ECNU at SemEval-2017 Task 5: An Ensemble of Regression Algorithms with Effective Features for Fine-Grained Sentiment Analysis in Financial Domain

Abstract

This paper describes our systems submitted to the Fine-Grained Sentiment Analysis on Financial Microblogs and News task (i.e., Task 5) in SemEval-2017. This task includes two subtasks in microblogs and news headline domain respectively. To settle this problem, we extract four types of effective features, including linguistic features, sentiment lexicon features, domain-specific features and word embedding features. Then we employ these features to construct models by using ensemble regression algorithms. Our submissions rank 1st and rank 5th in subtask 1 and subtask 2 respectively.

SemEval-2017 Task 5: Fine-Grained Sentiment Analysis on Financial Microblogs and News

“…The Hybrid (ML, Lex) technique by Jiang et al (2017) ranked first for this track, whereas the Hybrid (DL, Lex) technique by Ghosal et al (2017) ranked second. The system placing third (Deborah et al, 2017) adopted a ML technique.…”

“…• Artificial Neural Network (ANN) -adopted by • Naive Bayes (NB) -adopted by Seyeditabari et al (2017) • Multi-Kernel Gaussian Process (MKGP) -adopted by Deborah et al (2017) • XGBoost Regressor (XGB) -adopted by Nasim (2017); Jiang et al (2017) • Boosted Decision Tree Regression (BDTR) -adopted by Symeonidis et al (2017) • AdaBoost Regressor (ABR) -adopted by Jiang et al (2017) • Bagging Regressor (BR) -adopted by Jiang et al (2017) • Gradient Boosting Regressor (GBR) -adopted by Jiang et al (2017) • Least Absolute Shrinkage and Selection Operator (LASSO) -adopted by Jiang et al (2017) The most common ML techniques used overall -by 4 participants-were RF, SVM and SVR. The SVR was part of the ensemble regression model used by the system that ranked first for this track (Jiang et al, 2017).…”

“…• SentiWordNet 7 -adopted by Cabanski et al • Amazon Laptop Sentiment Lexicon 11 -adopted by Chen et al (2017) • Macquarie Semantic Orientation Lexicon 12 -adopted by Chen et al (2017) 9 http://saifmohammad.com/Lexicons/ Sentiment140-Lexicon-v0.1.zip 10 http://saifmohammad.com/Lexicons/Yelprestaurant-reviews.zip 11 http://saifmohammad.com/Lexicons/ Amazon-laptop-electronics-reviews.zip • AFINN 14 -adopted by Jiang et al (2017) • Corpus of Business News (Pivovarova et al, 2013) -adopted by Pivovarova et al (2017) The following three lexica listed are the ones mostly used overall: (i) the Loughran and McDonald Sentiment Word (rank 2), (ii) Opinion Lexicon (rank 1, 2) and (iii) MPQA Subjectivity Lexicon (rank 1, 2). An interesting observation is that the systems that ranked first (Jiang et al, 2017) and second (Ghosal et al, 2017) in this track utilised all three lexicons (ranked system using the particular lexicon represented next to each name), whereby lexica (ii) and (iii) were used by both.…”

“…An interesting observation is that the systems that ranked first (Jiang et al, 2017) and second (Ghosal et al, 2017) in this track utilised all three lexicons (ranked system using the particular lexicon represented next to each name), whereby lexica (ii) and (iii) were used by both. Seyeditabari et al (2017) extended Loughran and McDonald's word list of positive and negative words with 10,000 financial reports containing a summary of the company's performances in order to add features to the training dataset In Cabanski et al (2017), the authors, besides using the sentiment lexica identified above, also built and used a custom Financial Sentiment Lexicon.…”

“…Some participants also performed dependency parsing, POS tagging, stemming and URL resolution, besides other specific tasks, such as filtering out all named entities and keeping only "general" tokens given that they are generally the ones carrying the sentiment (Rotim et al, 2017). Same as track 1, NLTK was the tool mostly used (Ghosal et al, 2017;Deborah et al, 2017;Kumar et al, 2017;Symeonidis et al, 2017;Jiang et al, 2017) for pre-processing, whereas Stanford CoreNLP 15 was used for performing NER, sentence breaking and parsing. (Nasim, 2017;Rotim et al, 2017;Schouten et al, 2017;Chen et al, 2017;Jiang et al, 2017) Figure 7 shows all the techniques used by each system of the 17 participants.…”