New Text Classification Strategy Based on a Word Embedding and Noise-Words Removal

In the intricate realm of automatic text classification, striking an optimal balance between accuracy and efficiency remains a paramount objective. While recent advancements have significantly improved precision, recall, and f-score, a glaring gap persists: many models grapple with ensuring efficiency, especially concerning time and computational resources. This inefficiency in both time and space dimensions underscores a pressing need for innovative solutions that adeptly tackle dimensionality reduction. With this challenge in focus, this paper endeavors to pioneer a fresh approach in text classification. At the heart of our methodology lies a robust neural network model, seamlessly complemented by the Word2vec model and Cosine similarity. This potent combination diligently discerns text terms imbued with public domain meanings, subsequently sidestepping them during the embedding vector generation process. To ascertain the efficacy of this strategy, a rigorous evaluation was conducted across diverse datasets, namely BBC, ECO, and SPAM, each subjected to distinct conditions and threshold values. The outcomes were not just promising but also illuminating. We observed a marked reduction ratio: 45% for the BBC dataset, 37% for the ECO, and 36% for SPAM when juxtaposed against baseline measures. What adds intrigue to these results is the realization that this dimensionality reduction was consummated without any detrimental impact on accuracy. On the contrary, in several instances, our model demonstrated an enhanced accuracy metric, a testament to its prowess.