SSN-SPARKS at SemEval-2019 Task 9: Mining Suggestions from Online Reviews using Deep Learning Techniques on Augmented Data

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 1237–1241 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 1237 SSN-SPARKS at SemEval-2019 Task 9: Mining Suggestions from Online Reviews using Deep Learning Techniques on Augmented Data Rajalakshmi S, Angel Deborah S, S Milton Rajendram, Mirnalinee T T Department of Computer Science and Engineering SSN College of Engineering Chennai 603 110, Tamil Nadu, India rajalakshmis@ssn.edu.in, angeldeborahs@ssn.edu.in miltonrs@ssn.edu.in, mirnalineett@ssn.edu.in Abstract This paper describes the work on mining the suggestions from online reviews and forums. Opinion mining detects whether the comments are positive, negative or neutral, while sugges- tion mining explores the review content for the possible tips or advice. The system developed by SSN-SPARKS team in SemEval-2019 for task 9 (suggestion mining) uses a rule-based approach for feature selection, SMOTE tech- nique for data augmentation and deep learning technique (Convolutional Neural Network) for classification. We have compared the results with Random Forest classifier (RF) and Multi- Layer Perceptron (MLP) model. Results show that the CNN model performs better than other models for both the subtasks. 1 Introduction Sentiment analysis is a process of computationally identifying and categorizing the opinions from un- structured data. This can be used to identify a user’s perspective of a product — positive, nega- tive or neutral. Opinion mining is used to identify whether the product is a success in the market or not. Suggestion mining finds out ways to enhance the product to satisfy the customers. Review texts are mainly used to identify the sen- timents of the user. Besides sentiments, review texts also contain valuable information such as ad- vice, recommendations, tips and suggestions on a variety of points of interest (Negi and Buitelaar, 2017a). These suggestions will help other cus- tomers make their choices, on the one hand, and the sellers improve their products, on the other hand. Suggestion mining is relatively a young field of research compared to sentiment analysis. While mining for suggestions, the propositional aspects like mood, modality, sarcasm, and compound statements have to be considered. It is observed that, in some cases, grammatical properties of the sentence alone can be used to identify the label, while in other cases semantics play a significant role in label classification (Negi and Buitelaar, 2017b). “Task 9 – Suggestion mining from online re- views and forums” has two subtasks (Negi et al., 2019). Subtask A is to classify a sentence into a suggestion or a non-suggestion. Subtask B is a cross-domain testing in which the model learned from a domain-specific dataset is used to classify dataset from a new domain. We have built clas- sifiers using MultiLayer Perceptron (MLP), Ran- dom Forest (RF) and Convolutional Neural Net- work (CNN) models. However, due to the im- balance in the data, we have augmented it us- ing Synthetic Minority Over-sampling TEchnique (SMOTE). We found that the CNN model per- forms better compared to RF and MLP classifiers for both the subtasks. 2 Related Work Ramanand et al. (2010) discusses the rule-based method …

Content truncated to 500 words