SA-UZH: Verb-based Sentiment Analysis

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 503–507, Dublin, Ireland, August 23-24, 2014. SA-UZH: Verb-based Sentiment Analysis Nora Hollenstein, Michi Amsler, Martina Bachmann, Manfred Klenner Institute of Computational Linguistics, University of Zurich Binzmuehlestrasse 14, CH-8050 Zurich, Switzerland {hollenstein,mamsler,bachmann,klenner}@ifi.uzh.ch Abstract This paper describes the details of our system submitted to the SemEval-2014 shared task about aspect-based sentiment analysis on review texts. We participated in subtask 2 (prediction of the polarity of aspect terms) and 4 (prediction of the polarity of aspect categories). Our ap- proach to determine the sentiment of as- pect terms and categories is based on lin- guistic preprocessing, including a com- positional analysis and a verb resource, task-specific feature engineering and su- pervised machine learning techniques. We used a Logistic Regression classifier to make predictions, which were ranked above-average in the shared task. 1 Introduction Aspect-based sentiment analysis refers to the problem of predicting the polarity of an explicit or implicit mention of a target in a sentence or text. The SemEval-2014 shared task required sen- timent analysis of laptop and restaurant reviews on sentence level and comprised four subtasks (Pontiki et al., 2014). The organizers created and shared manually labelled domain-specific training and test data sets. Two of the four subtasks dealt with determining the sentiment of a given aspect term (explicitly mentioned) or aspect category (ex- plicitly or implicitly mentioned) in a sentence. The subtasks we participated in do not include the recognition of aspects. Given the sentence “The sushi rolls were perfect, but overall it was too ex- pensive.”, “sushi rolls” is an aspect term, and the corresponding aspect categories are “food” and This work is licenced under a Creative Commons Attribution 4.0 International License. Page numbers and proceedings footer are added by the organizers. License details: http: //creativecommons.org/licenses/by/4.0/ “price”. The correct predictions would be the fol- lowing: • Subtask 2 (aspect terms): {sushi rolls ¨ pos- itive} • Subtask 4 (aspect categories): {food ¨ posi- tive, price ¨ negative} To solve these tasks, we introduce a Logis- tic Regression Model for target-specific sentiment analysis. Features are derived from a fine-grained polarity lexicon, a verb resource specifying expec- tations and effects of the verbs functional roles, and a compositional analysis. In our experiments on the restaurant and laptop reviews data for the SemEval-2014 shared task, we found that im- provements over the baseline are possible for all classes except “conflict”. 2 Related Work We focus on the question whether fine-grained lin- guistic sentiment analysis improves target-specific polarity classification. Existing approaches to aspect-based sentiment detection have focused on different aspects of this task, e.g. the identifi- cation of targets and their components (Popescu and Etzioni, 2005) and sentence-level composition (Moilanen and Pulman, 2007). Ding et al. (2008) and Hu and Liu (2004) produced lexicon-based approaches, which perform quite well in a large number of domains, and Blair-Goldensohn et al. (2008) combined lexicon-based methods and su- pervised learning. Jiang et al. (2011) used a depen- dency parser to generate a set of …

Content truncated to 500 words