Abstract
Word2Vec has become one of the most relevant neural networks to generate word embeddings for NLP applications. Despite that, little has been investigated in terms of its sensitivity to the word vectors' length (n) and the window size (w). Thus, the present paper performs a sensitivity analysis of Word2Vec when applied to generate word embeddings for a deep neural architecture used to classify emotion states in tweets. Furthermore, we present a computational performance analysis to investigate how the system scales as a function of n and w in different computing environments. The results show that a window size of approximately half the tweet length (8 words) and a value of n = 50 suffices to find good performances. Also, by increasing these values one may unnecessarily increase the computational cost.