Global Context Mechanism for Sequence Labeling
- python==3.8.13
- torch==1.13.0
- transformers==4.27.1
- tqdm==4.64.0
- numpy==1.22.4
- Aspect-Based sentiment analysis (ABSA): all datasets for ABSA can be found in (Li, 2019)
- Named Entity Recognition: Conll2003, wnut2017, weibo
- Part of Speech Tagging: Universal Dependencies
| Layers | Rest14 | Rest15 | Rest16 | Laoptop14 | Conll2003 | Wnut2017 | Conll2003 | UD | |
|---|---|---|---|---|---|---|---|---|---|
| BERT | 1E-5 | 1E-5 | 1E-5 | 1E-5 | 1E-5 | 1E-5 | 1E-5 | 1E-5 | 1E-5 |
| BiLSTM | 5E-4 | 1E-3 | 5E-4 | 5E-4 | 1E-3 | 1E-3 | 1E-3 | 1E-3 | 1E-3 |
| context | 1E-3 | 1E-3 | 1E-5 | 1E-5 | 1E-3 | 1E-3 | 1E-3 | 1E-4 | 1E-3 |
| classification | 1E-4 | 1E-4 | 1E-4 | 1E-4 | 1E-4 | 1E-4 | 1E-4 | 1E-4 | 1E-4 |
bert-base-chinese and bert-base-cased is used for English datasets and Chinese datasets respectively. batch size:
- ABSA: Rest14 32, Rest15 16, Rest16 32, Laptop14 16.
- NER: 16 is applied for all datasets.
- POS Tagging: 16 is applied for all datasets.
python main.py --dataset_type absa --dataset_name rest14 --use_tagger True --use_context True
- model_name: pretrained model name. default: bert-base-cased
- cache_dir: the directory to save pretrained model.
- use_tagger: using BiLSTM or not. default: True
- use_context: using context mechanism or not. default: False
- context_mechanism: which context mechanism will be used. default: global
- mode: using pretrained language or not. default: pretrained
- tagger_size: dimension of BiLSTM output. default 600
In case of that you have specific dataset format, making a new reader function which is a parameter to construct the Dataset classes.
Rename the files under each dataset to train.txt, valid.txt and test.txt respectively. the format samples are given under each dataset directory.
| Layers | Rest14 | Rest15 | Rest16 | Laoptop14 | Conll2003 | Wnut2017 | Conll2003 | UD | |
|---|---|---|---|---|---|---|---|---|---|
| BERT | 69.75 | 57.07 | 65.95 | 58.49 | 91.51 | 43.59 | 68.09 | 95.56 | 96.85 |
| BERT-BiLSTM | 73.47 | 61.14 | 71.05 | 61.12 | 91.85 | 46.95 | 68.86 | 95.66 | 95.90 |
| BERT-BiLSTM-context | 73.84 | 63.24 | 71.51 | 62.92 | 91.91 | 48.02 | 69.84 | 95.62 | 97.01 |