SentiSys at SemEval-2016 Task 5: Opinion Target Extraction and Sentiment Polarity Detection

Proceedings of SemEval-2016, pages 350–355, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics SentiSys at SemEval-2016 Task 5: Opinion Target Extraction and Sentiment Polarity Detection Hussam Hamdan Aix-Marseille University hamdan.hussam@gmail.com Abstract This paper describes our contribution in Opin- ion Target Extraction and Sentiment Polarity sub-tasks of SemEval 2016 ABSA task. Conditional Random Field model has beenA adopted for opinion target extraction. Logistic Regression model with a weightingA schema of positive and negative labels has been used for sentiment polarity. Our sub- mission for opinion target extraction is ranked second among the constrained systems which do not use additional resources and sixth over 19 submissions among the constrained and un- constrained systems in English restaurant re- views. Our submission for Sentiment Polarity is ranked eighth over 22 submissions on the laptop reviews. 1 Introduction Classifying opinion texts at document or sentence levels is not sufficient for applications which need to identify the opinion targets. Even if the docu- ment is about one entity, many applications need to determine the opinion about each aspect of the en- tity. A user may express a positive opinion towards a restaurant, but he may have a negative opinion to- wards some aspects as the ambiance. Therefore, we need to identify the aspects and determine whether the sentiment is positive, negative or neutral towards each one. This task is called Aspect-Based Senti- ment Analysis or Feature-Based opinion mining as called in the early work (Hu and Liu, 2004). Aspect-Based Sentiment Analysis is composed of four subtasks: 1. Opinion Target Expression Extraction Opinion Target Expression is a linguistic ex- pression used in a given text to refer to an as- pect of the reviewed entity. This subtask aims at identifying all the aspect terms present in a given set of sentences with pre-identified en- tities such as restaurants, laptops. An opinion target names a particular aspect of the target en- tity. For example: ”I liked the service and the staff, but not the food” ”The hard disk is very noisy” The service, staff and food are opinion target expressions. hard disk is multi-word opinion target expression which will be treated as a sin- gle term. 2. Aspect Sentiment Detection Each identified opinion target has to be as- signed to one of the following polarity labels: positive, negative or neutral. For example: Input: ”I hated their fajitas, but their salads were great” Output: {fajitas: negative, salads: positive} 3. Aspect Category Detection This subtask aims at identifying the aspect categories discussed in a given sentence from a predefined set of aspect categories such as 350 price and food. Aspect categories are typically coarser than the opinion targets, and they do not necessarily occur as terms in a given sentence. For example, given the set of aspect categories of restaurant entity {food, service, price, am- biance}: ”The restaurant was too expensive” The aspect category is {price}. ”The restaurant was expensive, but the menu was great” The aspect categories are {price, food}. 4. Aspect Category Polarity Given a …

Content truncated to 500 words