Using word embeddings for ontology-driven aspect-based sentiment analysis
Nowadays, the Web is the main platform to gather information. The growing amount of freely available unstructured data has increased the interest in sentiment analysis, where the goal is to extract opinions from text. In this paper we focus on review-level aspect-based sentiment analysis, where we predict the sentiment of a certain aspect in a review. We propose a two-stage sentiment analysis algorithm. In the first stage a domain ontology is utilized to predict the sentiment. If the domain ontology stage is inconclusive, a back-up stage based on an SVM bag-of-words model is employed. Furthermore, the use of word embeddings to improve the domain ontology coverage in the first stage by finding semantically similar words is investigated. We find that the two-stage approach significantly outperforms two baseline methods and achieves competitive results for the SemEval-2016 data. Furthermore, by not employing the back-up stage, we still perform significantly better than the baselines. Lastly, we find that employing word embeddings improves the accuracy when the domain ontology size is relatively small.
|Keywords||Aspect-based sentiment analysis, Domain ontology, Review-level sentiment analysis, Word embeddings|
|Persistent URL||dx.doi.org/10.1145/3341105.3373848, hdl.handle.net/1765/126164|
|Conference||35th Annual ACM Symposium on Applied Computing, SAC 2020|
De Kok, S. (Sophie), & Frasincar, F. (2020). Using word embeddings for ontology-driven aspect-based sentiment analysis. In Proceedings of the ACM Symposium on Applied Computing (pp. 834–842). doi:10.1145/3341105.3373848