This result is due to the natureof emotion, that is inherently ambiguous both interms of the emotion classes and the naturallanguage words that representing them. Manyinstances a sentence may be found to exhibit morethan one emotion (the word “‫”صرخ‬ can relate tothree categories ( "‫غضب‬ ،‫حسن‬ ،‫سعبده‬" ), more than oneemotion can be assigned to one sentence, that theterms in the anger category tend to share disgustterms and also fear and sadness terms, a sentencealso could not be attributed to any determinedcategories. Cohen‟s kappa is used to measure theagreement in classifying items into determinedcategories.Based on the annotated dataset, the emotionlexicon for the six basic emotions is constructed. Tothe best of our knowledge, at present, there are nolarge Arabic emotion lexicons. The developedArabic lexicon is an affective lexical repository ofwords referring to emotional states relevant to thesix basic emotion categories extracted from theannotated dataset and with direct translation ofEnglish emotional lexicons. This lexicon containsemotional words for direct and indirect affectivewords. The direct emotional words refer directlyemotional states. While indirect emotional wordshave indirect reference that depends on the context.Testing dataset is selected from this datasetrandomly, includes 35 documents, this allows tomeasure how well our process marks the sentencesfrom which we have obtained our Arabic lexicon. Automatic Emotion DetectionIn the first step in the procedure the textgoes through preprocessing steps which are:stopwords removal and word stemming. These stepshelp the significant linguistic components of text tobe focused and considered by removing unimportantdata.Starting with the motion words lexicon,which contain six lists of affective words with bothdirect and indirect emotion words for each category,the Vector Space Model (VSM) is used to computethe similarities between sentences and the basic sixemotion lexicon. With the emotional word intensityvalues and term frequency, the emotion weightmatrix can be measured. An emotional word isrepresented as a vector to record its intensity of thesix basic emotional classes. The matrix l, whoseentire l (i, j) are the emotion weight of the term i insentence j. Terms are encoded as vectors in a matrix,whose components are co-occurrence frequencies ofwords in the sentences for each documentscorresponding to the six emotional categories. Thematrix element means the importance of the term inrepresenting the emotional meaning of the sentence.Sentences in our categorical approach are convertedto a vector whose components are co-occurrencefrequencies of emotional words in sentences, forexample a sentence vector is represented as vectorvalues corresponding to the basic six emotions:s= (0.0, 0.4, 0.7, 0.0, 0.0, 0,0)Both the sentences and emotional categories arerepresented in a common vector space as in fig.1.
  5. 5. Amira F. El Gohary, Torky I. Sultan, Maha A. Hana, Mohamed M. El Dosoky / InternationalJournal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.100-107104 | P a g eFig.1 sentences and emotion categories arerepresented as vectors in the vector space modelDistances between the sentences andemotional categories can then be measured by usingthe cosine angle between the input sentence and anemotional vector as a similarity measure to labeleach sentence with the closest category using (1). Ifthe cosine similarity does not exceed a threshold, theinput sentence is labeled as “neutral”, the absence ofemotion. Otherwise it is labeled with one emotionassociated with the closest emotional vector havingthe highest similarity value. A predeterminedthreshold is (0.6) is used for the purpose ofvalidating a strong emotional relation between twovectors. The values produced by this technique areto be used for the classification task that convertseach sentence into predefined categories. Theclassification consists of finding the closest emotioncategory to the sentence. One class with themaximum score is selected as the final emotionclass.(1)After that, each document can berepresented by summing up the normalized vectorsof all the sentences contained in it. The similaritybetween each document and the emotionalcategories is computed also using the cosinesimilarity and assign a document a category with thehighest value.Fig.3 the results of measuring the similaritybetween the normalized documents‟ vectors in ourdataset with the emotional categories to determinewhich category each document belongs to. Eachdocument has different values with respect to eachemotion from the six basic emotions; a document islabeled with the emotion with the highest valuewhich indicate the closest vector to that document.Fig.2. similarity values for various documentsAfter that, K-mean method is used in theclassification of the documents, which proposeclusters for the six emotional categories. In the firststep random vectors have been assigned arbitrary to10 classes, then compute the similarity betweenthese classes and the normalized vectors of thedocuments. The documents will be clustered indifferent classes. For each class a new class valuewill be computed getting the average of the vectorsfor all the documents containing in it. Repeat theabove steps with the new vector values for thecategories until no change occur in the values of theclasses vectors. Each category contains some of thedocuments that are related to this category. This willfacilitate the information retrieval processes thatmaking queries for specific documents easier.Fig.2 shows the results of applying the K-meansmethod on the test dataset using the three emotionalwords (all, direct, and indirect) in this case themajority class is “‫,”السعبدة‬ in the three emotionalwords. This is due to the sentences in the childrenstories tending to have more happy emotional terms.024681012--Fig.2 no of documents within different emotionalcategoriesThe goal of affect detection is to predict a singleemotional label to a given input sentence. In thisdataset of stories the majority class is happiness.
