text2vec, Text to Vector.
文本向量表征工具,把文本转化为向量矩阵,是文本进行计算机处理的第一步。
text2vec实现了Word2Vec、RankBM25、BERT、Sentence-BERT、CoSENT等多种文本表征、文本相似度计算模型,并在文本语义匹配(相似度计算)任务上比较了各模型的效果。
Guide
- Word2Vec:通过腾讯AI Lab开源的大规模高质量中文词向量数据(800万中文词轻量版) (文件名:light_Tencent_AILab_ChineseEmbedding.bin 密码: tawe)实现词向量检索,本项目实现了句子(词向量求平均)的word2vec向量表示
- SBERT(Sentence-BERT):权衡性能和效率的句向量表示模型,训练时通过有监督训练上层分类函数,文本匹配预测时直接句子向量做余弦,本项目基于PyTorch复现了Sentence-BERT模型的训练和预测
- CoSENT(Cosine Sentence):CoSENT模型提出了一种排序的损失函数,使训练过程更贴近预测,模型收敛速度和效果比Sentence-BERT更好,本项目基于PyTorch实现了CoSENT模型的训练和预测
- 英文匹配数据集的评测结果:
Arch | Backbone | Model Name | English-STS-B |
---|---|---|---|
GloVe | glove | Avg_word_embeddings_glove_6B_300d | 61.77 |
BERT | bert-base-uncased | BERT-base-cls | 20.29 |
BERT | bert-base-uncased | BERT-base-first_last_avg | 59.04 |
BERT | bert-base-uncased | BERT-base-first_last_avg-whiten(NLI) | 63.65 |
SBERT | sentence-transformers/bert-base-nli-mean-tokens | SBERT-base-nli-cls | 73.65 |
SBERT | sentence-transformers/bert-base-nli-mean-tokens | SBERT-base-nli-first_last_avg | 77.96 |
SBERT | xlm-roberta-base | paraphrase-multilingual-MiniLM-L12-v2 | 84.42 |
CoSENT | bert-base-uncased | CoSENT-base-first_last_avg | 69.93 |
CoSENT | sentence-transformers/bert-base-nli-mean-tokens | CoSENT-base-nli-first_last_avg | 79.68 |
- 中文匹配数据集的评测结果:
Arch | Backbone | Model Name | ATEC | BQ | LCQMC | PAWSX | STS-B | Avg | QPS |
---|---|---|---|---|---|---|---|---|---|
CoSENT | hfl/chinese-macbert-base | CoSENT-macbert-base | 50.39 | 72.93 | 79.17 | 60.86 | 80.51 | 68.77 | 3008 |
CoSENT | Langboat/mengzi-bert-base | CoSENT-mengzi-base | 50.52 | 72.27 | 78.69 | 12.89 | 80.15 | 58.90 | 2502 |
CoSENT | bert-base-chinese | CoSENT-bert-base | 49.74 | 72.38 | 78.69 | 60.00 | 80.14 | 68.19 | 2653 |
SBERT | bert-base-chinese | SBERT-bert-base | 46.36 | 70.36 | 78.72 | 46.86 | 66.41 | 61.74 | 3365 |
SBERT | hfl/chinese-macbert-base | SBERT-macbert-base | 47.28 | 68.63 | 79.42 | 55.59 | 64.82 | 63.15 | 2948 |
CoSENT | hfl/chinese-roberta-wwm-ext | CoSENT-roberta-ext | 50.81 | 71.45 | 79.31 | 61.56 | 81.13 | 68.85 | - |
SBERT | hfl/chinese-roberta-wwm-ext | SBERT-roberta-ext | 48.29 | 69.99 | 79.22 | 44.10 | 72.42 | 62.80 | - |
- 本项目release模型的中文匹配评测结果:
Arch | Backbone | Model Name | ATEC | BQ | LCQMC | PAWSX | STS-B | Avg | QPS |
---|---|---|---|---|---|---|---|---|---|
Word2Vec | word2vec | w2v-light-tencent-chinese | 20.00 | 31.49 | 59.46 | 2.57 | 55.78 | 33.86 | 23769 |
SBERT | xlm-roberta-base | paraphrase-multilingual-MiniLM-L12-v2 | 18.42 | 38.52 | 63.96 | 10.14 | 78.90 | 41.99 | 3138 |
CoSENT | hfl/chinese-macbert-base | shibing624/text2vec-base-chinese | 31.93 | 42.67 | 70.16 | 17.21 | 79.30 | 48.25 | 3008 |
CoSENT | hfl/chinese-lert-large | GanymedeNil/text2vec-large-chinese | - | - | - | - | - | - | - |
说明:
- 结果值均使用spearman系数
