Where are the emotional sentences in this review? The authors of this paper offer a working solution to answer this question. They apply machine learning based on data annotated with emotional categories and the fusion of classification results, a useful extra step that is shown to be productive in improving the accuracy of identifying emotional sentences.

Performance of the extreme learning machine (ELM) classifier, to classify sentences as either emotional or unemotional, is compared to a baseline of support vector machine (SVM) classification performance. Results show that ELM achieves a similar F-measure performance (a balanced combination of precision and recall) to SVM, in a considerably faster time. Reasonable accuracy is obtained through classification using emotional words, although linguistically significant aspects such as context and part of speech are not accounted for. The methodology is otherwise sound and well grounded.

The authors adopted a target of binary classification, emotional or neutral, instead of a finer-grained identification of emotion type or sentiment (positive or negative emotion). This is somewhat disappointing given the depth of information in the annotated data used for training, a corpus of fairy tales. The corpus is annotated to identify emotional sentences and label them with seven types of emotion. Only two annotators performed the annotation. More annotators would probably have made for a more stable training dataset. Nevertheless, this is a rich training set that could feasibly be explored for more ambitious learning and classification aspirations.

The paper does not discuss the results in detail or make many detailed conclusions, though the results are presented in a tabular form for readers to draw their own conclusions. I suspect that the methodology has generated more significant results than the paper draws attention to. Though the format of the conference paper may well restrict a lengthier presentation of results, space could have been found by reporting fewer details of information that can be found elsewhere (such as the details of ELM and the explanation of evaluation criteria). I also suspect that this methodology has the potential to produce more fine-grained evaluations of emotional sentences, and could be used to attempt more ambitious tasks such as identifying how emotional a sentence is or what type of emotion is being displayed. Sadly, this paper only hints at such potential.