  6. 6. Amira F. El Gohary, Torky I. Sultan, Maha A. Hana, Mohamed M. El Dosoky / InternationalJournal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.100-107105 | P a g eClassification accuracy is usually measured in termsof precision, recall, and F-measure.Precision =FPTPTP, Recall =FNTPTP(2)Precision is the number of correctly labeledsentences retrieved by the algorithm divided by allthe sentences retrieved by the algorithm.Recall is the number of correctly labeled sentencesretrieved by the given algorithm divided by thesentences annotated as correct. After precision andrecall are calculated, the values are used to calculatethe f-score, the harmonic mean of precision andrecall that functions as a weighted average equation.f - score = 2 *callecisioncallecisionRePrRe*Pr(3)Table 1 Overall average results for the three datasetsusing precicion, recall, and f-scorsEmotions Precision RecallF –measure‫السعبدة‬ 0.794 0.771 0.782‫الحزن‬ 0.675 0.612 0.642‫الخوف‬ 0.714 0.585 0.643‫الغضب‬ 0.6 0.6 0.6‫اإلشمئزاز‬ 0.4 0.333 0.363‫المفبجأة‬ 0.867 0.8 0.828Table1 shows the results of measuring the accuracyof our approach to automatic emotion detection withthe test dataset. Our method obtained average 65% f-measure for recognizing the six basic emotions forArabic emotion sentences. This showed promisingresults compared with the existing classificationmethods used with other languages.From the above results, it has been found that theemotions computed by our method are in agreementwith the emotion annotated by annotators.4. CONCLUSIONEmotions recognition on text has wideapplications. This study investigates a computationalapproach for analyzing and tracking emotions inArabic children stories based on emotion words.This implies a development of Arabic lexicon usedfor determining emotional words in the input text.Arabic differs from European languagesmorphologically, syntactically, and semantically.The Arabic language is difficult to deal with due toits orthographic variations and its complexmorphological structure. The Vector Space Model(VSM) is used to compute the similarities betweensentences and the basic six emotion lexicon, theinput sentence is labeled with the closest category.K-mean method is used in the classification of thedocuments, which propose clusters for the sixemotional categories. Each category contains someof the documents that are related to this category.The results show that the majority class is “‫.”السعبدة‬Detecting emotions in Arabic texts is useful for anumber of purposes including: identifying blogs thatexpress specific emotions towards the topic ofinterest, identifying what emotions a newspaperheadline evoke, and create automatic dialoguesystems that respond automatically andappropriately to different emotional states of theuser. The resulting document level emotion taggercan be used in an emotion based informationretrieval system. The experiments show thatemotional words and sentences play important rolefor emotion recognition, but more linguisticexpressions such as negative word, conjunctions,and punctuations should be considered for moreaccurate recognition. That the affective meaning isnot simply expressed by the lexicon used as themodel assumes, it is also an effect of the linguisticstructure.5. FUTURE WORKResearch efforts will be directed towardsapplying machine learning techniques that work withunlabeled