Span-Based Joint Entity and Relation Extraction Method

In response to the low accuracy of relationship extraction triplets in the current field of natural language processing, this paper proposes an improved model based on span for joint entity and relation extraction. Firstly, after receiving a text corpus, sentence and word representations are obtaine...

Full description

Saved in:
Bibliographic Details
Published in2024 4th Asia-Pacific Conference on Communications Technology and Computer Science (ACCTCS) pp. 123 - 127
Main Authors Du, Meiyu, Zhao, Yahui, Jin, Guozhe, Ren, Yiping, Cui, Rongyi, Yin, Fei
Format Conference Proceeding
LanguageEnglish
Published IEEE 24.02.2024
Subjects
Accuracy
Communications technology
Computational modeling
Computer science
entity classification
Information filters
Natural language processing
relation extraction
span-based
SpERT model
Training
Online AccessGet full text
DOI10.1109/ACCTCS61748.2024.00029

Cover

More Information
Summary:In response to the low accuracy of relationship extraction triplets in the current field of natural language processing, this paper proposes an improved model based on span for joint entity and relation extraction. Firstly, after receiving a text corpus, sentence and word representations are obtained through BERT preprocessing, and candidate entities are obtained through the span classifier; Rate different entities and filter them through span filters to obtain the real entities. Secondly, we need to concatenate entity pairs and their context into a relation classifier, add an additional linear layer to enter the bilinear layer classifier and then predict each predefined relation category through an activation function. Finally, by setting a confidence threshold to filter out the relations between entity pairs, a relation triplet is obtained, which extracts valuable hidden information from the text to support the implementation and improvement of other tasks. After training the proposed model on a predefined dataset, it was validated on the CoNLL04 dataset. The experimental results showed that the improved method proposed in this paper outperformed the baseline model, achieving an F1 value of 86.93% for entity recognition and 71.65% for relationship extraction.
DOI:10.1109/ACCTCS61748.2024.00029