Abstract

In this research note we present a language independent system to modelOpinion Target Extraction (OTE) as a sequence labelling task. The systemconsists of a combination of clustering features implemented on top of a simpleset of shallow local features. Experiments on the well known Aspect BasedSentiment Analysis (ABSA) benchmarks show that our approach is very competitiveacross languages, obtaining best results for six languages in seven differentdatasets. Furthermore, the results provide further insights into the behaviourof clustering features for sequence labelling tasks. The system and modelsgenerated in this work are available for public use and to facilitatereproducibility of results.

 

Quick Read (beta)