DeepMiner at SemEval-2018 Task 1: Emotion Intensity Recognition Using Deep Representation Learning

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 305–312 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics DeepMiner at SemEval-2018 Task 1: Emotion Intensity Recognition Using Deep Representation Learning Habibeh Naderi Dalhousie University, Canada habibeh.naderi@dal.ca Behrouz H. Soleimani Dalhousie University, Canada behrouz.hajisoleimani@dal.ca Svetlana Kiritchenko, Saif M. Mohammad National Research Council Canada {svetlana.kiritchenko, saif.mohammad}@nrc-cnrc.gc.ca Stan Matwin Dalhousie University, Canada stan@cs.dal.ca Abstract In this paper, we propose a regression system to infer the emotion intensity of a tweet. We develop a multi-aspect feature learning mech- anism to capture the most discriminative se- mantic features of a tweet as well as the emo- tion information conveyed by each word in it. We combine six types of feature groups: (1) a tweet representation learned by an LSTM deep neural network on the training data, (2) a tweet representation learned by an LSTM network on a large corpus of tweets that contain emo- tion words (a distant supervision corpus), (3) word embeddings trained on the distant super- vision corpus and averaged over all words in a tweet, (4) word and character n-grams, (5) fea- tures derived from various sentiment and emo- tion lexicons, and (6) other hand-crafted fea- tures. As part of the word embedding train- ing, we also learn the distributed representa- tions of multi-word expressions (MWEs) and negated forms of words. An SVR regressor is then trained over the full set of features. We evaluate the effectiveness of our ensemble fea- ture sets on the SemEval-2018 Task 1 datasets and achieve a Pearson correlation of 72% on the task of tweet emotion intensity prediction. 1 Introduction The widespread use of micro-blogging and social networking websites such as Twitter for convey- ing information, sharing opinions, and expressing feelings, makes the sentiment analysis of tweets an attractive area of research. However, senti- ment analysis is challenging because people often convey their emotions indirectly and creatively, rather than explicitly stating how they feel. Sen- timent analysis of tweets is additionally challeng- ing because of the frequent occurrences of non- standard language and poor grammatical structure. Tweets also often contain misspellings, abbrevia- tions, hashtags, and emoticons. Various machine learning approaches have been developed for Twitter sentiment classification. Most of these algorithms train a classifier over tweets with manually annotated sentiment inten- sity labels and learn the most discriminative fea- tures. Hence, designing an effective feature engi- neering algorithm can improve classification per- formance, greatly. Mohammad et al. (2013; 2014) used many different sentiment lexicons (manually created and automatically generated), as well as a variety of hand-crafted features to build the top- ranked system for Twitter sentiment classification tasks in SemEval-2013 and SemEval-2014. Sen- timent lexicons, either hand-crafted or algorith- mically generated, consist of words and their as- sociated polarity scores. However, since feature engineering is labour intensive and usually needs domain-specific knowledge, sentiment classifica- tion algorithms with less dependency on feature engineering are attracting considerable interest. Socher et al. (2013) proposed a feature learning algorithm to discover explanatory factors in sen- …

Content truncated to 500 words