- 结果均只用该数据集的train训练,在test上评估得到的表现,没用外部数据
- shibing624/text2vec-base-chinese模型,是用CoSENT方法训练,基于MacBERT在中文STS-B数据训练得到,并在中文STS-B测试集评估达到SOTA,运行examples/training_sup_text_matching_model.py代码可复现结果,模型文件已经上传到huggingface的模型库shibing624/text2vec-base-chinese,中文语义匹配任务推荐使用
SBERT-macbert-base
模型,是用SBERT方法训练,运行examples/training_sup_text_matching_model.py代码复现结果paraphrase-multilingual-MiniLM-L12-v2
模型名称是sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2,是用SBERT训练,是paraphrase-MiniLM-L12-v2
模型的多语言版本,支持中文、英文等w2v-light-tencent-chinese
是腾讯词向量的Word2Vec模型,CPU加载使用,适用于中文字面匹配任务和缺少数据的冷启动情况- 各预训练模型均可以通过transformers调用,如MacBERT模型:
--model_name hfl/chinese-macbert-base
或者roberta模型:--model_name uer/roberta-medium-wwm-chinese-cluecorpussmall
- 中文匹配数据集下载链接见下方
- 中文匹配任务实验表明,pooling最优是
first_last_avg
,即 SentenceModel 的EncoderType.FIRST_LAST_AVG
,其与EncoderType.MEAN
的方法在预测效果上差异很小 - QPS的GPU测试环境是Tesla V100,显存32GB
Official Demo: https://www.mulanai.com/product/short_text_sim/
HuggingFace Demo: https://huggingface.co/spaces/shibing624/text2vec
run example: examples/gradio_demo.py to see the demo:
python examples/gradio_demo.py
pip install torch # conda install pytorch
pip install -U text2vec
or
pip install torch # conda install pytorch
pip install -r requirements.txt
git clone https://github.com/shibing624/text2vec.git
cd text2vec
pip install --no-deps .
基于pretrained model
计算文本向量:
>>> from text2vec import SentenceModel
>>> m = SentenceModel()
>>> m.encode("如何更换花呗绑定银行卡")
Embedding shape: (768,)
example: examples/computing_embeddings_demo.py
import sys
sys.path.append('..')
from text2vec import SentenceModel
from text2vec import Word2Vec
def compute_emb(model):
# Embed a list of sentences
sentences = [
'卡',
'银行卡',
'如何更换花呗绑定银行卡',
'花呗更改绑定银行卡',
'This framework generates embeddings for each input sentence',
'Sentences are passed as a list of string.',
'The quick brown fox jumps over the lazy dog.'
]
sentence_embeddings = model.encode(sentences)
print(type(sentence_embeddings), sentence_embeddings.shape)
# The result is a list of sentence embeddings as numpy arrays
for sentence, embedding in zip(sentences, sentence_embeddings):
print("Sentence:", sentence)
print("Embedding shape:", embedding.shape)
print("Embedding head:", embedding[:10])
print()
if __name__ == "__main__":
# 中文句向量模型(CoSENT),中文语义匹配任务推荐,支持fine-tune继续训练
t2v_model = SentenceModel("shibing624/text2vec-base-chinese")
compute_emb(t2v_model)
# 支持多语言的句向量模型(Sentence-BERT),英文语义匹配任务推荐,支持fine-tune继续训练
sbert_model = SentenceModel("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2")
compute_emb(sbert_model)
# 中文词向量模型(word2vec),中文字面匹配任务和冷启动适用
w2v_model = Word2Vec("w2v-light-tencent-chinese")
compute_emb(w2v_model)
output:
<class 'numpy.ndarray'> (7, 768)
Sentence: 卡
Embedding shape: (768,)
Sentence: 银行卡
Embedding shape: (768,)
...
