YNU-HPCC at SemEval-2018 Task 1: BiLSTM with Attention based Sentiment Analysis for Affect in Tweets

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 273–278 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics YNU-HPCC at SemEval-2018 Task 1: BiLSTM with Attention Based Sentiment Analysis for Affect in Tweets You Zhang, Jin Wang and Xuejie Zhang School of Information Science and Engineering Yunnan University Contact:Kunming, P.R. China xjzhang@ynu.edu.cn Abstract This paper describes the system we built as the YNU-HPCC team in the SemEval-2018 com- petition. As participants of Task 1, named Af- fect in Tweets, we implemented the sentimen- t system for all five subtasks in English and Spanish. All subtasks involved predicting e- motion or sentiment intensity (regression and ordinal classification) and determining emo- tions (multi-label classification). Our system mainly applied the bidirectional long-short ter- m memory (BiLSTM) model with an atten- tion mechanism. We used BiLSTM in order to extract word information from both direc- tions. The attention mechanism was used to find the contribution of each word to improv- ing the scores. Furthermore, based on the BiLSTM with an attention mechanism, a few deep-learning algorithms were employed for different subtasks. For regression and ordinal classification tasks, we used domain adapta- tion and ensemble learning methods to lever- age the base model, while a single base model was used for the multi-label task. Our system achieved very competitive results on the offi- cial leaderboard. 1 Introduction Sentiment analysis is an area of natural language processing (NLP), which aims to systematically i- dentify and study affective state, and to quantify subjective sentiment expressed in texts. Tweets in Twitter always constitute a challenging task a- mong NLP problems because of the colorful writ- ing styles used. In previous work on sentiment analysis tasks, researchers usually used a variety of hand-crafted features and sentiment lexicons to generate the so- lution system by combining traditional method- s such as naive Bayes, support vector machines (SVMs) (Mohammad et al., 2013), and decision trees (Blake, 2007). Recently, many ensemble learning models based on these traditional meth- ods (Giorgis et al., 2016) have attracted the interest of researcher and have shown good results. These approaches require long-term studies to gather in- formation from massive or unstructured dataset- s, and often result in redundant or missing fea- tures. In contrast, the novel deep learning method (Socher et al., 2013) has immediately and shown exceptionally good results in NLP. In this paper, we primarily present a deep learn- ing system for the SemEval-2018 shared Task 1: Affect in Tweets. We employ the bidirectional long short-term memory with an attention mech- anism (BiLSTMATT ) as a base model. For the re- gression and ordinal classification tasks, we used fine-tuning methods on the base model, combined with multi-tasking and AdaBoost algorithm. We use a simple BiLSTM with an attention mechanis- m for the multi-label task. Our contributions are as follows: • We propose a base model combining the BiL- STM with an attention mechanism for the sentiment analysis problem. • Using the base model, a domain adaptation method of fine-tuning …

Content truncated to 500 words