ELMo 说明解析及用法
ELMo(Embeddings from Language Models) 的全面解析,包括原理、用法、解决的问题以及代码实现示例:
一. ELMo 简介
ELMo(Embeddings from Language Models)是由 AllenNLP 在 2018 年提出的 上下文相关的词嵌入模型。与传统静态词嵌入(如 Word2Vec、GloVe)不同,ELMo 生成的词向量会随上下文动态变化,解决了多义词和复杂语境下的语义表示问题。
二. ELMo 的核心思想
- 双向语言模型(BiLM): ELMo 通过双向 LSTM 分别建模正向和反向的语言模型,捕捉上下文信息。
- 正向语言模型:根据前文预测当前词。
- 反向语言模型:根据后文预测当前词。
- 多层表示融合: 整合 LSTM 不同层的隐藏状态(浅层捕捉语法,深层捕捉语义),生成动态词向量。
结构图
示意图
三、实现过程
四. ELMo 解决的问题
问题类型 | 传统方法缺陷 | ELMo 的改进 |
---|---|---|
多义词歧义 | Word2Vec 对多义词只有单一表示 | 根据上下文生成不同嵌入(如 "bank" 在金融/河流场景不同) |
复杂语境理解 | 忽略句子结构信息 | 通过双向 LSTM 捕捉前后文依赖关系 |
任务特定特征提取 | 需从头训练模型 | 提供预训练嵌入,支持下游任务微调 |
五. ELMo 的用法
- 安装依赖
#导入
pip install allennlp allennlp-models
- 自定义 ELMo 嵌入提取
from allennlp.commands.elmo import ElmoEmbedder# 加载预训练 ELMo
elmo = ElmoEmbedder()# 提取单句词向量
sentence = ["I", "ate", "an", "apple"]
vectors = elmo.embed_sentence(sentence) # 返回三层 LSTM 的输出(每层 1024 维)
print(vectors.shape) # (3, 4, 1024): 3 层 x 4 词 x 1024 维# 提取批量句子
batch = [["Hello", "world"], ["ELMo", "is", "awesome"]]
batch_vectors = elmo.embed_sentences(batch)
import torch
from allennlp.modules.elmo import Elmo# 配置 ELMo
options_file = "path/to/options.json"
weight_file = "path/to/weights.hdf5"
elmo = Elmo(options_file, weight_file, num_output_representations=1)# 模拟输入
input_ids = torch.randn(2, 10, 50) # 假设已转换为字符 ID
embeddings = elmo(input_ids)["elmo_representations"][0] # (2, 10, 1024)
六. 使用场景 分类、命名实体识别(NER)和语义相似度计算等任务
-
文本分类(Text Classification) 利用 ELMo 的动态词向量增强输入表示,提升分类效果(如情感分析、新闻分类)
from allennlp.modules.elmo import Elmo, batch_to_ids
import torch
import torch.nn as nn# 配置 ELMo
options_file = "https://allennlp.s3.amazonaws.com/models/elmo/2x4096_512_2048cnn/2x4096_512_2048cnn_elmo_options.json"
weight_file = "https://allennlp.s3.amazonaws.com/models/elmo/2x4096_512_2048cnn/2x4096_512_2048cnn_elmo_weights.hdf5"# 定义分类模型
class ELMoTextClassifier(nn.Module):def __init__(self, num_classes):super().__init__()self.elmo = Elmo(options_file, weight_file, num_output_representations=1, dropout=0)self.lstm = nn.LSTM(input_size=1024, hidden_size=256, batch_first=True)self.classifier = nn.Linear(256, num_classes)def forward(self, sentences):# 生成 ELMo 嵌入character_ids = batch_to_ids(sentences) # 将文本转为字符IDelmo_emb = self.elmo(character_ids)["elmo_representations"][0] # (batch, seq_len, 1024)# 通过LSTM和分类器lstm_out, _ = self.lstm(elmo_emb)logits = self.classifier(lstm_out[:, -1, :]) # 取最后时间步return logits# 示例使用
model = ELMoTextClassifier(num_classes=2)
sentences = [["I", "love", "this", "movie"], ["This", "is", "terrible"]]
output = model(sentences)
print(output.shape) # torch.Size([2, 2])
-
命名实体识别(Named Entity Recognition, NER) 利用 ELMo 捕捉上下文敏感的实体边界(如人名、地名)