- 返回值
embeddings
是numpy.ndarray
类型,shape为(sentences_size, model_embedding_size)
,三个模型任选一种即可,推荐用第一个。 shibing624/text2vec-base-chinese
模型是CoSENT方法在中文STS-B数据集训练得到的,模型已经上传到huggingface的 模型库shibing624/text2vec-base-chinese, 是text2vec.SentenceModel
指定的默认模型,可以通过上面示例调用,或者如下所示用transformers库调用, 模型自动下载到本机路径:~/.cache/huggingface/transformers
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
模型是Sentence-BERT的多语言句向量模型, 适用于释义(paraphrase)识别,文本匹配,通过text2vec.SentenceModel
和sentence-transformers库都可以调用该模型w2v-light-tencent-chinese
是通过gensim加载的Word2Vec模型,使用腾讯词向量Tencent_AILab_ChineseEmbedding.tar.gz
计算各字词的词向量,句子向量通过单词词 向量取平均值得到,模型自动下载到本机路径:~/.text2vec/datasets/light_Tencent_AILab_ChineseEmbedding.bin
Without text2vec, you can use the model like this:
First, you pass your input through the transformer model, then you have to apply the right pooling-operation on-top of the contextualized word embeddings.
example: examples/use_origin_transformers_demo.py
import os
import torch
from transformers import AutoTokenizer, AutoModel
os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
# Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
token_embeddings = model_output[0] # First element of model_output contains all token embeddings
input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('shibing624/text2vec-base-chinese')
model = AutoModel.from_pretrained('shibing624/text2vec-base-chinese')
sentences = ['如何更换花呗绑定银行卡', '花呗更改绑定银行卡']
# Tokenize sentences
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
# Compute token embeddings
with torch.no_grad():
model_output = model(**encoded_input)
# Perform pooling. In this case, max pooling.
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
print("Sentence embeddings:")
print(sentence_embeddings)
sentence-transformers is a popular library to compute dense vector representations for sentences.
Install sentence-transformers:
pip install -U sentence-transformers
Then load model and predict:
from sentence_transformers import SentenceTransformer
m = SentenceTransformer("shibing624/text2vec-base-chinese")
sentences = ['如何更换花呗绑定银行卡', '花呗更改绑定银行卡']
sentence_embeddings = m.encode(sentences)
print("Sentence embeddings:")
print(sentence_embeddings)
提供两种Word2Vec
词向量,任选一个:
- 轻量版腾讯词向量 百度云盘-密码:tawe 或 谷歌云盘,二进制文件,111M,是简化后的高频143613个词,每个词向量还是200维(跟原版一样),运行程序,自动下载到
~/.text2vec/datasets/light_Tencent_AILab_ChineseEmbedding.bin
- 腾讯词向量-官方全量, 6.78G放到:
~/.text2vec/datasets/Tencent_AILab_ChineseEmbedding.txt
,腾讯词向量主页:https://ai.tencent.com/ailab/nlp/zh/index.html 词向量下载地址:https://ai.tencent.com/ailab/nlp/en/download.html 更多查看腾讯词向量介绍-wiki
example: examples/semantic_text_similarity_demo.py
import sys
sys.path.append('..')
