Cross-Lingual Word Vectors for Deep Sentiment Analysis

There is a long history of inducing vector-based representations of linguistic utterances. In our work, we induce sentiment embeddings custom-tailored for the task of sentiment analysis. Deep neural networks normally require large annotated training corpora for each combination of language, domain, and genre. We conjecture that encoding the prior sentiment polarity of words in different domains into their word vectors may mitigate this challenge.