EAEPrompt

code for paper "Event Argument Extraction with Enriched Prompts"

1.运行代码

（同RAMS原始数据集建议评估方式）：

nohup python3 -u code_1.py --gold_file dev.jsonlines --pred_file 1.json --ontology_file event_role_multiplicities.txt  --do_all --reuse_gold_format > r_1.txt 2>&1 &

2.代码说明

• role-prompt:

3. 损失增强

def dice_loss(target, predictive, ep=1e-8):
    t = 0.0
    for i in range(len(target)):
        intersection = 2 * torch.sum(predictive[i] * target[i]) + ep
        union = torch.sum(predictive[i]) + torch.sum(target[i]) + ep
        t += 1 - intersection / union
    loss = t/len(target)
    return loss

loss = weight1 * (loss_func(out_1.cuda(), label_begin.cuda()) + loss_func(out_2.cuda(), label_end.cuda())) + weight2 * dice_loss(target=label.cuda(), predictive=(out_1.softmax(dim=-1).cuda()+out_2.softmax(dim=-1).cuda())/2)

4.LLM prompt

You are now an expert in event argument extraction for natural language processing tasks. Below is a news document, and your task is to extract the arguments based on the roles:

【document】

Now, you need to extract the arguments for the event type 【event_type】, where the event trigger is 【trigger】. The roles to extract include 【role1】, 【role2】, and 【role3】. Please note the following:

Some roles may not have corresponding arguments; in such cases, return null.
The extracted arguments should consist of one or more consecutive words that can be found in the news document.
The format of your output should be {}.