from text2vec import Similarity
# Two lists of sentences
sentences1 = ['如何更换花呗绑定银行卡',
'The cat sits outside',
'A man is playing guitar',
'The new movie is awesome']
sentences2 = ['花呗更改绑定银行卡',
'The dog plays in the garden',
'A woman watches TV',
'The new movie is so great']
sim_model = Similarity()
for i in range(len(sentences1)):
for j in range(len(sentences2)):
score = sim_model.get_score(sentences1[i], sentences2[j])
print("{} \t\t {} \t\t Score: {:.4f}".format(sentences1[i], sentences2[j], score))
output:
如何更换花呗绑定银行卡 花呗更改绑定银行卡 Score: 0.9477
如何更换花呗绑定银行卡 The dog plays in the garden Score: -0.1748
如何更换花呗绑定银行卡 A woman watches TV Score: -0.0839
如何更换花呗绑定银行卡 The new movie is so great Score: -0.0044
The cat sits outside 花呗更改绑定银行卡 Score: -0.0097
The cat sits outside The dog plays in the garden Score: 0.1908
The cat sits outside A woman watches TV Score: -0.0203
The cat sits outside The new movie is so great Score: 0.0302
A man is playing guitar 花呗更改绑定银行卡 Score: -0.0010
A man is playing guitar The dog plays in the garden Score: 0.1062
A man is playing guitar A woman watches TV Score: 0.0055
A man is playing guitar The new movie is so great Score: 0.0097
The new movie is awesome 花呗更改绑定银行卡 Score: 0.0302
The new movie is awesome The dog plays in the garden Score: -0.0160
The new movie is awesome A woman watches TV Score: 0.1321
The new movie is awesome The new movie is so great Score: 0.9591
句子余弦相似度值
score
范围是[-1, 1],值越大越相似。
一般在文档候选集中找与query最相似的文本,常用于QA场景的问句相似匹配、文本相似检索等任务。
example: examples/semantic_search_demo.py
import sys
sys.path.append('..')
from text2vec import SentenceModel, cos_sim, semantic_search
embedder = SentenceModel()
# Corpus with example sentences
corpus = [
'花呗更改绑定银行卡',
'我什么时候开通了花呗',
'A man is eating food.',
'A man is eating a piece of bread.',
'The girl is carrying a baby.',
'A man is riding a horse.',
'A woman is playing violin.',
'Two men pushed carts through the woods.',
'A man is riding a white horse on an enclosed ground.',
'A monkey is playing drums.',
'A cheetah is running behind its prey.'
]
corpus_embeddings = embedder.encode(corpus)
# Query sentences:
queries = [
'如何更换花呗绑定银行卡',
'A man is eating pasta.',
'Someone in a gorilla costume is playing a set of drums.',
'A cheetah chases prey on across a field.']
for query in queries:
query_embedding = embedder.encode(query)
hits = semantic_search(query_embedding, corpus_embeddings, top_k=5)
print("\n\n======================\n\n")
print("Query:", query)
print("\nTop 5 most similar sentences in corpus:")
hits = hits[0] # Get the hits for the first query
for hit in hits:
print(corpus[hit['corpus_id']], "(Score: {:.4f})".format(hit['score']))
output:
Query: 如何更换花呗绑定银行卡
Top 5 most similar sentences in corpus:
花呗更改绑定银行卡 (Score: 0.9477)
我什么时候开通了花呗 (Score: 0.3635)
A man is eating food. (Score: 0.0321)
A man is riding a horse. (Score: 0.0228)
Two men pushed carts through the woods. (Score: 0.0090)
======================
Query: A man is eating pasta.
Top 5 most similar sentences in corpus:
A man is eating food. (Score: 0.6734)
A man is eating a piece of bread. (Score: 0.4269)
A man is riding a horse. (Score: 0.2086)
A man is riding a white horse on an enclosed ground. (Score: 0.1020)
A cheetah is running behind its prey. (Score: 0.0566)
======================
Query: Someone in a gorilla costume is playing a set of drums.
Top 5 most similar sentences in corpus:
A monkey is playing drums. (Score: 0.8167)
A cheetah is running behind its prey. (Score: 0.2720)
A woman is playing violin. (Score: 0.1721)
A man is riding a horse. (Score: 0.1291)
A man is riding a white horse on an enclosed ground. (Score: 0.1213)
======================
Query: A cheetah chases prey on across a field.
