Search

V teoretičnem delu magistrskega dela so predstavljena najprej čustva na splošno, nato pa čustvo jeze, pri čemer so podrobneje predstavljeni prepoznavanje, izražanje in uravnavanje jeze, specifične strategije za uravnavanje jeze ter jeza v šolskem prostoru. Empirični del je zasnovan kvalitativno in kvantitativno. Vzorec zajema šestindvajset učencev 3. in 4. razreda osnovne šole, s katerimi so bili opravljeni polstrukturirani intervjuji pred in po treningu uravnavanja čustev in izvedeno strukturirano opazovanje izražanja čustev učencev med treningom. Trening so sestavljale aktivnosti: reševanje delovnih listov o prepoznavanju, izražanju in uravnavanju čustev s poudarkom na jezi, učenje novih strategij uravnavanja čustev s poudarkom na jezi, pogovorov o tem in izvajanje različnih tehnik sproščanja. Na začetku in ob koncu vsakodnevnih aktivnosti treninga je vsak učenec tudi prepoznal in komentiral svoje trenutno čustveno doživljanje. V raziskavi me je zanimalo, na katere načine so učenci prepoznavali in izražali jezo pred celotnim treningom in med vsakodnevnimi aktivnostmi treninga, na katere načine so učenci uravnavali jezo pred in po celotnem treningu in kako uspešni so bili učenci pri uravnavanju jeze med vsakodnevnimi aktivnostmi treninga. Poleg tega me je zanimalo tudi, katere tehnike sproščanja so prepoznali kot najbolj učinkovite pri uravnavanju jeze. Rezultati pred treningom so pokazali, da učenci znajo prepoznavati čustvo jeze in se zavedajo vzrokov zanjo, pri čemer kot glavni vzrok jeze navajajo nasilje vseh vrst. Pri doživljanju jeze se vsi učenci počutijo neprijetno in si želijo svoje počutje čim prej spremeniti, pri skoraj vseh učencih pa se ob jezi pojavljajo agresivni vzgibi. Pred treningom je bilo največkrat prepoznano čustvo veselja, jezo na začetku aktivnosti pa so učenci med samo aktivnostjo znali uravnati in so po koncu aktivnosti doživljali druga čustva. Večina učencev je bila pred našim treningom navajena izražati jezo na neprimeren način z agresivnim vedenjem. Učenci so poročali, da svojo jezo večinoma uravnavajo s preusmeritvijo pozornosti, umikom iz situacije ali pogovorom. Rezultati intervjujev z učenci po treningu so pokazali, da se učenci v situacijah, ki jih razjezijo, v večini najprej ustavijo in razmislijo, kako odreagirati, uporabljajo dihalne vaje, rešujejo konflikte s pogovorom, prakticirajo čuječnost in strategijo štetja. Rezultati opazovanja med treningom so pokazali, da so učenci svojo jezo dokaj uspešno uravnavali, nekateri pa so med izvajanjem aktivnosti potrebovali nekaj opozoril. Vzroki za opozorila in pogostost opozoril so se iz tedna v teden spreminjali. Med najuporabnejše tehnike sproščanja so učenci uvrstili dihalne vaje, čuječnost, jogo, masažo, gibalne aktivnosti in vizualizacijo. In the theoretical part of the master thesis there are firstly presented emotions in general, then emotion of anger with the focus on recognition, expression and regulation of anger, specific strategies of anger regulation and emotion of anger in school environment. Empirical part includes qualitative and quantitative research. Twenty-six pupils of third and fourth grade took part in the research, in which they took part in a half-structured interview before and after our training of emotion regulation and structured observation of their emotional expressions during our training. Training consisted of activities: solving worksheets about recognition, expression, and regulation of emotions, especially anger learning new strategies of emotion regulation, especially anger conversations about it and trying various relaxation techniques. At the beginning and at the end of everyday activity of the training each student recognised and commented on his current emotional state. In my research I wanted to figure out the ways in which the students will recognise and express anger before and during our training, in which ways will they regulate their anger before and after our training and how successful will they be at regulating anger during everyday activities in training. Furthermore, I wanted to find out which relaxation techniques will students recognize as the most efficient for regulating anger. Results gathered before training show that students can recognise emotion of anger and reasons for it with the main reason of their anger being violence of all kinds. When they are dealing with anger they all feel uncomfortable and want their emotions to improve as soon as possible. In almost all students anger is accompanied with the violent impulses. Before our training emotion of happiness was the most frequently reported, while the emotion of anger when reported at the beginning of the activity was regulated during the activity and students felt other, more pleasant emotions at the end of the activity. Students reported they mostly regulate their anger by redirecting their attention, moving out of the situation or by conversation. Results of the interviews after training showed that when dealing with anger most students use strategy where they first stop and think about their reaction, they use breathing exercises, they are solving conflicts by conversation, practise mindfullness and by counting strategies. Results of observation during our training showed that students were quite successful in regulating anger, some of the students needed a few warnings during the activities. Reasons and frequency of the warnings were changing throughout the training. Students believe that most efficient relaxation techniques are breathing exercises, mindfullness, yoga, massage, activities that include moving and visualization.

