LT3: Applying Hybrid Terminology Extraction to Aspect-Based Sentiment Analysis

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 719–724, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics LT3: Applying Hybrid Terminology Extraction to Aspect-Based Sentiment Analysis Orph´ee De Clercq, Marjan Van de Kauter, Els Lefever and V´eronique Hoste LT3, Language and Translation Technology Team Department of Translation, Interpreting and Communication – Ghent University Groot-Brittanni¨elaan 45, 9000 Ghent, Belgium Firstname.Lastname@UGent.be Abstract The LT3 system perceives ABSA as a task consisting of three main subtasks, which have to be tackled incrementally, namely aspect term extraction, classification and polarity classification. For the first two steps, we see that employing a hybrid terminology extrac- tion system leads to promising results, espe- cially when it comes to recall. For the polar- ity classification, we show that it is possible to gain satisfying accuracies, even on out-of- domain data, with a basic model employing only lexical information. 1 Introduction There exists a large interest in sentiment analysis of user-generated content. Until recently, the main research focus has been on discovering the overall polarity of a certain text or phrase. A noticeable shift has occurred to consider a more fine-grained approach, known as aspect-based sentiment analysis (ABSA). For this task the goal is to automatically identify the aspects of given target entities and the sentiment expressed towards each of them. In this paper, we present the LT3 system that participated in this year’s SemEval 2015 ABSA task. Though the focus was on the same domains (restaurants and laptops) as last year’s task (Pontiki et al., 2014), it differed in two ways. This time, entire reviews were to be annotated and for one subtask the systems were confronted with an out-of-domain test set, unknown to the participants. The task ran in two phases. In the first phase (Phase A), the participants were given two test sets (one for the laptops and one for the restaurants do- main). The restaurant sentences were to be anno- tated with automatically identified <target, aspect category> tuples, the laptop sentences only with the identified aspect categories. In the second phase (Phase B), the gold annotations for the above two datasets, as well as for a hidden domain, were given and the participants had to return the corresponding polarities (positive, negative, neutral). For more in- formation we refer to Pontiki et al. (2015). We tackled the problem by dividing the ABSA task into three incremental subtasks: (i) aspect term extraction, (ii) aspect term classification and (iii) as- pect term polarity estimation (Pavlopoulos and An- droutsopoulos, 2014). The first two are at the basis of Phase A, whereas the final one constitutes Phase B. For the first step, viz. extracting terms (or tar- gets), we wanted to test our in-house hybrid termi- nology extraction system (Section 2). Next, we per- formed a multiclass classification task relying on a feature space containing both lexical and semantic information to aggregate the previously identified terms into the domain-specific and predefined as- pects (or aspect categories) (Section 3). Finally, we performed polarity classification …

Content truncated to 500 words