UMDuluth-CS8761-12: A Novel Machine Learning Approach for Aspect Based Sentiment Analysis

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 742–747, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics UMDuluth-CS8761-12: A Novel Machine Learning Approach for Aspect Based Sentiment Analysis Akshay Reddy Koppula, Ranga Reddy Pallelra, Ravikanth Repaka, Venkata Subhash Movva Department of Computer Science University of Minnesota Duluth 320 Heller Hall 1114 Kirby Drive Duluth, MN 55812-2496, USA {koppu001, palle015, repak003, movva002}@d.umn.edu Abstract This paper provides a detailed description of the approach of our system for the Aspect- Based Sentiment Analysis task of SemEval- 2015. The task is to identify the Aspect Cate- gory (Entity and Attribute pair), Opinion Tar- get and Sentiment of the reviews. For the In-domain subtask that is provided with the training data, the system is developed using a supervised technique Support Vector Machine and for the Out-of-domain subtask for which the training data is not provided, it is imple- mented based on the sentiment score of the vo- cabulary. For In-domain subtask, our system is developed specifically for restaurant data. 1 Introduction With the increase in usage of internet, most of the users record their experiences of a particular product or item in the form of online reviews. Users might express their opinion about many different aspects of an item in a review. While most of existing systems try to extract the overall polarity of a sentence, Semeval 2015 con- ducted a task on Aspect-Based Sentiment Anal- ysis and the requirement was to extract entities (e.g., Food, Price, Service for Restaurant data), attributes(e.g., Quality, Style) for each sentence and then to determine the polarity for each entity- attribute pair. The fajitas were delicious, but expensive. In the above example, there are two opinions.The first opinion has FOOD#QUALITY as the entity- attribute pair with positive polarity and second has FOOD#PRICES with negative polarity. The target for both these opinions is fajitas. Since there are two opinions with two different polarities, it is useful to identify entities, attributes and targets for each sen- tence. Our system tries a new approach of trying to split the sentence to find out more than one opinion in a sentence. Initially, all the unnecessary words are removed and then sentences are split in a way such that each split sentence has an opinion. These split sentences are given to a classifier for identifying en- tities and attributes. Later, these entities are used to extract opinion targets. Polarity is found using a classifier and voting mechanisms. The rest of the paper is structured as follows: Section 2 presents the description of SemEval- Task Aspect-Based Sentiment Analysis. Section 3 presents the description of our system. Section 4 discusses the results of our system and analyze them. Section 5 presents a conclusion to the paper. 2 SemEval Task Description The SemEval Task is divided into two subtasks. 2.1 Subtask 1 Following are the slots in the Subtask 1 2.1.1 Slot 1 - Aspect Category (Entity and Attribute) It specifies the category of the domain to which the review refers. …

Content truncated to 500 words