Pred kratkim so v študiji ugotovili, da lahko z uporabo nekaterih novih značilk občutno izboljšajo modele za klasifikacijo in regresijo glasbenih besedil glede na valenco in aktivnost čustev. Na podlagi te študije smo implementirali svoj sistem za tovrstno avtomatsko analizo glasbenih besedil. Delali smo z glasbenimi besedili v angleškem jeziku. Najprej smo s spleta pridobili vsa glasbena besedila. Ta besedila smo z našim procesorjem besedil preoblikovali v primerno obliko za ekstrakcijo značilk. Pripravili smo vrsto funkcij za pridobivanje značilk, s katerimi smo nato iz besedil izluščili različne tipe značilk. Nato smo značilke, s pomočjo algoritmov za selekcijo značilk, razvrstili po pomembnosti in izbrali le najpomembnejše. Z naključnim iskanjem parametrov smo optimizirali še parametre učnih algoritmov, ki smo jih učili na izbranih značilkah. Za učenje klasifikatorjev in regresorjev smo uporabili metodo podpornih vektorjev in Gradient Boosting. Uspešnost naših modelov smo na koncu ocenili s stratificiranim 10-kratnim prečnim preverjanjem. V delu predstavimo metode, ki smo jih uporabili za izgradnjo sistema, končno rešitev in rezultate, ki jih naš sistem dosega v primerjavi s sistemom iz prej omenjene študije. In a recent study on Lyrics Music Emotion Recognition a new set of features was proposed. These new features were proved to increase accuracy of existing models for classification and regression based on valence and arousal of emotion in lyrics. Based on the findings of this study we have implemented our own system for automatically acquiring, analyzing and classifying lyrics. We only dealt with lyrics in English. First we acquired the lyrics from the web. Then we prepared the lyrics for feature extraction, using the preprocessor we implemented. A variety of functions were implemented for feature extraction, which were then used to extract features from the preprocessed lyrics. Using feature selection algorithms we ranked the features and selected only the best. Using randomized hyper-parameter optimization we optimized the parameters of learning methods for our models. For classification and regression we used two learning algorithms, Support Vector Machine and Gradient Boosting. In the end we evaluated our models with stratified 10-fold cross validation. In this work we present the methods that we used to build our system, final solution and the results that we achieved in comparison to the study we based our system on.

V tem magistrskem delu smo se ukvarjali s študijem postopkov za razpoznavanje človeških čustev na osnovi digitalnih posnetkov. V praktičnem delu te magistrske naloge smo razvili izviren računalniški razpoznavalni sistem, ki temelji na teksturnih značilkah. Algoritem v prvem koraku poišče grob položaj obraza na vhodni sliki, zatem pa v dobljeni regiji računamo ujemanje obraza z modelom AAM (Active Appearance Models). Nato rotiramo obraz za izračunan kot v pravilen frontalni položaj. V zadnji fazi algoritma s pomočjo koordinat iz modela AAM natančno izrežemo obraz iz slike, zatem pa obraz aproksimiramo z elipso, s čimer odstranimo odvečne informacije z slike. Obrezan obraz na koncu popišemo s teksturno značilko HOG (Histogram of Oriented Gradients). Vmesni rezultat je histogram, ki ga posredujemo v stroje SVM (Support Vector Machines) za klasifikacijo, pri čemer za vsako od šestih osnovnih emocij naučimo lasten SVM. Razvili smo dve varianti algoritma, in sicer algoritem na osnovi statičnih 2D slik in algoritem na osnovi slik razlik. Prvi algoritem uporablja za razpoznavanje čustev le trenutno sliko opazovane osebe, medtem ko drugi algoritem detektira spremembe obraza pri izražanju čustev. Algoritem na osnovi statičnih 2D slik smo nadgradili z algoritmom Adaboost, algoritem na osnovi slik razlik pa smo razširili s tremi variantami, kjer pri gradnji histograma s pomočjo interpolacije vnesemo še vmesne korake pri spreminjanju obraza iz nevtralnega v obraz z izkazanim čustvom. Razvite algoritme smo testirali na dveh javno dostopnih testnih podatkovnih bazah: bazi MMI Facial Expression Database (MMI) in bazi Cohn-Kanade. Z algoritmom na osnovi statičnih 2D slik smo na bazi MMI dosegli najvišjo uspešnost 76,31 %, na bazi CK pa 91,49 %. Z algoritmom na osnovi slik razlik pa smo na bazi MMI dosegli 74,63 % uspešnost, na bazi CK pa se je uspešnost prepoznavanja čustev povzpela kar na 95,64 %. In our work, we have been studying the procedures of human emotion recognition based on digital images. We have developed an original computer recognitional sistem, which is based on feature descriptor, in the practical part of our work. The first step is the algorithm defining the approximate position of the face in the input image, which is followed by calculations of matching of the face with the AAM model (Active Appearance Model) in the region. Then we rotate the face according to the calculated angle to its frontal position. The last step of the algorithm is cutting the face out of the image, using coordinates, and approximating it with an ellipse in order to remove the abundant information from the image, which is followed by using the feature descriptor HOG (Histogram of Oriented Gradients). The intermediate result is a histogram, which needs to be classified by SVM machines (Support Vector Machines), whereas one machine is taught one particular emotion. We have developed two versions of the algorithm. The first one is based on static 2D images, and the other is based on images of differences. The first one uses a current image of the observed person, whereas the second one detects changes of the face when expressing emotions. We have upgraded the algorithm based on static 2D images with algorithm Adaboost, and expanded the one based on images of differences with three versions, which include interpolating additional intermediate changes of the face from neutral to the one with the revealed emotion, to construction of the histograms. We have tested the algorithms on two public test databases: MMI Facial Expression Database (MMI) and Cohn-Kanade database. The 2D static images based algorithm's highest score of effectiveness was 76.31% on MMI database and 91.49% on CK database. The images of differences based algorithm's highest score of effectiveness was 74.63% on MMI and 95.64% on CK.

