Performance Evaluation of Word and Sentence Embeddings for Finance Headlines Sentiment Analysis

Nowadays, tremendous number of financial online articles are published every day. Numerous natural language processing (NLP) algorithms and methodologies have arose, not only for correct, but also for fast financial sentiment extraction. Currently, word and sentence encoders are popular topic in NLP field, due to their ability to represent them as dense vectors in a continuous real numbers space, referred to as embeddings. These low dimensional embedding vectors are appropriate for deep neural networks (DNN) inputs, and their invention boosted the performance of multiple of NLP tasks. In this paper, we evaluate different word and sentence embeddings in combination with standard machine learning and deep-learning classifiers for financial texts sentiment extraction. Our evaluation shows the BiGRU+Attention architecture with word embedding as features, give the best score in overall evaluation.