Emotion Recognition on Twitter Using Neural Networks

Niko Colnerič (2019) Emotion Recognition on Twitter Using Neural Networks. PhD thesis.

Preview

PDF Download (3438Kb)

Abstract

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.

Actions (login required)