Тип публикации: доклад, тезисы доклада, статья из сборника материалов конференций
Конференция: SWS International Scientific Conference on Arts and Humanities 2019; Albena, Bulgaria; Albena, Bulgaria
Год издания: 2019
Идентификатор DOI: 10.5593/SWS.ISCAH.2019.1/S14.082
Ключевые слова: emoji, feature selection, sentiment analysis, Lövheim's Cube
Аннотация: When dealing with multiclass classification problem one has to search for reliable and robust predictors to increase the quality of performance that decreases drastically with growth of classes number. In our research, we propose the sentiment analysis method backed by Lövheim Cube Emotional model including 8 emotional classes unliПоказать полностьюke traditional two (negative/ positive). To increase the accuracy in case of multi-class classification we need to extract more features from the text than it is usually demanded for binary classifier. One of the most promising features is using emojis on par with text as predictors, as they (by definition) express emotions. To prove the hypothesis we ran our classifier using three feature sets transformed to Bag-Of-Words representation: 1) pure emoji BOW where each occurring emoji was represented as a discrete word (Emojis model), 2) normalized words without emojis (Text model), 3) normalized words including emojis, where each occurring emoji was represented as a discrete word (Text + emojis model). In spite of our expectations, the Emojis model has demonstrated the worst results, but the Text model has performed rather well, showing better total benchmarks than the other two models. The analysis of false positives (6,8 % from all sample) those sentiment was predicted triply (by text only, by both the emojis and the text, by trained before classifier) but was still inconsistent with the real sentiment, has revealed 4 types of situations engendering the discrepancy between text sentiment and emojis sentiment. Firstly, due to the sarcastic or ironical tonality texts express sentiment opposite to those of emoji. Secondly, texts express sentiment different (but not obviously opposite) to those of emoji to realize the speaker's face-saving communicative strategy. Then, text has no emotions, but there are one or more emoji used for making the message more attractive for recipients. Finally, text tonality is ambiguous and emoji's emotional value is unequivocal. These results led us to the conclusion that emoji does not serve as complement for the text, but should be considered as text's meta-data which cannot be straightforwardly processed like an ordinary text token.
Журнал: 6th SWS International Scientific Conference on Arts and Humanities 2019
Выпуск журнала: 6. Issue 1
Номера страниц: 645-652
Издатель: Общество с ограниченной ответственностью СТЕФ92 Технолоджи