UMDSub at SemEval-2018 Task 2: Multilingual Emoji Prediction Multi-channel Convolutional Neural Network on Subword Embedding

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 395–399 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics UMDSub at SemEval-2018 Task 2: Multilingual Emoji Prediction Multi-channel Convolutional Neural Network on Subword Embedding Zhenduo Wang and Ted Pedersen Department of Computer Science University of Minnesota Duluth, MN 55812, USA {wang7211,tpederse}@d.umn.edu Abstract This paper describes the UMDSub system that participated in Task 2 of SemEval-2018. We developed a system that predicts an emoji given the raw text in a English tweet. The system is a Multi-channel Convolutional Neu- ral Network based on subword embeddings for the representation of tweets. This model im- proves on character or word based methods by about 2%. Our system placed 21st of 48 par- ticipating systems in the official evaluation. 1 Introduction People use emojis on social media with text in or- der to express their emotions or reveal meanings hidden by figurative language. However, under- standing the semantic relationship between texts and emojis is a challenging task because emojis often have different interpretations that depend on the reader. (Barbieri et al., 2018) shows that a carefully designed machine learning system can perform better than humans on an emoji prediction task. The system which gives the best predictions in that paper is a long short term memory (LSTM) network with pre-trained character embeddings as the text representation method. (Barbieri et al., 2018) provides a new task in the context of Twitter and describes several powerful models in text sequence modeling and classifica- tion. Our work explores other possible configura- tions for the system, including subword embed- ding as a text representation method and multi- channel Convolutional Neural Networks as a fea- ture extraction method. We conduct experiments on the Semeval-2018 Task 2 Multilingual Emoji Prediction data set and compare our results with other methods. Our results show that subword em- bedding based Convolutional Neural Network sys- tem is effective. It improves on character embed- ding by 2.1% and word embedding by 1.8%. 2 Task Description SemEval–2018 Task 2 has two subtasks which are similar but in different languages. Subtask A is emoji prediction on English tweets, while subtask B is on Spanish. Both tasks are to predict 1 out of 20 emojis (19 in Spanish subtask) given only the text of a tweet. There are ∼500K tweets in the En- glish training set and ∼100K tweets in the Spanish training set. The organizers use the macro F-score to evaluate the performance of the systems. This is by definition a text classification task. It is more challenging than tasks such as senti- ment analysis or authorship identification because it is on Twitter data which is extremely large and changes frequently. We build UMDSub for the English subtask and conduct several experiments with different settings. 3 System Description UMDSub is a multi-channel Convolutional Neu- ral Network based on subword embedding as the text representation. We use byte-pair-encoding for subword segmentation. 3.1 Word Segmentation The goal of the word segmentation step is to break …

Content truncated to 500 words