data, rather than annotating data which istime consuming and also susceptible to error.Integrate the in-depth semantic analysis of the text,the affective lexicon used need to be enhanced that itshould include the entire emotional adjective, nouns,verbs, and adverbs, emoticons, and abbreviations tohandle abbreviated language.The emotions of a text can be affected by manyfactors: emotion words, negative words,conjunctions, modifiers, punctuations, contexts inthis study not all of these factors are included, infuture research it is useful to include the impact ofthe function words in changing the emotional valuesof the text.REFERENCES[1] W. Peter, M. Roger, F. Petra, C. Marc andW. Yorick, Emotion in Human-AgentInterfaces (January 2006).[2] S. Turkle, The Second Self: Computers andthe Human Spirit (Simon & Schuster,1984).[3] S. Mingli , Y. Mingyu, L.Na, Ch. Chun, Arobust multimodal approach for emotionrecognition (College of Computer Science,Zhejiang University, China, 2008).[4] P. Ekman. “Are there basic emotions?”,Psychological Review, 99, pp. 550-553.[5] N. Frijda, “Emotion, Cognitive Structure,and Action Tendency,” Cognition andEmotion, vol. 1, pp. 115-143, 1987.[6] O. Zeljko , G. Nestor , L. Miguel ,F.Inmaculada , C.Idoia, An Ontology forDescription of Emotional Cues, (ComputerScience Faculty; University of the BasqueCountry, 2005).
  7. 7. Amira F. El Gohary, Torky I. Sultan, Maha A. Hana, Mohamed M. El Dosoky / InternationalJournal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.100-107106 | P a g e[7] B. Nadia, C. Paul, C. Anna, J. Charlene,and K. Woong. On posture as a modalityfor expressing and recognizing emotions.UCLIC, University College London,Remax House, Alfred Place, LondonWC1E 7DP, UK.[8] C. Elizabeth, N. Shami, P. Christian, Let’sget emotional: Emotion research in humancomputer interaction, (California, USA,April 2007).[9] S. Mingli , Y. Mingyu, L.Na, Ch. Chun, Arobust multimodal approach for emotionrecognition (College of Computer Science,Zhejiang University, China, 2008).[10] Shiv Naresh Shivhare, Saritha Khethawat,“Emotion Detection from Text”, CS & IT05, 2012, pp.371-377.[11] C.E. Osgood,W.H. May, and M.S. Miron,Cross-Cultural Universals of AffectiveMeaning (Univ. of Illinois Press, 1975).[12] C. Lutz and G. White, “The Anthropologyof Emotions,” Ann. Rev. Anthropology, vol.15, pp. 405-436, 1986.[13] A.V. Samsonovich and G.A. Ascoli,Cognitive Map Dimensions of the HumanValue System Extracted from NaturalLanguage, Proc. AGI Workshop Advancesin Artificial General Intelligence: Concepts,Architectures and Algorithms, pp. 111-124,2006.[14] M.A. Cohn, M.R. Mehl, and J.W.Pennebaker, “Linguistic Markers ofPsychological Change SurroundingSeptember 11, 2001”, PsychologicalScience, vol. 15, pp. 687-693, Oct. 2004.[15] J. Pennebaker, M. Mehl, and K.Niederhoffer, “Psychological Aspects ofNatural Language Use: Our Words, OurSelves,” Ann. Rev. Psychology, vol. 54, pp.547-577, 2003.[16] J. Kahn, R. Tobin, A. Massey, and J.Anderson, “Measuring EmotionalExpression with the Linguistic Inquiry andWord Count,” Am. J. Psychology, vol. 120,pp. 263-286, Summer 2007.[17] J. Pennebaker, M. Francis, and R. Booth,Linguistic Inquiry and Word Count(LIWC): A Computerized Text AnalysisProgram. Erlbaum Publishers, 2001.[18] J. Hancock, C. Landrigan, and C. Silver,Expressing Emotion inText-BasedCommunication, Proc. SIGCHI, 2007.