Deep learning has recently revolutionised many fields of natural language processing but has not yet been applied to emotion recognition. Most recent studies of emotion recognition on tweets used simple classifiers on a combination of bag-of-words and human-engineered features. Hence, we worked on improving emotion-recognition algorithms using neural networks. To this end, we created three large emotion-labelled data sets corresponding to Ekman's, Plutchik's, and POMS's emotions by exploiting Twitter's popular self-annotation mechanism — hashtags. We compared the performance of bag-of-words and latent semantic indexing models with the performance of neural networks. We trained several word- and character-based, recurrent and convolutional neural networks. Further, we investigated the transferability of final hidden state representations of neural networks: how appropriate is the representation trained on one classification for recognising another one? Finally, we developed a single model for recognising all three emotion classifications from a shared representation. We show that neural networks can surpass traditional text classification approaches for emotion recognition. Recurrent neural network working directly on characters without any text preprocessing in a completely end-to-end fashion was the most successful architecture. Although models trained on single data sets have revealed poor transferability, we improved the generality of final hidden state representation in the unison model. When training the unison model, the standard training heuristic yielded unbalanced performance, due to the vast difference in data set sizes. However, the newly proposed training strategy produced a unison model with performance comparable to that of single models. Čeprav prodira globoko učenje na vsa področja procesiranja naravnega jezika, do zdaj še ni bilo uporabljeno za prepoznavanje čustev. Večina dosedanjih študij prepoznavanja čustev na tvitih uporablja preproste klasifikatorje na značilkah, ki pripadajo modelu vreče besed ali pa jih raziskovalci konstruirajo ročno. Glavna tema disertacije je izboljšava modelov za prepoznavanje čustev v tvitih z uporabo nevronskih mrež. V ta namen najprej ustvarimo tri velike podatkovne množice, sestavljene iz učnih primerov, ki so označeni glede na to, katero čustvo po Ekmanovi, Plutchikovi ali POMS-ovi kategorizaciji izražajo. Čustvene oznake pridelamo avtomatsko z uporabo Twitterjevega mehanizma za samooznačevanje vsebine, s t. i. tematskimi oznakami (angl. hashtags). Nato primerjamo natančnost klasifikatorjev z uporabo modelov vreče besed in latentnega semantičnega indeksiranja z natančnostjo nevronskih mrež, tako rekurenčnih kot konvolucijskih, ki na vhodu sprejmejo besede ali znake. Nadalje smo raziskovali prenosljivost reprezentacij končnih skritih stanj modelov nevronskih mrež, natančneje, ali je reprezentacija, naučena pri treniranju modela za neko klasifikacijo čustev, lahko koristna za napovedovanje druge klasifikacije. Zaključimo z učenjem skupnega modela, ki je sposoben prepoznavati čustva vseh treh omenjenih klasifikacij, pri tem pa je omejen na uporabo skupne reprezentacije. Eksperimentalno pokažemo, da so nevronske mreže natančnejše od klasičnih pristopov k prepoznavanju čustev. Kot najnatančnejše se izkažejo rekurenčne mreže, ki na vhodu sprejemajo znake in tako predstavljajo celosten pristop k učenju (angl. end-to-end learning). Čeprav je prenosljivost reprezentacij modelov, ki so trenirani na eni podatkovni množici, precej slaba, se ta drastično izboljša pri skupnem modelu. Pri učenju skupnega modela z znanimi metodami opazimo, da je natančnost zelo neuravnotežena glede na podatkovne množice, predvsem zaradi velike razlike v številu učnih primerov znotraj posamezne množice. Zato zasnujemo novo strategijo treniranja takšnih skupnih modelov, s katero naučimo model, katerega natančnost je uravnotežena čez vse tri podatkovne množice.