Learning Polarity Embedding

Our goal is to extract the political-polarity information from the politicians’ speech, which could make supporters of different parties use totally different expressions referring to the same concept. We formalize this problem as learning the polarity embedding of the words. Uulike the standard methods where all aspects of a word’s meaning composites the word vector representation, we want it to specifically capture two parts: (1) the political embedding, and (2) the non-political, semantic embedding. In theory we can use any existing model as our base model. For instance, word2vec, GloVe, ELMo, BERT, XLNet. Based on that, we propose to add auxiliary tasks to specifically force the embeddings to end up in what we want. At the current stage, we have the previouslycrawled tweets, the most-recent 3, 000 tweets for the 115 and 116 representatives before March, 2019. Our experiments show that there still remain some challenges to conquer, and also show that the existing approaches are not good enough in this scenario. Our code and current-stage data are released at https://github.com/PatriciaXiao/ CS263_PolarityWordEmbedding.