Top 5 most similar sentences in corpus:
A cheetah is running behind its prey. (Score: 0.9147)
A monkey is playing drums. (Score: 0.2655)
A man is riding a horse. (Score: 0.1933)
A man is riding a white horse on an enclosed ground. (Score: 0.1733)
A man is eating food. (Score: 0.0329)
similarities库[推荐]
文本相似度计算和文本匹配搜索任务,推荐使用 similarities库 ,兼容本项目release的 Word2vec、SBERT、Cosent类语义匹配模型,还支持字面维度相似度计算、匹配搜索算法,支持文本、图像。
安装:
pip install -U similarities
句子相似度计算:
from similarities import Similarity
m = Similarity()
r = m.similarity('如何更换花呗绑定银行卡', '花呗更改绑定银行卡')
print(f"similarity score: {float(r)}") # similarity score: 0.855146050453186
CoSENT(Cosine Sentence)文本匹配模型,在Sentence-BERT上改进了CosineRankLoss的句向量方案
Network structure:
Training:
Inference:
训练和预测CoSENT模型:
- 在中文STS-B数据集训练和评估
CoSENT
模型
example: examples/training_sup_text_matching_model.py
cd examples
python training_sup_text_matching_model.py --model_arch cosent --do_train --do_predict --num_epochs 10 --model_name hfl/chinese-macbert-base --output_dir ./outputs/STS-B-cosent
- 在蚂蚁金融匹配数据集ATEC上训练和评估
CoSENT
模型
支持这些中文匹配数据集的使用:'ATEC', 'STS-B', 'BQ', 'LCQMC', 'PAWSX',具体参考HuggingFace datasets https://huggingface.co/datasets/shibing624/nli_zh
python training_sup_text_matching_model.py --task_name ATEC --model_arch cosent --do_train --do_predict --num_epochs 10 --model_name hfl/chinese-macbert-base --output_dir ./outputs/ATEC-cosent
- 在自有中文数据集上训练模型
example: examples/training_sup_text_matching_model_selfdata.py
python training_sup_text_matching_model_selfdata.py --do_train --do_predict
- 在英文STS-B数据集训练和评估
CoSENT
模型
example: examples/training_sup_text_matching_model_en.py
cd examples
python training_sup_text_matching_model_en.py --model_arch cosent --do_train --do_predict --num_epochs 10 --model_name bert-base-uncased --output_dir ./outputs/STS-B-en-cosent
- 在英文NLI数据集训练
CoSENT
模型,在STS-B测试集评估效果
example: examples/training_unsup_text_matching_model_en.py
cd examples
python training_unsup_text_matching_model_en.py --model_arch cosent --do_train --do_predict --num_epochs 10 --model_name bert-base-uncased --output_dir ./outputs/STS-B-en-unsup-cosent
Sentence-BERT文本匹配模型,表征式句向量表示方案
Network structure:
Training:
Inference:
- 在中文STS-B数据集训练和评估
SBERT
模型
example: examples/training_sup_text_matching_model.py
cd examples
python training_sup_text_matching_model.py --model_arch sentencebert --do_train --do_predict --num_epochs 10 --model_name hfl/chinese-macbert-base --output_dir ./outputs/STS-B-sbert
- 在英文STS-B数据集训练和评估
SBERT
模型
example: examples/training_sup_text_matching_model_en.py
cd examples
