update ee

example/ee/standard/README.md

@@ -12,6 +12,7 @@
   ```bash
   python==3.8
   pip install -r requirements.txt
+  pip install hydra-core==1.3.1 # ignore the conlict with deepke
   ```
 
 ## Download Code
@@ -30,30 +31,39 @@ Follow the instruction [here](./data/DuEE/README.md)
 
 ## Train
 
-Modify the parameters in `./conf/train.yaml`.
+Modify the parameters in `./conf/train.yaml`. Select different dataset by set `data_name`, and change the `model_name_or_path` for different dataset.
 
-- Trigger
+- Trigger (Event Detection or Trigger Classification)
+
+  First train trigger classification model, and predict the trigger of each instance.
+
   Set `task_name` to `trigger`.
-  Select different dataset by set `data_name`.
   Then run the following command:
+
   ```bash
   python run.py
   ```
+
+  The prediction will be conducted after the training, and the result will be in `exp/xx/trigger/xxx/eval_pred.json`.
 
-- Role
-  Here we train the event arguments extraction model with the gold trigger.
+Then train the event arguments extraction model by the gold trigger.
+- Role (Event Arguments Extraction)
+  Then, we train the event arguements extraction models, here we train the event arguments extraction model with the gold trigger.
   Set `task_name` to `role`.
-  Select different dataset by set `data_name`.
   Then run the following command:
+
   ```bash
   python run.py
   ```
 
-## Predict
+## Predict (Event Arguments Extraction)
+
+The trigger prediction has been conducted during training, and the result is in the `output_dir`. Here we predict the event arguments extraction results with pred trigger result.
+
+Modify the parameters in `./conf/predict.yaml`. Set the `model_name_or_path` to the trained role model path, and `do_pipeline_predict=True` to do the pipeline prediction.
 
-The trigger prediction has been conducted during training, and the result is in the `output_dir`.Here we predict the event arguments extraction results with pred trigger result.
-Modify the parameters in `./conf/predict.yaml`.
 Then run the following command:
 ```bash
   python predict.py
-```
+```
+The final result will be in `eval_pred.json` of the role model path folder.

example/ee/standard/README_CN.md

@@ -12,6 +12,7 @@
   ```bash
   python==3.8
   pip install -r requirements.txt
+  pip install hydra-core==1.3.1 # ignore the conlict with deepke
   ```
 
 ## 克隆代码
@@ -31,30 +32,39 @@ cd DeepKE/example/ee/standard
 
 ## 训练
 
-在`./conf/train.yaml`中修改模型参数
+在`./conf/train.yaml`中修改训练参数。可以通过更改`data_name`参数来选择不同的数据集，并将`model_name_or_path`改成对应的模型。
+
+整个事件抽取的训练分为两部分，第一部分是训练触发词分类的模型，第二部分是训练事件角色抽取的模型。
 
 - Trigger 触发词
+
+  首先是完成每个instance的触发词的抽取，
+
   将`task_name`设置为`trigger`。
-  可以通过更改`data_name`参数来选择不同的数据集。
   然后运行下述命令
+
   ```bash
   python run.py
   ```
-
-- 事件角色
-  在这里我们用正确的trigger训练事件元素抽取模型
+
+  在训练后会完成触发词的预测，结果保存在`exp/xx/trigger/xxx/eval_pred.json`。
+
+- 事件元素 (Event Arguments Extraction)
+  然后是训练一个事件角色抽取的模型，在这里我们用正确的trigger训练事件元素抽取模型。
   将`task_name`设置为`role`。
-  可以通过更改`data_name`参数来选择不同的数据集。
   然后运行下述命令
+
   ```bash
   python run.py
   ```
 
-## 预测
+## 预测 (Event Arguments Extraction)
 
-触发词的预测在训练的过程中会完成，预测的结果在`output_dir`中。在这里我们使用预测得到的触发词来抽取事件抽取元素。
-在`./conf/predict.yaml`中修改模型参数。
+触发词的预测在训练的过程中会完成，预测的结果在`output_dir`中。在这里我们使用预测得到的触发词来完成pipeline的事件元素抽取。
+在`./conf/predict.yaml`中修改模型参数。将`model_name_or_path`设置为训练好的事件元素抽取模型的路径，并且设置`do_pipeline_predict=True`来完成pipeline的事件抽取。
 然后运行下述命令
+
 ```bash
   python predict.py
-```
+```
+最后的预测结果会在role模型对应目录的`eval_pred.json`下。

example/ee/standard/conf/predict.yaml

@@ -1,8 +1,8 @@
 defaults:
   - train
 
-data_name: ACE # [ACE, DuEE]
-model_name_or_path: ./exp/ACE/role/bert-base-uncased
+data_name: DuEE # [ACE, DuEE]
+model_name_or_path: ./exp/DuEE/role/bert-base-chinese
 task_name: role # the trigger prediction is done during the training process.
 do_train: False
 do_eval: True
@@ -12,5 +12,5 @@ do_pipeline_predict: True
 overwrite_cache: True
 
