【论文笔记】SDCL: Self-Distillation Contrastive Learning for Chinese Spell Checking

本文介绍: 论文提出了一种token–level的自蒸馏对比学习(self–dist illa t ion contrastive learning)方法。作者并没有直接使用BERT的输出作为token embedding，而是使用点乘的方式融合了BERT的输出和word embeddings。

论文地址：https://ar x i v.org/pdf/2210.17168.pdf

论文提出了一种token–level的自蒸馏对比学习(s elf–di s tillat ion contrastive learning)方法。

在这里插入图片描述

传统方法使用BERT后，会对confus ion chars进行聚类，但使用作者提出的方法，会让其变得分布更均匀。

作者提取特征的方式：① 先用MacBERT得到hidden state s，然后用word embedding和hidden state s进行点乘。写成公式为：

import torch
import torch.nn as nn
from transformers import BertTokenizerFast, BertForMaskedLM
import torch.nn.functional as F


class SDCLModel(nn.Module):

    def __init__(self):
        super(SDCLModel, self).__init__()
        self.tokenizer = BertTokenizerFast.from_pretrained('hfl/chinese-macbert-base')
        self.model = BertForMaskedLM.from_pretrained('hfl/chinese-macbert-base')

        self.alpha = 1
        self.beta = 0.5
        self.temperature = 0.9

    def forward(self, inputs, targets=None):
    	"""
    	inputs: 为tokenizer对原文本编码后的输入，包括input_ids, attention_mask等
    	targets：与inputs相同，只不过是对目标文本编码后的结果。
    	"""
        if targets is not None:
        	# 提取labels的input_ids
            text_labels = targets['input_ids'].clone()
            text_labels[text_labels == 0] = -100  # -100计算损失时会忽略
        else:
            text_labels = None
		
        word_embeddings = self.model.bert.embeddings.word_embeddings(inputs['input_ids'])
        hidden_states = self.model.bert(**inputs).last_hidden_state
        logits = self.model.cls(hidden_states * word_embeddings)

        if targets:
            loss = F.cross_entropy(logits.view(logits.shape[0] * logits.shape[1], logits.shape[2]), text_labels.view(-1))
        else:
            loss = 0.

        return logits, hidden_states, loss

    def extract_outputs(self, outputs):
        logits, _, _ = outputs
        return logits.argmax(-1)

    def compute_loss(self, outputs, targets, inputs, detect_targets, *args, **kwargs):
        logits_x, hidden_states_x, loss_x = outputs
        logits_y, hidden_states_y, loss_y = self.forward(targets, targets)

        # FIXME
        anchor_samples = hidden_states_x[detect_targets.bool()]
        positive_samples = hidden_states_y[detect_targets.bool()]
        negative_samples = hidden_states_x[~detect_targets.bool() &amp; inputs['attention_mask'].bool()]

        # 错字和对应正确的字计算余弦相似度
        positive_sim = F.cosine_similarity(anchor_samples, positive_samples)
        # 错字与所有batch内的所有其他字计算余弦相似度
        # （FIXME，这里与原论文不一致，原论文说的是与当前句子的其他字计算，但我除了for循环，不知道该怎么写）
        negative_sim = F.cosine_similarity(anchor_samples.unsqueeze(1), negative_samples.unsqueeze(0), dim=-1)

        sims = torch.concat([positive_sim.unsqueeze(1), negative_sim], dim=1) / self.temperature
        sim_labels = torch.zeros(sims.shape[0]).long().to(self.args.device)

        loss_c = F.cross_entropy(sims, sim_labels)

        self.loss_c = float(loss_c)  # 记录一下

        return loss_x + self.alpha * loss_y + self.beta * loss_c

    def get_optimizer(self):
        return torch.optim.AdamW(self.parameters(), lr=7e-5)

    def predict(self, src):
        src = ' '.join(src.replace(" ", ""))
        inputs = self.tokenizer(src, return_tensors='pt').to(self.args.device)
        outputs = self.forward(inputs)
        outputs = self.extract_outputs(outputs)[0][1:-1]
        return self.tokenizer.decode(outputs).replace(' ', '')