from allennlp.modules.elmo import Elmo, batch_to_ids
import torch
import torch.nn as nnclass ELMoForNER(nn.Module):def __init__(self, num_tags):super().__init__()self.elmo = Elmo(options_file, weight_file, num_output_representations=1, dropout=0)self.lstm = nn.LSTM(input_size=1024, hidden_size=256, batch_first=True, bidirectional=True)self.classifier = nn.Linear(512, num_tags) # 双向LSTM输出拼接def forward(self, sentences):character_ids = batch_to_ids(sentences)elmo_emb = self.elmo(character_ids)["elmo_representations"][0] # (batch, seq_len, 1024)# 双向LSTMlstm_out, _ = self.lstm(elmo_emb) # (batch, seq_len, 512)# 每个词对应的标签logitstag_logits = self.classifier(lstm_out) # (batch, seq_len, num_tags)return tag_logits# 示例使用
model = ELMoForNER(num_tags=5) # 假设5种实体类型
sentences = [["Apple", "is", "based", "in", "Cupertino"]]
output = model(sentences)
print(output.shape) # torch.Size([1, 5, 5])
- 语义相似度计算(Semantic Similarity)
计算句子对的语义相似度(如问答匹配、 paraphrase 检测)。
代码实现:
from allennlp.modules.elmo import Elmo, batch_to_ids
import torch
import torch.nn.functional as Fdef elmo_sentence_similarity(sentence1, sentence2):# 初始化ELMoelmo = Elmo(options_file, weight_file, num_output_representations=1, dropout=0)# 生成句子嵌入char_ids = batch_to_ids([sentence1, sentence2])embeddings = elmo(char_ids)["elmo_representations"][0] # (2, seq_len, 1024)# 取句子整体嵌入(均值池化)sent1_emb = torch.mean(embeddings[0], dim=0) # (1024,)sent2_emb = torch.mean(embeddings[1], dim=0) # (1024,)# 计算余弦相似度similarity = F.cosine_similarity(sent1_emb.unsqueeze(0), sent2_emb.unsqueeze(0), dim=1)return similarity.item()# 示例使用
sentence1 = ["The", "cat", "sat", "on", "the", "mat"]
sentence2 = ["A", "feline", "is", "sitting", "on", "a", "rug"]
similarity = elmo_sentence_similarity(sentence1, sentence2)
print(f"Similarity: {similarity:.4f}") # 输出范围 [-1, 1]
- 词义消歧(Word Sense Disambiguation)
根据上下文动态区分多义词的不同含义。
代码实现:
from allennlp.modules.elmo import ElmoEmbedderdef disambiguate_word_sense(word, context):elmo = ElmoEmbedder()embeddings = elmo.embed_sentence(context) # (3 layers, seq_len, 1024)# 获取目标词的ELMo嵌入(所有层拼接)word_index = context.index(word)word_embedding = torch.cat([torch.tensor(embeddings[i][word_index]) for i in range(3)], dim=0) # 3072维return word_embedding# 示例:区分 "bank" 的不同含义
context1 = ["He", "went", "to", "the", "bank", "to", "deposit", "money"] # 金融机构
context2 = ["They", "fished", "by", "the", "bank", "of", "the", "river"] # 河岸embedding1 = disambiguate_word_sense("bank", context1)
embedding2 = disambiguate_word_sense("bank", context2)similarity = F.cosine_similarity(embedding1.unsqueeze(0), embedding2.unsqueeze(0), dim=1)
print(f"Similarity between 'bank' senses: {similarity.item():.4f}") # 预期较低(不同含义)