Distributed representations of sentences have been developed recently torepresent their meaning as real-valued vectors. However, it is not clear howmuch information such representations retain about the polarity of sentences.To study this question, we decode sentiment from unsupervised sentencerepresentations learned with different architectures (sensitive to the order ofwords, the order of sentences, or none) in 9 typologically diverse languages.Sentiment results from the (recursive) composition of lexical items andgrammatical strategies such as negation and concession. The results aremanifold: we show that there is no `one-size-fits-all' representationarchitecture outperforming the others across the board. Rather, the top-rankingarchitectures depend on the language and data at hand. Moreover, we find thatin several cases the additive composition model based on skip-gram word vectorsmay surpass supervised state-of-art architectures such as bidirectional LSTMs.Finally, we provide a possible explanation of the observed variation based onthe type of negative constructions in each language.
展开▼
