/similarities

Similarities: a toolkit for similarity calculation and semantic search. 语义相似度计算、匹配搜索工具包,支持文本和图像,开箱即用。

Primary LanguagePythonApache License 2.0Apache-2.0

PyPI version Downloads Contributions welcome GitHub contributors License Apache 2.0 python_version GitHub issues Wechat Group

Similarities

Similarities is a toolkit for similarity calculation and semantic search, supports text and image.

similarities:相似度计算、语义匹配搜索工具包。

similarities 实现了多种相似度计算、匹配搜索算法,支持文本、图像,python3开发,pip安装,开箱即用。

Guide

Feature

文本相似度计算(文本匹配)

  • 余弦相似(Cosine Similarity):两向量求余弦
  • 点积(Dot Product):两向量归一化后求内积
  • 汉明距离(Hamming Distance),编辑距离(Levenshtein Distance),欧氏距离(Euclidean Distance),曼哈顿距离(Manhattan Distance)等

语义模型

字面模型

图像相似度计算(图像匹配)

语义模型

特征提取

  • pHash【推荐】, dHash, wHash, aHash
  • SIFT, Scale Invariant Feature Transform(SIFT)
  • SURF, Speeded Up Robust Features(SURF)(doing)

图文相似度计算

匹配搜索

  • SemanticSearch:向量相似检索,使用Cosine Similarty + topk高效计算,比一对一暴力计算快一个数量级

Demo

Compute similarity score Demo: https://huggingface.co/spaces/shibing624/text2vec

Semantic Search Demo: https://huggingface.co/spaces/shibing624/similarities

Evaluation

文本匹配和文本检索

中文文本匹配模型评测结果

Model ATEC BQ LCQMC PAWSX STS-B Avg QPS
Word2Vec 20.00 31.49 59.46 2.57 55.78 33.86 10283
SBERT-multi 18.42 38.52 63.96 10.14 78.90 41.99 2371
Text2vec 31.93 42.67 70.16 17.21 79.30 48.25 2572

结果值使用spearman系数

Model:

  • Cilin
  • Hownet
  • SimHash
  • TFIDF

Install

pip3 install torch # conda install pytorch
pip3 install -U similarities

or

git clone https://github.com/shibing624/similarities.git
cd similarities
python3 setup.py install

Usage

1. 文本语义相似度计算

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

Similarity的默认方法:

Similarity(corpus: Union[List[str], Dict[str, str]] = None, 
           model_name_or_path="shibing624/text2vec-base-chinese",
           max_seq_length=128)

返回值:余弦值score范围是[-1, 1],值越大越相似

corpus表示:搜索用的doc集,仅搜索时需要,输入格式:句子列表List[str]或者{corpus_id: sentence}的Dict[str, str]格式

model_name_or_path表示:模型名称或者模型路径,默认使用中文表征式匹配模型shibing624/text2vec-base-chinese,可以替换为多语言表征模型sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2

max_seq_length表示:输入句子的最大长度,最大为匹配模型支持的最大长度,BERT系列是512

2. 文本语义匹配搜索

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/base_demo.py

import sys

sys.path.append('..')
from similarities import Similarity

# 1.Compute cosine similarity between two sentences.
sentences = ['如何更换花呗绑定银行卡',
             '花呗更改绑定银行卡']
corpus = [
    '花呗更改绑定银行卡',
    '我什么时候开通了花呗',
    '俄罗斯警告乌克兰反对欧盟协议',
    '暴风雨掩埋了东北部;新泽西16英寸的降雪',
    '**情报局局长访问以色列叙利亚会谈',
    '人在巴基斯坦基地的炸弹袭击中丧生',
]
model = Similarity(model_name_or_path="shibing624/text2vec-base-chinese")
print(model)
similarity_score = model.similarity(sentences[0], sentences[1])
print(f"{sentences[0]} vs {sentences[1]}, score: {float(similarity_score):.4f}")

print('-' * 50 + '\n')
# 2.Compute similarity between two list
similarity_scores = model.similarity(sentences, corpus)
print(similarity_scores.numpy())
for i in range(len(sentences)):
    for j in range(len(corpus)):
        print(f"{sentences[i]} vs {corpus[j]}, score: {similarity_scores.numpy()[i][j]:.4f}")

print('-' * 50 + '\n')
# 3.Semantic Search
model.add_corpus(corpus)
res = model.most_similar(queries=sentences, topn=3)
print(res)
for q_id, c in res.items():
    print('query:', sentences[q_id])
    print("search top 3:")
    for corpus_id, s in c.items():
        print(f'\t{model.corpus[corpus_id]}: {s:.4f}')

output:

如何更换花呗绑定银行卡 vs 花呗更改绑定银行卡, score: 0.8551
...

