SNAP: A Multi-Stage XML-Pipeline for Aspect Based Sentiment Analysis

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 578–584, Dublin, Ireland, August 23-24, 2014. SNAP: A Multi-Stage XML-Pipeline for Aspect Based Sentiment Analysis Clemens Schulze Wettendorf and Robin Jegan and Allan K¨orner and Julia Zerche and Nataliia Plotnikova and Julian Moreth and Tamara Schertl and Verena Obermeyer and Susanne Streil and Tamara Willacker and Stefan Evert Friedrich-Alexander-Universit¨at Erlangen-N¨urnberg Department Germanistik und Komparatistik Professur f¨ur Korpuslinguistik Bismarckstr. 6, 91054 Erlangen, Germany {clemens.schulze.wettendorf, robin.jegan, allan.koerner, julia.zerche, nataliia.plotnikova, julian.moreth, tamara.schertl, verena.obermeyer, susanne.streil, tamara.willacker, stefan.evert}@fau.de Abstract This paper describes the SNAP system, which participated in Task 4 of SemEval- 2014: Aspect Based Sentiment Analysis. We use an XML-based pipeline that com- bines several independent components to perform each subtask. Key resources used by the system are Bing Liu’s senti- ment lexicon, Stanford CoreNLP, RFTag- ger, several machine learning algorithms and WordNet. SNAP achieved satisfactory results in the evaluation, placing in the top half of the field for most subtasks. 1 Introduction This paper describes the approach of the SemaN- tic Analyis Project (SNAP) to Task 4 of SemEval- 2014: Aspect Based Sentiment Analysis (Pontiki et al., 2014). SNAP is a team of undergraduate students at the Corpus Linguistics Group, FAU Erlangen-N¨urnberg, who carried out this work as part of a seminar in computational linguistics. Task 4 was divided into the four subtasks As- pect term extraction (1), Aspect term polarity (2), Aspect category detection (3) and Aspect category polarity (4), which were evaluated in two phases (A: subtasks 1/3; B: subtasks 2/4). Subtasks 1 and 3 were carried out on two different datasets, one of laptop reviews and one of restaurant reviews. Subtasks 2 and 4 only made use of the latter. This work is licensed under a Creative Commons Attribution 4.0 International Licence. Page numbers and proceedings footer are added by the organisers. Licence details: http: //creativecommons.org/licenses/by/4.0/ Task Dataset Rank Score Best 1 Lap 10 of 21 0.624 0.746 1 Res 20 of 21 0.465 0.840 3 Res 6 of 15 0.782 0.886 2 Lap 7 of 23 0.641 0.705 2 Res 12 of 24 0.708 0.810 4 Res 11 of 18 0.696 0.829 Table 1: Ranking among constrained systems. The developed system consists of one module per subtask, in addition to a general infrastruc- ture and preprocessing module, All modules ac- cept training and test data in the XML format spec- fied by the task organizers. The modules can be combined into a pipeline, where each step adds new annotation corresponding to one of the four subtasks. Table 1 shows our ranking among all con- strained systems (counting only the best run from each team), the score achieved by SNAP (accu- racy or F-score, depending on subtask), and the score achieved by the best system in the respec- tive subtask. Because of a preprocessing mistake that was only discovered after phase A of the eval- uation had ended, results for subtasks 1 and 3 are significantly lower than the results achieved dur- ing development of the system. 2 …

Content truncated to 500 words