AUEB-ABSA at SemEval-2016 Task 5: Ensembles of Classifiers and Embeddings for Aspect Based Sentiment Analysis

Proceedings of SemEval-2016, pages 312–317, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics AUEB-ABSA at SemEval-2016 Task 5: Ensembles of Classifiers and Embeddings for Aspect Based Sentiment Analysis Dionysios Xenos, Panagiotis Theodorakakos, John Pavlopoulos, Prodromos Malakasiotis and Ion Androutsopoulos NLP Group, Department of Informatics Athens University of Economics and Business Patission 76, GR-104 34 Athens, Greece http://nlp.cs.aueb.gr Abstract This paper describes our submissions to the Aspect Based Sentiment Analysis task of SemEval-2016. For Aspect Category Detec- tion (Subtask1/Slot1), we used multiple en- sembles, based on Support Vector Machine classifiers. For Opinion Target Expression extraction (Subtask1/Slot2), we used a se- quence labeling approach with Conditional Random Fields. For Polarity Detection (Sub- task1/Slot3), we used an ensemble of two su- pervised classifiers, one based on hand crafted features and one based on word embeddings. Our systems were ranked in the top 6 positions in all the tasks we participated. The source code of our systems is publicly available. 1 Introduction The amount of user-generated content on the web has grown rapidly in recent years, leading to in- creased interest in sentiment analysis and, more gen- erally, opinion mining. The task of Aspect Based Sentiment Analysis of SemEval-2014 (SE-ABSA14) and SemEval-2015 (SE-ABSA15) was concerned with identifying the aspects of given target entities and extracting the sentiment expressed towards each aspect (Pontiki et al., 2014; Pontiki et al., 2015). The task of Aspect Based Sentiment Analysis of SemEval-2016 (SE-ABSA16) is a continuation of those tasks (Pontiki et al., 2016). We participated in Aspect Category Detection (ACD, Subtask1/Slot1), Opinion Target Expression (OTE, Subtask1/Slot2), and Polarity Detection (PD, Subtask1/Slot3). In ACD, we participated in the English language, for both Laptops and Restaurants, submitting both constrained and unconstrained systems. Our con- strained system used only the provided training data for the corresponding domain. Features were ex- tracted from lexicons created from the training data. One Support Vector Machine (SVM) classifier (Vap- nik and Vapnik, 1998) was trained for each Entity and Attribute category (called E and A respectively). Our unconstrained system used word embeddings as additional resources (Mikolov et al., 2013). For each category (E or A), we used an ensemble of two sys- tems: our constrained system and one new system, which was based on word embeddings. In OTE, we participated with both a constrained and an unconstrained system.1 The task is to iden- tify aspects of given target entities. We addressed the problem as a sequential labeling task (Toh and Wang, 2014), assigning one label to each word in a sentence, indicating whether the word was an aspect term or not. In this task, we used Conditional Ran- dom Fields (Lafferty et al., 2001). Similarly to ACD, our unconstrained system differed in that it also used word embeddings as features. In PD, we participated only with an unconstrained system, in both domains, in the English language. We used an ensemble of two classifiers. The first classifier used hand crafted features and sentiment lexicons with scores. The second classifier was based on IDF-weighted …

Content truncated to 500 words