[19] G. Miller, R. Beckwith, C. Fellbaum, D.Gross, and K. Miller, “Introduction toWordnet: An On-Line Lexical Database”, J.Lexicography, vol. 3, pp. 235-244, 1990.[20] C. Strapparava and A. Valitutti, “WordNet-Affect: An Affective Extension ofWordNet,” Proc. Int’l Conf. LanguageResources and Evaluation, vol. 4, pp. 1083-1086, 2004.[21] M.M. Bradley and P.J. Lang, “AffectiveNorms for English Words (ANEW):Instruction Manual and Affective Ratings,”Technical Report C1, The Center forResearch in Psychophysiology, Univ. ofFlorida, 1999.[22] R.A. Stevenson, J.A. Mikels, and T.W.James, “Characterization of the AffectiveNorms for English Words by DiscreteEmotional Categories,” Behavior ResearchMethods, vol. 39, pp. 1020-1024, 2007[23] .M.M. Bradley and P.J. Lang, “AffectiveNorms for English Text (ANET): AffectiveRatings of Text and InstructionManual,”Technical Report No. D-1, Univ.of Florida, 2007.[24] Gill, R. French, D. Gergle, and J.Oberlander, “Identifying EmotionalCharacteristics from Short Blog Texts,”Proc. 30th Ann. Conf. Cognitive ScienceSoc., B.C.Love,K.McRae,and V.M.Sloutsky, eds., pp. 2237-2242, 2008.[25] T. Landauer, D. McNamara, S. Dennis, andW. Kintsch, Handbook of Latent SemanticAnalysis. Erlbaum, 2007.[26] K. Lund and C. Burgess, “Producing High-Dimensional Semantic Spaces from LexicalCo-Occurrence”, Behavior ResearchMethods Instruments and Computers, vol.28, pp. 203-208, 1996.[27] E. Breck, Y. Choi, and C. Cardie,“Identifying Expressions of Opinion inContext,” Proc. 20th Int’l Joint Conf.Artificial Intelligence, 2007.[28] H. Liu, H. Lieberman, and S. Selker, “AModel of Textual Affect Sensing UsingReal-World Knowledge,” Proc. Eighth Int’lConf. Intelligent User Interfaces, pp. 125-132, 2003.[29] M. Shaikh, H. Prendinger, and M. Ishizuka,“Sentiment Assessment of Text byAnalyzing Linguistic Features andContextual Valence Assignment”, AppliedArtificial Intelligence, vol. 22, pp. 558-601,2008.[30] B. Pang and L. Lee, “Opinion Mining andSentiment Analysis,” Foundations andTrends in Information Retrieval, vol. 2, pp.1-135, 2008.[31] S. D‟Mello and A. Graesser, “MultimodalSemi- Automated Affect Detection fromConversational Cues, Gross BodyLanguage, and Facial Features,” UserModeling and User-Adapted Interaction,vol. 10, pp. 147-187, 2010.[32] H. H. Binali, V. Potdar, and C. Wu, "AState Of The Art Opinion Mining And Its
  8. 8. Amira F. El Gohary, Torky I. Sultan, Maha A. Hana, Mohamed M. El Dosoky / InternationalJournal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.100-107107 | P a g eApplication Domains," in IEEE ICIT09,Gippsland, Victoria, Australia, 2009.[33] H. Cunningham, D. Maynard, K.Bontcheva, V. Tablan, C. Ursu, M.Dimitrov, M. Dowman, N. Aswani, I.Roberts, Y. Li, A. Sha, and A. Funk,"Developing Language ProcessingComponents with GATE Version 5 (a UserGuide)," 2009.[34] F. Sebastiani, "Machine Learning inAutomated Text Categorization," ACMComputing Surveys, vol. 34, No.1, pp. 1-47,2002.[35] J. Read, "Recognising Affect in Text usingPointwise-Mutual Information ": Master ofScience, University of Sussex, 2004.[36] E. C.-C. Kao, L. Chun-Chieh, Y. Ting-Hao,H. Chang-Tai, and S. Von-Wun, "TowardsText-based Emotion Detection," inInternational Conference on InformationManagement and Engineering, 2009, pp.70-74.[37] Pang Bo, and Lillian Lee. 2008.Foundations and Trends in InformationRetrieval 2, vol. 1-2, pp. 1–135.

×