如何更换花呗绑定银行卡 vs 花呗更改绑定银行卡, score: 0.8551
如何更换花呗绑定银行卡 vs 我什么时候开通了花呗, score: 0.7212
如何更换花呗绑定银行卡 vs 俄罗斯警告乌克兰反对欧盟协议, score: 0.1450
如何更换花呗绑定银行卡 vs 暴风雨掩埋了东北部;新泽西16英寸的降雪, score: 0.2167
如何更换花呗绑定银行卡 vs **情报局局长访问以色列叙利亚会谈, score: 0.2517
如何更换花呗绑定银行卡 vs 人在巴基斯坦基地的炸弹袭击中丧生, score: 0.0809
花呗更改绑定银行卡 vs 花呗更改绑定银行卡, score: 1.0000
花呗更改绑定银行卡 vs 我什么时候开通了花呗, score: 0.6807
花呗更改绑定银行卡 vs 俄罗斯警告乌克兰反对欧盟协议, score: 0.1714
花呗更改绑定银行卡 vs 暴风雨掩埋了东北部;新泽西16英寸的降雪, score: 0.2162
花呗更改绑定银行卡 vs **情报局局长访问以色列叙利亚会谈, score: 0.2728
花呗更改绑定银行卡 vs 人在巴基斯坦基地的炸弹袭击中丧生, score: 0.1279

query: 如何更换花呗绑定银行卡
search top 3:
	花呗更改绑定银行卡: 0.8551
	我什么时候开通了花呗: 0.7212
	**情报局局长访问以色列叙利亚会谈: 0.2517

余弦score的值范围[-1, 1],值越大,表示该query与corpus的文本越相似。

英文语义相似度计算和匹配搜索

example: examples/base_english_demo.py

3. 快速近似语义匹配搜索

支持Annoy、Hnswlib的近似语义匹配搜索,常用于百万数据集的匹配搜索任务。

example: examples/fast_sim_demo.py

4. 基于字面的文本相似度计算和匹配搜索

支持同义词词林(Cilin)、知网Hownet、词向量(WordEmbedding)、Tfidf、SimHash、BM25等算法的相似度计算和字面匹配搜索,常用于文本匹配冷启动。

example: examples/literal_sim_demo.py

from similarities.literalsim import SimHashSimilarity, TfidfSimilarity, BM25Similarity, \
    WordEmbeddingSimilarity, CilinSimilarity, HownetSimilarity

text1 = "如何更换花呗绑定银行卡"
text2 = "花呗更改绑定银行卡"

corpus = [
    '花呗更改绑定银行卡',
    '我什么时候开通了花呗',
    '俄罗斯警告乌克兰反对欧盟协议',
    '暴风雨掩埋了东北部;新泽西16英寸的降雪',
    '**情报局局长访问以色列叙利亚会谈',
    '人在巴基斯坦基地的炸弹袭击中丧生',
]

queries = [
    '我的花呗开通了?',
    '乌克兰被俄罗斯警告'
]
m = TfidfSimilarity()
print(text1, text2, ' sim score: ', m.similarity(text1, text2))

m.add_corpus(corpus)
res = m.most_similar(queries, topn=3)
print('sim search: ', res)
for q_id, c in res.items():
    print('query:', queries[q_id])
    print("search top 3:")
    for corpus_id, s in c.items():
        print(f'\t{m.corpus[corpus_id]}: {s:.4f}')

output:

如何更换花呗绑定银行卡 花呗更改绑定银行卡  sim score:  0.8203384355246909

sim search:  {0: {2: 0.9999999403953552, 1: 0.43930041790008545, 0: 0.0}, 1: {0: 0.7380483150482178, 1: 0.0, 2: 0.0}}
query: 我的花呗开通了?
search top 3:
	我什么时候开通了花呗: 1.0000
	花呗更改绑定银行卡: 0.4393
	俄罗斯警告乌克兰反对欧盟协议: 0.0000
...

