This study investigated the automatic recognition of emotion in English
and Arabic text. We perform experiments with a new method of classi-
fication for recognising emotions using the Prediction by Partial Matching
(PPM) character-based text compression scheme. These experiments involve
both document level classification (whether a text of document is emotional
or not) and also fine-grained classification such as recognising Ekman's six
basic emotions (Anger, Disgust, Fear, Happiness, Sadness, Surprise). Experimental results with three English datasets (the LiveJournal's blogs dataset, Aman's blogs dataset, and Alm's fairy tales dataset) show that the new
method signicantly outperforms the traditional word-based text classification
methods. The results show that the PPM compression-based classification method is able to distinguish between emotional and non-emotional
text with high accuracy, between texts involving Happiness and Sadness emotions (with 79.1% accuracy for Aman's dataset and 76.9% for Alm's datasets)
and texts involving Ekman's six basic emotions for the LiveJournal dataset
(87.4% accuracy). Results also show that the method outperforms traditional
feature-based classifiers such as Naive Bayes and SMO in most cases
in terms of accuracy, precision, recall and F-measure. In order to see how well the classifier performs on another language not related to English and also in order to create another Arabic benchmark corpus for future emotion classification experiments, we created a new Iraqi Arabic Emotion Corpus (IAEC) dataset annotated according to Ekman's basic emotions. This dataset is composed of Facebook posts written in the Iraqi dialect. We evaluated the quality of this dataset using four external judges which resulted in an average inter-annotation agreement of 0.751. We then explored six different supervised machine learning methods to test the new dataset. We used standard Weka classifiers ZeroR, J48, Naive Bayes, Multinomial Naive Bayes for Text and SMO. We compared these results with our compression-based classifier PPM. Our study reveals that the PPM classifier significantly outperforms the other classifiers for the new dataset achieving the highest results in terms of accuracy, precision, recall, and Fmeasure.
We also designed and investigated another new classification technique
motivated by information divergence to recognize Ekman's emotions in text.
We used the three datasets written in the English Language and the one in
the Arabic Language to evaluate the new method. The new method was able
to achieve a better result for Alm's dataset in terms of accuracy, precision,
recall and F-measure than PPM and standard Weka classifiers. The new
method also outperforms all standard Weka classifiers for all four datasets.
Finally, these results show that our proposed technique is promising as an
alternative technique for English and Arabic text categorization in general.