Multi-label Relation Classification using BERT

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering

Final Year Thesis Oral Defense

Title: "Multi-label Relation Classification using BERT"

by

WANG Yili

Abstract:

Relation classification is an important natural language processing task 
in information extraction area. It provides a link between raw texts and 
structured data which is useful for many other NLP application like 
reading comprehension and question answering. General relation 
classification is defined as, given texts and a pair of entities as input, 
a classifier will output a label indicating the type of relation between 
entity tokens. In our project, we extend traditional relation extraction 
to end-to-end multi-label learning in which we do not rely on the entity 
pairs but directly predict all the relations in the input sentences. We 
also explore the performance of pre-trained language model BERT and 
previous neural networks models on Chinese dataset. We show that by 
adapting BERT structure to multi-label classification, we can achieve high 
performance in terms of exact matching accuracy and F1 score.


Date            : 14 May 2020 (Thursday)

Time            : 11:00 - 11:40

Zoom Meeting    : https://hkust.zoom.us/j/592856162

Advisor         : Prof. LIN Fangzhen

2nd Reader      : Dr. SONG Yangqiu