5. 图像相似度计算和匹配搜索

支持CLIP、pHash、SIFT等算法的图像相似度计算和匹配搜索。

example: examples/image_demo.py

import sys
import glob
from PIL import Image

sys.path.append('..')
from similarities.imagesim import ImageHashSimilarity, SiftSimilarity, ClipSimilarity


def sim_and_search(m):
    print(m)
    # similarity
    sim_scores = m.similarity(imgs1, imgs2)
    print('sim scores: ', sim_scores)
    for (idx, i), j in zip(enumerate(image_fps1), image_fps2):
        s = sim_scores[idx] if isinstance(sim_scores, list) else sim_scores[idx][idx]
        print(f"{i} vs {j}, score: {s:.4f}")
    # search
    m.add_corpus(corpus_imgs)
    queries = imgs1
    res = m.most_similar(queries, topn=3)
    print('sim search: ', res)
    for q_id, c in res.items():
        print('query:', image_fps1[q_id])
        print("search top 3:")
        for corpus_id, s in c.items():
            print(f'\t{m.corpus[corpus_id].filename}: {s:.4f}')
    print('-' * 50 + '\n')

image_fps1 = ['data/image1.png', 'data/image3.png']
image_fps2 = ['data/image12-like-image1.png', 'data/image10.png']
imgs1 = [Image.open(i) for i in image_fps1]
imgs2 = [Image.open(i) for i in image_fps2]
corpus_fps = glob.glob('data/*.jpg') + glob.glob('data/*.png')
corpus_imgs = [Image.open(i) for i in corpus_fps]

# 2. image and image similarity score
sim_and_search(ClipSimilarity())  # the best result
sim_and_search(ImageHashSimilarity(hash_function='phash'))
sim_and_search(SiftSimilarity())

output:

Similarity: ClipSimilarity, matching_model: CLIPModel
sim scores:  tensor([[0.9580, 0.8654],
        [0.6558, 0.6145]])

data/image1.png vs data/image12-like-image1.png, score: 0.9580
data/image3.png vs data/image10.png, score: 0.6145

sim search:  {0: {6: 0.9999999403953552, 0: 0.9579654932022095, 4: 0.9326782822608948}, 1: {8: 0.9999997615814209, 4: 0.6729235649108887, 0: 0.6558331847190857}}

query: data/image1.png
search top 3:
	data/image1.png: 1.0000
	data/image12-like-image1.png: 0.9580
	data/image8-like-image1.png: 0.9327

image_sim

6. 图文互搜

CLIP 模型不仅支持以图搜图,还支持图文互搜:

import sys
import glob
from PIL import Image
sys.path.append('..')
from similarities.imagesim import ImageHashSimilarity, SiftSimilarity, ClipSimilarity

m = ClipSimilarity()
print(m)
# similarity score between text and image
image_fps = ['data/image3.png',  # yellow flower image
             'data/image1.png']  # tiger image
texts = ['a yellow flower', 'a tiger']
imgs = [Image.open(i) for i in image_fps]
sim_scores = m.similarity(imgs, texts)

print('sim scores: ', sim_scores)
for (idx, i), j in zip(enumerate(image_fps), texts):
    s = sim_scores[idx][idx]
    print(f"{i} vs {j}, score: {s:.4f}")

output:

sim scores:  tensor([[0.3220, 0.2409],
        [0.1677, 0.2959]])
data/image3.png vs a yellow flower, score: 0.3220
data/image1.png vs a tiger, score: 0.2959

Contact

  • Issue(建议) :GitHub issues
  • 邮件我:xuming: xuming624@qq.com
  • 微信我: 加我微信号:xuming624, 备注:姓名-公司-NLP 进NLP交流群。

Citation

如果你在研究中使用了similarities,请按如下格式引用:

APA:

Xu, M. Similarities: Compute similarity score for humans (Version 1.0.1) [Computer software]. https://github.com/shibing624/similarities

BibTeX:

@misc{Xu_Similarities_Compute_similarity,
  title={Similarities: similarity calculation and semantic search toolkit},
  author={Xu Ming},
  year={2022},
  howpublished={\url{https://github.com/shibing624/similarities}},
}

License

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similarities的链接和授权协议。

Contribute

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:

  • tests添加相应的单元测试
  • 使用python -m pytest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。

Reference