Real-Time Sentiment Analysis Over Social Network Streams Using Deep Learning Models in Edge-Cloud Architectures

Abstract

This paper describes a real-time sentiment analysis system dedicated to social media streams, specifically from Twitter, using deep learning methods in an edge-cloud infrastructure. The recommendation system we develop aims to the classification of textual information through a hybrid deep model, consisting of both Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks. At first, a large-scale data collection approach was carried out to retrieve tweets about different domains, where emphasis was placed on real-world, high volume and noisy datasets. Our classification was composed of two stages: the first of which filters tweets according to the relevance of their topic (given a set of topics), and the second of which classifies the polarity of sentiment in the tweet (positive, negative, neutral, irrelevant). This method extends standard three-class sentiment classification to a four-label model that includes a fourth category to indicate topic irrelevance, resulting in more context-aware analysis of streaming data. In order to support real-time processing and reduce the end-to-end delay, the system was unfolded with a hybrid edge-cloud architecture, that performs initial filtering and light computation at the edge meanwhile deep learning inference and storage. Experiments show that the proposed CNN-LSTM model is superior to some baseline models and is robust with unstructured and noisy social network text. The framework is established by experimentation on a real-world use case, that visualizes Twitter data insights into near-real time, proving its practical benefits for smart observation and public opinion observation. Although encouraging results have been achieved, more improvements are still desired in dynamic visualization and sensitivity to topic drift, indicating that real-time sentiment analysis is still a challenging and growing research topic.