BIDINLTK/dev/grammar/dependencyParsing.py

import torch
import torch.nn as nn
import torch.optim as optim

# 定义状态类
class DependencyParserState:
    def __init__(self, sentence):
        self.sentence = sentence
        self.buffer = list(range(len(sentence)))  # 待分析的词索引列表
        self.stack = []  # 用于构建依存树的栈
        self.heads = [-1] * len(sentence)  # 词的依存头词索引列表
        self.transitions = []  # 转移动作序列

    def is_terminal(self):
        return len(self.buffer) == 0 and len(self.stack) == 1

# 定义转移动作类
class TransitionSystem:
    def __init__(self):
        self.actions = ['SHIFT', 'LEFT', 'RIGHT']

    def legal_actions(self, state):
        actions = []
        if len(state.buffer) > 0:
            actions.append('SHIFT')
        if len(state.stack) >= 2:
            actions.append('LEFT')
        if len(state.stack) >= 2:
            actions.append('RIGHT')
        return actions

import torch.nn.functional as F

class DependencyParser(nn.Module):
    def __init__(self, input_dim, hidden_dim, output_dim):
        super(DependencyParser, self).__init__()
        self.embedding = nn.Embedding(input_dim, hidden_dim)
        self.linear = nn.Linear(hidden_dim * 2, output_dim)

    def forward(self, state):
        if len(state.stack) < 2:
            return torch.tensor([0, 0, 1])  # 不执行转移
        head_word = state.stack[-1]
        modifier_word = state.stack[-2]
        head_embed = self.embedding(torch.tensor(head_word))
        modifier_embed = self.embedding(torch.tensor(modifier_word))
        combined = torch.cat((head_embed, modifier_embed), dim=0)
        output = self.linear(combined)
        return F.softmax(output, dim=0)

def train_model(train_data, model, transition_system, optimizer, criterion, num_epochs):
    for epoch in range(num_epochs):
        total_loss = 0.0
        for sentence in train_data:
            state = DependencyParserState(sentence)
            while not state.is_terminal():
                legal_actions = transition_system.legal_actions(state)
                action_scores = model(state)
                action_index = torch.argmax(action_scores).item()
                action = transition_system.actions[action_index]
                print(action,",",legal_actions)
                if action in legal_actions:
                    if action == 'SHIFT':
                        state.stack.append(state.buffer.pop(0))
                    elif action == 'LEFT':
                        state.heads[state.stack[-2]] = state.stack[-1]
                        state.stack.pop(-2)
                    elif action == 'RIGHT':
                        state.heads[state.stack[-1]] = state.stack[-2]
                        state.stack.pop()
                    state.transitions.append(action_index)
                else:
                    state.transitions.append(action_index)
                    print("====")
                    break
            if len(state.transitions)>0:
                optimizer.zero_grad()
                print(action_scores.view(1, -1))
                print("===")
                print(torch.tensor(state.transitions))
                loss = criterion(action_scores.view(1, -1), torch.tensor(state.transitions))
                loss.backward()
                optimizer.step()
                total_loss += loss.item()
        print(f'Epoch {epoch+1}/{num_epochs}, Loss: {total_loss/len(train_data)}')


# 示例数据
train_data = [[('I', 'PRP'), ('love', 'VBP'), ('natural', 'JJ'), ('language', 'NN'), ('processing', 'NN')],
              [('Dependency', 'NN'), ('syntax', 'NN'), ('analysis', 'NN'), ('is', 'VBZ'), ('important', 'JJ')],
              [('The', 'DT'), ('cat', 'NN'), ('chased', 'VBD'), ('the', 'DT'), ('mouse', 'NN')]]

# 超参数
input_dim = 10  # 输入维度（词汇量大小）
hidden_dim = 5  # 隐藏层维度
output_dim = 3  # 输出维度（转移动作数量）
num_epochs = 10
learning_rate = 0.01

# 初始化模型、优化器和损失函数
model = DependencyParser(input_dim, hidden_dim, output_dim)
optimizer = optim.SGD(model.parameters(), lr=learning_rate)
criterion = nn.CrossEntropyLoss()

# 训练模型
transition_system = TransitionSystem()
train_model(train_data, model, transition_system, optimizer, criterion, num_epochs)