BEC-BiLSTM: Real-time Informative Tweet Prediction And Knowledge Graph Construction For Crisis Events

Nowadays,it is a common practice for the general public to use different social media platforms mainly Twitter to share ideas,opinions,and information.During crisis situations,like short-term disasters or longer-term events such as pandemics(COVID-19),Twitter can be a valuable source of information.Social media can often provide facts about changes far faster than traditional sources such as official news,and can also provide personal viewpoints on events,such as opinions or unique requirements because of this during a crisis and mass emergencies,Twitter is increasingly being used as a popular source of information to communicate and share about the situation at the crisis environment,report about the affected people and casualties.A recent study in this field has confirmed that such social media information can be used for several crisis response tasks.As a result,dealing with crisis-related tweets and predicting informative tweets is critical for improving crisis response and reducing human and financial loss.Even though the availability of social media Information is huge,open,and free,there are substantial limitations in making sense of this social media data due to its tremendous volume,diversity,velocity,value,and variability.During the crisis event,people post a huge number of informative and non-informative tweets.Affected people and individuals,infrastructure damage,availability,and resource requirements are all included in informative tweets.Therefore,Informative tweets provide useful and helpful information.In contrast,non-informative tweets cannot provide helpful and useful information related to either humanitarian organizations or victims.Therefore,it is very important to build a system that accurately identifies informative tweets.However,it is very challenging to build an accurate predictive model to identify informative Tweets in crisis situations because of the short length of a tweet,and lack of sufficient context.In addition,Crisis related Tweets and regular ones can be hard to identify because of word ambiguity.By considering these challenges in this thesis,we present a deep learning model for identifying Informative tweets from regular tweets during crisis events.The model we propose in this thesis is called Bidirectional Encoder with convolutional layer and Bidirectional short-term memory(BEC-Bi LSTM).The proposed model contains three layers the first layer is the BERT layer which is used to generate contextual word embeddings from Tweets,the second layer is a convolutional layer which is used for local feature extraction,and the last layer is the Bi LSTM layer used for memory to link the extracted context features.Additionally,in this thesis,we propose a knowledge graph for displaying related insights and generating real-time visual information on Informative tweets.To check our model and compare it to its peers,we performed an experiment on the Crisislex-T6 dataset.The results clearly indicate that the proposed BEC-Bi LSTM outperforms its peers in terms of accuracy and F1-scores.