An Attention Arousal Space for Mapping Twitter Data

Abstract

Not every trending tweet commands your attention! Overflow of social media data will soon require mechanisms to identify the severity of a situation to evaluate if a tweet must be given ample attention. News about bomb blasts, flood, epidemic outbreak affects several sectors and impacts the local economy. Such news is covered by media and government sources which forms meta-information with tweets. Most of the algorithms, both traditional and deep learning can decipher if the opinion is positive, negative, or neutral, they cannot allocate attention to the opinion. Compared to the binary classification of positive and negative sentiments, this article proposes a glove-text-CNN algorithm and attaches meta-information to predict the attention accuracy of sentence-level data. Correlation is being established among the various parameters of a trending hashtag and the most important parameter is being reported. The primary contribution is a modified neural algorithm to define an Attention Arousal Space that is designed to capture five major insights of sentence-level documents by attaching meta-information, to include media score, time decay, retweet, favorite scores and government source to modify original algorithm. The sentiment classification algorithm used for this research is Text-CNN with glove word embeddings, which was trained on 2,000,000 sentiment marked tweets from SemEval datasets. Sentiment classification training accuracy was 98.6% and validation set accuracy was 87.13%. Our experiments show accuracy over Hierarchical Attention Networks (HAN) which proves the effectiveness of this modified algorithm with meta-information for capturing attention.