Accurate disaster entity recognition based on contextual embeddings in self-attentive BiLSTM-CRF

Area of Science:

• Natural Language Processing
• Information Extraction
• Disaster Management

Background:

• Automated extraction of disaster-related named entities is vital for timely crisis information gathering.
• Online news is a key source for disaster information during emergencies and response.
• Effective disaster management relies on accurate and accessible crisis data.

Purpose of the Study:

• To investigate the automatic extraction of disaster-related named entities from online news articles.
• To develop and evaluate a novel contextualized deep Bi-directional LSTM network for NER.
• To compare the proposed model's performance against existing word embedding approaches.

Main Methods:

• Constructed a novel word embedding model inspired by Word2vec for contextual vector representations.
• Utilized a deep Bi-directional LSTM network with self-attention and CRF layers for encoding features.
• Annotated a dataset of 1000 online news articles with 14 crisis-specific entities.

Main Results:

• The proposed NER model achieved sentence-level Precision of 92%, Recall of 91%, Accuracy of 87%, and F1-score of 92% on an independent test set.
• The context-sensitive optimized model outperformed general, non-contextual word embeddings, and BERT models.
• Demonstrated superior performance in extracting disaster-related named entities.

Conclusions:

• The developed context-sensitive NER model offers a significant advancement in automated disaster information extraction.
• This approach enhances the reliability and timeliness of data crucial for disaster management.
• The model's effectiveness highlights the importance of contextual word embeddings in crisis informatics.