Research On Social Media Oriented Disaster Information Identification

In recent years,social media has become an important source for disaster-related information,but it also contains a large amount of irrelevant or invalid information,which poses a problem of information filtering for emergency response and rescue organizations.Therefore,it is important to identify effective disaster-related information from the massive social media data for disaster situational awareness and relief needs assessment.Most of the existing studies are based on machine learning and deep learning approaches,but there are still some shortcomings.In this thesis the following improvements are proposed to address this problem:(1)This thesis uses pre-trained language models to improve the performance of disaster effective information recognition.Pre-trained language models are able to learn rich language features by pre-training on a large-scale corpus.In this thesis,we adapt different pre-trained language models to the task of disaster valid information recognition by fine-tuning them on different pre-trained language models,and verify their effectiveness on this task.(2)In this thesis,Delta-tuning fine-tuning method is used to reduce the time and resource consumption of fine-tuning pre-trained language models.The traditional Fine-tuning method needs to update all parameters of the pre-trained language model,which consumes a lot of computational resources.Delta-tuning,on the other hand,only needs to update a small number of parameters to achieve similar or even better results than Fine-tuning.In this thesis,we improve two fine-tuning methods,Prefix-tuning and LoRA-tuning,and apply them to different pre-trained language models.(3)In this thesis,we propose a PLMEF(Pretraining Language Model Ensemble Framework),an integrated framework for disaster effective information recognition based on pre-trained language models,to further improve the recognition performance.Since different pre-training language models adopt different training strategies in the pre-training stage and learn different text features,combining them can obtain richer semantic features.In this thesis,three different pre-trained language models,Ro BERTa,BERT and XLNet,are used as base learners using soft voting method as an integration strategy and trained using a delta-tuning based fine-tuning method.The experimental results show that PLMEF outperforms 11 baseline models on the disaster information recognition task.