python training_sup_text_matching_model_en.py --model_arch sentencebert --do_train --do_predict --num_epochs 10 --model_name bert-base-uncased --output_dir ./outputs/STS-B-en-sbert
- 在英文NLI数据集训练
SBERT
模型,在STS-B测试集评估效果
example: examples/training_unsup_text_matching_model_en.py
cd examples
python training_unsup_text_matching_model_en.py --model_arch sentencebert --do_train --do_predict --num_epochs 10 --model_name bert-base-uncased --output_dir ./outputs/STS-B-en-unsup-sbert
BERT文本匹配模型,原生BERT匹配网络结构,交互式句向量匹配模型
Network structure:
Training and inference:
训练脚本同上examples/training_sup_text_matching_model.py。
由于text2vec训练的模型可以使用sentence-transformers库加载,此处复用其模型蒸馏方法distillation。
- 模型降维,参考dimensionality_reduction.py使用PCA对模型输出embedding降维,可减少milvus等向量检索数据库的存储压力,还能轻微提升模型效果。
- 模型蒸馏,参考model_distillation.py使用蒸馏方法,将Teacher大模型蒸馏到更少layers层数的student模型中,在权衡效果的情况下,可大幅提升模型预测速度。
提供两种部署模型,搭建服务的方法: 1)基于Jina搭建gRPC服务【推荐】;2)基于FastAPI搭建原生Http服务。
采用C/S模式搭建高性能服务,支持docker云原生,gRPC/HTTP/WebSocket,支持多个模型同时预测,GPU多卡处理。
-
安装:
pip install jina
-
启动服务:
example: examples/jina_server_demo.py
from jina import Flow
port = 50001
f = Flow(port=port).add(
uses='jinahub://Text2vecEncoder',
uses_with={'model_name': 'shibing624/text2vec-base-chinese'}
)
with f:
# backend server forever
f.block()
该模型预测方法(executor)已经上传到JinaHub,里面包括docker、k8s部署方法。
- 调用服务:
from jina import Client
from docarray import Document, DocumentArray
port = 50001
c = Client(port=port)
data = ['如何更换花呗绑定银行卡',
'花呗更改绑定银行卡']
print("data:", data)
print('data embs:')
r = c.post('/', inputs=DocumentArray([Document(text='如何更换花呗绑定银行卡'), Document(text='花呗更改绑定银行卡')]))
print(r.embeddings)
批量调用方法见example: examples/jina_client_demo.py
-
安装:
pip install fastapi uvicorn
-
启动服务:
example: examples/fastapi_server_demo.py
cd examples
python fastapi_server_demo.py
- 调用服务:
curl -X 'GET' \
'http://0.0.0.0:8001/emb?q=hello' \
-H 'accept: application/json'
中文语义匹配数据集已经上传到huggingface datasets https://huggingface.co/datasets/shibing624/nli_zh
数据集使用示例:
pip install datasets
from datasets import load_dataset
dataset = load_dataset("shibing624/nli_zh", "STS-B") # ATEC or BQ or LCQMC or PAWSX or STS-B
print(dataset)
print(dataset['test'][0])
output:
DatasetDict({
train: Dataset({
features: ['sentence1', 'sentence2', 'label'],
num_rows: 5231
})
validation: Dataset({
features: ['sentence1', 'sentence2', 'label'],
num_rows: 1458
})
test: Dataset({
features: ['sentence1', 'sentence2', 'label'],
num_rows: 1361
})
})
{'sentence1': '一个女孩在给她的头发做发型。', 'sentence2': '一个女孩在梳头。', 'label': 2}
常见中文语义匹配数据集,包含ATEC、BQ、 LCQMC、PAWSX、STS-B共5个任务。 可以从数据集对应的链接自行下载,也可以从百度网盘(提取码:qkt6)下载。 其中senteval_cn目录是评测数据集汇总,senteval_cn.zip是senteval目录的打包,两者下其一就好。
文本向量方法介绍
文本向量表示咋做?文本匹配任务用哪个模型效果好?
