Multi-Granular Aspect Aggregation in Aspect-Based Sentiment Analysis

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 78–87, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Multi-Granular Aspect Aggregation in Aspect-Based Sentiment Analysis John Pavlopoulos and Ion Androutsopoulos Department of Informatics Athens University of Economics and Business Patission 76, GR-104 34 Athens, Greece http://nlp.cs.aueb.gr/ Abstract Aspect-based sentiment analysis estimates the sentiment expressed for each particu- lar aspect (e.g., battery, screen) of an en- tity (e.g., smartphone). Different words or phrases, however, may be used to re- fer to the same aspect, and similar as- pects may need to be aggregated at coarser or finer granularities to fit the available space or satisfy user preferences. We in- troduce the problem of aspect aggrega- tion at multiple granularities. We decom- pose it in two processing phases, to al- low previous work on term similarity and hierarchical clustering to be reused. We show that the second phase, where aspects are clustered, is almost a solved prob- lem, whereas further research is needed in the first phase, where semantic simi- larity measures are employed. We also introduce a novel sense pruning mecha- nism for WordNet-based similarity mea- sures, which improves their performance in the first phase. Finally, we provide pub- licly available benchmark datasets. 1 Introduction Given a set of texts discussing a particular en- tity (e.g., reviews of a laptop), aspect-based senti- ment analysis (ABSA) attempts to identify the most prominent (e.g., frequently discussed) aspects of the entity (e.g., battery, screen) and the average sentiment (e.g., 1 to 5 stars) for each aspect or group of aspects, as in Fig. 1. Most ABSA systems perform all or some of the following (Liu, 2012): subjectivity detection to retain only sentences (or other spans) expressing subjective opinions; as- pect extraction to extract (and possibly rank) terms corresponding to aspects (e.g., ‘battery’); aspect aggregation to group aspect terms that are near- synonyms (e.g., ‘price’, ‘cost’) or to obtain aspects Figure 1: Aspect groups and scores of an entity. at a coarser granularity (e.g., ‘chicken’,‘steak’, and ‘fish’ may be replaced by ‘food’ in restaurant reviews); and aspect sentiment score estimation to estimate the average sentiment for each aspect or group of aspects. In this paper, we focus on aspect aggregation, the least studied stage of the four. Aspect aggregation is needed to avoid reporting separate sentiment scores for aspect terms that are very similar. In Fig. 1, for example, showing sep- arate lines for ‘money’, ‘price’, and ‘cost’ would be confusing. The extent to which aspect terms should be aggregated, however, also depends on the available space and user preferences. On de- vices with smaller screens, it may be desirable to aggregate aspect terms that are similar, though not necessarily near-synonyms (e.g., ‘design’, ‘color’, ‘feeling’) to show fewer lines (Fig. 1), but finer as- pects may be preferable on larger screens. Users may also wish to adjust the granularity of aspects, e.g., by stretching or narrowing the height of Fig. 1 on a smartphone to view …

Content truncated to 500 words