 
-dev_trigger_pred_file: ./exp/ACE/trigger/bert-base-uncased/eval_pred.json
-test_trigger_pred_file: ./exp/ACE/trigger/bert-base-uncased/test_pred.json
+dev_trigger_pred_file: ./exp/DuEE/trigger/bert-base-chinese/eval_pred.json # change to your pred file of trigger classification
+test_trigger_pred_file: ./exp/DuEE/trigger/bert-base-chinese/test_pred.json

example/ee/standard/conf/train.yaml

@@ -1,6 +1,6 @@
-data_name: ACE # [ACE, DuEE]
-model_name_or_path: ../../../../models/bert-base-uncased # [bert-base-uncased, bert-base-chinese] english for ace, chinese for duee
-task_name: role # [trigger, role]
+data_name: DuEE # [ACE, DuEE]
+model_name_or_path:  /newdisk1/msy/models/bert-base-chinese # [bert-base-uncased, bert-base-chinese] english for ace, chinese for duee
+task_name: trigger # [trigger, role]
 model_type: bertcrf
 do_train: True
 do_eval: True
@@ -19,16 +19,16 @@ per_gpu_eval_batch_size: 16
 gradient_accumulation_steps: 1
 max_seq_length: 256
 max_grad_norm: 1.0
-num_train_epochs: 10
-max_steps: 5000
+num_train_epochs: 5
+max_steps: 500
 warmup_steps: 0
 logging_steps: 500
 save_steps: 500
 eval_all_checkpoints: False
 no_cuda: False
 n_gpu: 0
 overwrite_output_dir: True
-overwrite_cache: False
+overwrite_cache: True
 seed: 42
 fp16: False
 fp16_opt_level: "01"

example/ee/standard/predict.py

@@ -46,8 +46,8 @@ def main(args):
     args.data_dir = os.path.join(args.cwd, "./data/" + args.data_name + "/" + args.task_name)
     args.tag_path = os.path.join(args.cwd, "./data/" + args.data_name + "/schema")
     args.model_name_or_path = os.path.join(args.cwd, args.model_name_or_path)
-    args.dev_trigger_pred_file = os.path.join(args.cwd, args.dev_trigger_pred_file) if args.do_pipeline_predict else None
-    args.test_trigger_pred_file = os.path.join(args.cwd, args.test_trigger_pred_file) if args.do_pipeline_predict else None
+    args.dev_trigger_pred_file = os.path.join(args.cwd, args.dev_trigger_pred_file) if args.do_pipeline_predict and args.task_name=="role" else None
+    args.test_trigger_pred_file = os.path.join(args.cwd, args.test_trigger_pred_file) if args.do_pipeline_predict and args.task_name=="role" else None
     args.do_predict = True if args.data_name == "ACE" else False
 
     # Setup CUDA, GPU & distributed training
@@ -75,8 +75,10 @@ def main(args):
     args.model_type = args.model_type.lower()
     config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
 
+    config = config_class.from_pretrained(args.model_name_or_path)
     tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path, do_lower_case=args.do_lower_case)
-    model = model_class.from_pretrained(args.model_name_or_path)
+    model = model_class.from_pretrained(args.model_name_or_path, config=config)
+    logger.info(f"label_nums:{config.num_labels}")
     model.to(device)
 
     pad_token_label_id = -100
@@ -91,15 +93,15 @@ def main(args):
 
     raw_path = "/".join(args.data_dir.split("/")[:-1])
     if args.do_eval:
-        if args.dev_trigger_pred_file is not None:
+        if args.task_name=="role" and args.dev_trigger_pred_file is not None:
             processor.process_dev_with_pred_trigger(args, raw_path, "dev_with_pred_trigger.tsv")
             eval_examples = processor.get_examples(os.path.join(args.data_dir, "dev_with_pred_trigger.tsv"), "dev")
         else:
             eval_examples = processor.get_examples(os.path.join(args.data_dir, "dev.tsv"), "dev")
         eval_dataset = load_and_cache_examples(args, eval_examples , tokenizer, labels, pad_token_label_id, mode="dev")
 
     if args.do_predict:
-        if args.test_trigger_pred_file is not None:
+        if args.task_name=="role" and args.test_trigger_pred_file is not None:
             processor.process_test_with_pred_trigger(args, raw_path, "test_with_pred_trigger.tsv")
             test_examples = processor.get_examples(os.path.join(args.data_dir, "test_with_pred_trigger.tsv"), "test")
         else:

example/ee/standard/requirements.txt

@@ -1,11 +1,10 @@
 torch==1.10.0
 transformers==4.26.0
-hydra-core==1.3.1
-tensorboardx==2.4
+tensorboardx==2.5.1
 lxml==4.9.1
 beautifulsoup4==4.9.3
 bs4==0.0.1
 stanza==1.2
 sentencepiece==0.1.95
-ipdb==0.13.9
-deepke
+ipdb==0.13.11
+deepke==2.2.3