许多NLP任务的成功离不开训练优质有效的文本表示向量。特别是文本语义匹配(Semantic Textual Similarity,如paraphrase检测、QA的问题对匹配)、文本向量检索(Dense Text Retrieval)等任务。
- 基于TF-IDF、BM25、Jaccord、SimHash、LDA等算法抽取两个文本的词汇、主题等层面的特征,然后使用机器学习模型(LR, xgboost)训练分类模型
- 优点:可解释性较好
- 缺点:依赖人工寻找特征,泛化能力一般,而且由于特征数量的限制,模型的效果比较一般
代表模型:
- BM25
BM25算法,通过候选句子的字段对qurey字段的覆盖程度来计算两者间的匹配得分,得分越高的候选项与query的匹配度更好,主要解决词汇层面的相似度问题。
- 基于表征的匹配方式,初始阶段对两个文本各自单独处理,通过深层的神经网络进行编码(encode),得到文本的表征(embedding),再对两个表征进行相似度计算的函数得到两个文本的相似度
- 优点:基于BERT的模型通过有监督的Fine-tune在文本表征和文本匹配任务取得了不错的性能
- 缺点:BERT自身导出的句向量(不经过Fine-tune,对所有词向量求平均)质量较低,甚至比不上Glove的结果,因而难以反映出两个句子的语义相似度
主要原因是:
1.BERT对所有的句子都倾向于编码到一个较小的空间区域内,这使得大多数的句子对都具有较高的相似度分数,即使是那些语义上完全无关的句子对。
2.BERT句向量表示的聚集现象和句子中的高频词有关。具体来说,当通过平均词向量的方式计算句向量时,那些高频词的词向量将会主导句向量,使之难以体现其原本的语义。当计算句向量时去除若干高频词时,聚集现象可以在一定程度上得到缓解,但表征能力会下降。
代表模型:
- DSSM(2013)
- CDSSM(2014)
- ARC I(2014)
- Siamese Network(2016)
- InferSent(2017)
- BERT(2018)
- Sentence-BERT(2019)
- BERT-flow(2020)
- SimCSE(2021)
- ConSERT(2021)
- CoSENT(2022)
由于2018年BERT模型在NLP界带来了翻天覆地的变化,此处不讨论和比较2018年之前的模型(如果有兴趣了解的同学,可以参考中科院开源的MatchZoo 和MatchZoo-py)。
所以,本项目主要调研以下比原生BERT更优、适合文本匹配的向量表示模型:Sentence-BERT(2019)、BERT-flow(2020)、SimCSE(2021)、CoSENT(2022)。
- 基于交互的匹配方式,则认为在最后阶段才计算文本的相似度会过于依赖文本表征的质量,同时也会丢失基础的文本特征(比如词法、句法等),所以提出尽可能早的对文本特征进行交互,捕获更基础的特征,最后在高层基于这些基础匹配特征计算匹配分数
- 优点:基于交互的匹配模型端到端处理,效果好
- 缺点:这类模型(Cross-Encoder)的输入要求是两个句子,输出的是句子对的相似度值,模型不会产生句子向量表示(sentence embedding),我们也无法把单个句子输入给模型。因此,对于需要文本向量表示的任务来说,这类模型并不实用
代表模型:
- ARC II(2014)
- MV-LSTM(2015)
- MatchPyramid(2016)
- DRMM(2016)
- Conv-KNRM(2018)
- RE2(2019)
- Keyword-BERT(2020)
Cross-Encoder适用于向量检索精排。
- Issue(建议):
- 邮件我:xuming: [email protected]
- 微信我:加我微信号:xuming624, 备注:姓名-公司-NLP 进NLP交流群。
如果你在研究中使用了text2vec,请按如下格式引用:
APA:
Xu, M. Text2vec: Text to vector toolkit (Version 1.1.2) [Computer software]. https://github.com/shibing624/text2vec
BibTeX:
@misc{Text2vec,
author = {Xu, Ming},
title = {Text2vec: Text to vector toolkit},
year = {2022},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/shibing624/text2vec}},
}
授权协议为 The Apache License 2.0,可免费用做商业用途。请在产品说明中附加text2vec的链接和授权协议。
项目代码还很粗糙,如果大家对代码有所改进,欢迎提交回本项目,在提交之前,注意以下两点:
- 在
tests
添加相应的单元测试 - 使用
python -m pytest -v
来运行所有单元测试,确保所有单测都是通过的
之后即可提交PR。
- 将句子表示为向量(上):无监督句子表示学习(sentence embedding)
- 将句子表示为向量(下):无监督句子表示学习(sentence embedding)
- A Simple but Tough-to-Beat Baseline for Sentence Embeddings[Sanjeev Arora and Yingyu Liang and Tengyu Ma, 2017]
- 四种计算文本相似度的方法对比[Yves Peirsman]
- Improvements to BM25 and Language Models Examined
- CoSENT:比Sentence-BERT更有效的句向量方案
- 谈谈文本匹配和多轮检索
- Sentence-transformers