We present a novel unsupervised approach for multilingual sentiment analysisdriven by compositional syntax-based rules. On the one hand, we exploit some ofthe main advantages of unsupervised algorithms: (1) the interpretability oftheir output, in contrast with most supervised models, which behave as a blackbox and (2) their robustness across different corpora and domains. On the otherhand, by introducing the concept of compositional operations and exploitingsyntactic information in the form of universal dependencies, we tackle one oftheir main drawbacks: their rigidity on data that are structured differentlydepending on the language concerned. Experiments show an improvement both overexisting unsupervised methods, and over state-of-the-art supervised models whenevaluating outside their corpus of origin. Experiments also show how the samecompositional operations can be shared across languages. The system isavailable at http://www.grupolys.org/software/UUUSA/
展开▼
