NileTMRG at SemEval-2016 Task 5: Deep Convolutional Neural Networks for Aspect Category and Sentiment Extraction

Proceedings of SemEval-2016, pages 271–276, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics NileTMRG at SemEval-2016 Task 5: Deep Convolutional Neural Networks for Aspect Category and Sentiment Extraction Talaat Khalil and Samhaa R. El-Beltagy Center for informatics sciences Nile University Giza, Egypt t.maher@nu.edu.eg, samhaa@computer.org Abstract This paper describes our participation in the SemEval-2016 task 5, Aspect Based Sentiment Analysis (ABSA). We participated in two slots in the sentence level ABSA (Subtask 1) namely: aspect category extraction (Slot 1) and sentiment polarity extraction (Slot 3) in Eng- lish Restaurants and Laptops reviews. For Slot 1, we applied different models for each do- main. In the restaurants domain, we used an en- semble classifier for each aspect which is a combination of a Convolutional Neural Net- work (CNN) classifier initialized with pre- trained word vectors, and a Support Vector Ma- chine (SVM) classifier based on the bag of words model. For the Laptops domain, we used only one CNN classifier that predicts the as- pects based on a probability threshold. For Slot 3, we incorporated domain and aspect knowledge in one ensemble CNN classifier in- itialized with fine-tuned word vectors and used it in both domains. In the Restaurants domain, our system achieved the 2nd and the 3rd places in Slot 1 and Slot 3 respectively. However, we ranked the 8th in Slot 1 and the 5th in Slot 3 in the Laptops domain. Our extended experi- ments show our system could have ranked 2nd in the Laptops domain in Slot 1 and Slot 3, had we followed the same approach we followed in the Restaurants domain in slot 1 and trained each domain separately in Slot 3. 1 http://alt.qcri.org/semeval2016/task5/ 1 Introduction Due to the increasing numbers of user generated re- views written every day within e-commerce web- sites, a great interest has been shown in the senti- ment analysis research community to build intelli- gent systems that can accurately tackle the task of sentiment analysis in these reviews. In this context, the SemEval-2016 ABSA, task 51, Subtask 1 addresses a number of research problems related to this topic, including building systems that are able to extract aspect categories (Slot 1) and de- termine the sentiment polarity towards each aspect in each sentence (Slot-3) which were the two slots in which we participated. The best results for Slot 1 in SemEval-2015 ( Pontiki et al., 2015), were achieved by the NLANGP team (Toh and Su, 2015). The team tack- led the problem by modeling it as a multi-class clas- sification problem with binary classifiers for each aspect. They used a neural network with one hidden layer and features based on word n-grams, brown and k-means word clusters from Amazon and Yelp datasets and parsing features. For Slot 3, the best re- sults were achieved by the Sentiue team (Saias, 2015) who used a Maximum Entropy classifier with domain and aspect features and features based on word n-grams, lemmas, negation terms, exclama- tion and question marks, sentiment lexicons, and POS …

Content truncated to 500 words