Child-Sum (N2E2N)Tree-LSTMs: An interactive Child-Sum Tree-LSTMs to extract biomedical event
LSTM has been presented to overcome the problem of the gradient vanishing and explosion. Tree-LSTM could improve the parallel speed of LSTM, and incorporate relevant information from dependency or syntax trees. Tree-LSTM can update gate and memory vectors from the multiple sub-units. Learning edge f...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2024-12-01
|
| Series: | Systems and Soft Computing |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2772941924000048 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | LSTM has been presented to overcome the problem of the gradient vanishing and explosion. Tree-LSTM could improve the parallel speed of LSTM, and incorporate relevant information from dependency or syntax trees. Tree-LSTM can update gate and memory vectors from the multiple sub-units. Learning edge features can strengthen the expression ability of graph neural networks. However, the original Child-Sum Tree-LSTMs ignores edge features during aggregating the sub-nodes hidden states. To enhance node representation, we propose an interaction mechanism that can alternately updating nodes and edges vectors, thus the model can learn the richer nodes vectors. The interaction mechanism attaches the node embedding to its connected link at the first stage. Next, it superimposes the updated edge into the parent node once more. Repeat the above steps from bottom to top. We present five strategies during the alternant renewal process. Meanwhile, we adopt one constituent parser and one dependency parser to produce the diversified formats, and compare their performances in the experiment result. The proposed model achieves better performance than baseline methods on the BioNLP’09 and MLEE corpuses. The experimental results show that the simple event results are almost identical for each parser. But for complex events, Stanford Parser is better than MaltParser because of more frequent interactive behaviors. The different parsing formats have different results, and CoNLL’2008 Dependencies show competitive and superior performance for each parser. |
|---|---|
| ISSN: | 2772-9419 |