Contextual semantics using hierarchical attention network for sentiment classification in social internet-of-things

To steer impactful data-driven possibilities, Social Internet of Things (SIoT) comes with the collaboration of IoT and social networking and finds deeper insights into consumer behavior for better user experiences. Social networks can permeate intelligence to aid autonomous decision making by enhancing the service needs and communication among object peers in a SIoT. But the user-generated data has multiple layers of meaning which necessitate AI-driven solutions such as sentiment analysis to handle the dynamics. As people express opinions in complex ways and use rhetorical devices like sarcasm, irony, and implication etc., considering only the lexical content can be misleading. Moreover, intra-textual and sub-sentential reversals, topic drift, negation can further misrepresent sentiment, fostering the need to recognize and incorporate contextual semantics for increasing the sentiment classification accuracy. This research evaluates the use of hierarchical attention network (HAN) to classify sentiments in real-time Twitter data, including the multiple sentence tweets and multi-tweet threads. HAN allows differential contribution of various parts of tweet (tweet-sentence-word) to its essential meaning as it introduces two attentive mechanisms and the context-dependent importance of the parts of tweet are considered when constructing the representation of the document. The model is evaluated on two benchmark datasets and compares favorably to state-of-the-art approaches giving an effective solution to tweet-level analysis of sentiments in SIoT.