使用 Multilingual Universal Sentence Encoder 研究跨语言相似度和构建语义搜索引擎

在 TensorFlow.org 上查看 在 Google Colab 中运行 在 GitHub 上查看源代码 下载笔记本 查看 TF Hub 模型

此笔记本演示了如何访问 Multilingual Universal Sentence Encoder 模块，以及如何将它用于跨多种语言的句子相似度研究。本模块是原始 Universal Sentence Encoder 模块的扩展。

此笔记本分为以下两个部分：

  • 第一部分展示了成对语言之间句子的可视化。这是一项学术性较强的练习。
  • 在第二部分中，我们将展示如何从多种语言的 Wikipedia 语料库样本构建语义搜索引擎。

引用

研究论文在使用本 Colab 中探讨的模型时应引用以下内容：

###

Multilingual universal sentence encoder for semantic retrieval

Yinfei Yang, Daniel Cer, Amin Ahmad, Mandy Guo, Jax Law, Noah Constant, Gustavo Hernandez Abrego, Steve Yuan, Chris Tar, Yun-Hsuan Sung, Brian Strope, and Ray Kurzweil. 2019. arXiv preprint arXiv:1907.04307

安装

本部分将对访问 Multilingual Universal Sentence Encoder 模块的环境进行设置，并准备一组英语句子及其翻译。在以下部分中，多语模块将用于计算跨语言相似度。

Setup Environment

Setup common imports and functions

下面是附加的样板代码，我们在其中导入了预训练的 ML 模型，在此笔记本中我们将用它来对文本进行编码。

# The 16-language multilingual module is the default but feel free
# to pick others from the list and compare the results.
module_url = 'https://tfhub.dev/google/universal-sentence-encoder-multilingual/3'

model = hub.load(module_url)

def embed_text(input):
  return model(input)

呈现语言之间的文本相似度

现在有了句子嵌入向量，我们就能呈现不同语言之间的语义相似度。

计算文本嵌入向量

我们首先定义一组同时翻译成各种语言的句子。然后预计算所有句子的嵌入向量。

# Some texts of different lengths in different languages.
arabic_sentences = ['كلب', 'الجراء لطيفة.', 'أستمتع بالمشي لمسافات طويلة على طول الشاطئ مع كلبي.']
chinese_sentences = ['狗', '小狗很好。', '我喜欢和我的狗一起沿着海滩散步。']
english_sentences = ['dog', 'Puppies are nice.', 'I enjoy taking long walks along the beach with my dog.']
french_sentences = ['chien', 'Les chiots sont gentils.', 'J\'aime faire de longues promenades sur la plage avec mon chien.']
german_sentences = ['Hund', 'Welpen sind nett.', 'Ich genieße lange Spaziergänge am Strand entlang mit meinem Hund.']
italian_sentences = ['cane', 'I cuccioli sono carini.', 'Mi piace fare lunghe passeggiate lungo la spiaggia con il mio cane.']
japanese_sentences = ['犬', '子犬はいいです', '私は犬と一緒にビーチを散歩するのが好きです']
korean_sentences = ['개', '강아지가 좋다.', '나는 나의 개와 해변을 따라 길게 산책하는 것을 즐긴다.']
russian_sentences = ['собака', 'Милые щенки.', 'Мне нравится подолгу гулять по пляжу со своей собакой.']
spanish_sentences = ['perro', 'Los cachorros son agradables.', 'Disfruto de dar largos paseos por la playa con mi perro.']

# Multilingual example
multilingual_example = ["Willkommen zu einfachen, aber", "verrassend krachtige", "multilingüe", "compréhension du langage naturel", "модели.", "大家是什么意思" , "보다 중요한", ".اللغة التي يتحدثونها"]
multilingual_example_in_en =  ["Welcome to simple yet", "surprisingly powerful", "multilingual", "natural language understanding", "models.", "What people mean", "matters more than", "the language they speak."]

# Compute embeddings.
ar_result = embed_text(arabic_sentences)
en_result = embed_text(english_sentences)
es_result = embed_text(spanish_sentences)
de_result = embed_text(german_sentences)
fr_result = embed_text(french_sentences)
it_result = embed_text(italian_sentences)
ja_result = embed_text(japanese_sentences)
ko_result = embed_text(korean_sentences)
ru_result = embed_text(russian_sentences)
zh_result = embed_text(chinese_sentences)

multilingual_result = embed_text(multilingual_example)
multilingual_in_en_result = embed_text(multilingual_example_in_en)

可视化相似度

有了文本嵌入向量，我们就可以利用它们的点积来呈现句子在不同语言之间的相似程度。颜色越深，嵌入向量在语义上越相似。

多语言相似度

visualize_similarity(multilingual_in_en_result, multilingual_result,
                     multilingual_example_in_en, multilingual_example,  "Multilingual Universal Sentence Encoder for Semantic Retrieval (Yang et al., 2019)")

英语-阿拉伯语相似度

visualize_similarity(en_result, ar_result, english_sentences, arabic_sentences, 'English-Arabic Similarity')

英语-俄语相似度

visualize_similarity(en_result, ru_result, english_sentences, russian_sentences, 'English-Russian Similarity')

英语-西班牙语相似度

visualize_similarity(en_result, es_result, english_sentences, spanish_sentences, 'English-Spanish Similarity')

英语-意大利语相似度

visualize_similarity(en_result, it_result, english_sentences, italian_sentences, 'English-Italian Similarity')

意大利语-西班牙语相似度

visualize_similarity(it_result, es_result, italian_sentences, spanish_sentences, 'Italian-Spanish Similarity')

英语-中文相似度

visualize_similarity(en_result, zh_result, english_sentences, chinese_sentences, 'English-Chinese Similarity')

英语-韩语相似度

visualize_similarity(en_result, ko_result, english_sentences, korean_sentences, 'English-Korean Similarity')

中文-韩语相似度

visualize_similarity(zh_result, ko_result, chinese_sentences, korean_sentences, 'Chinese-Korean Similarity')

以及更多…

上面的示例可以扩展到英语、阿拉伯语、中文、荷兰语、法语、德语、意大利语、日语、韩语、波兰语、葡萄牙语、俄语、西班牙语、泰语和土耳其语中的任何语言对。编程愉快！

创建多语语义相似度搜索引擎

在前面的示例中，我们可视化了少量句子，而在本部分，我们将使用来自 Wikipedia 语料库的约 200,000 个句子构建一个语义搜索索引。其中，大约一半将使用英语，另一半使用西班牙语，以演示 Universal Sentence Encoder 的多语言功能。

将数据下载到索引

首先，我们将从 News-Commentary 语料库 [1] 中下载多种语言的新闻句子。在不失一般性的前提下，此方法应该也可以用来为其余支持的语言建立索引。

为了加快演示速度，我们将每种语言限制为 1000 个句子。

corpus_metadata = [
    ('ar', 'ar-en.txt.zip', 'News-Commentary.ar-en.ar', 'Arabic'),
    ('zh', 'en-zh.txt.zip', 'News-Commentary.en-zh.zh', 'Chinese'),
    ('en', 'en-es.txt.zip', 'News-Commentary.en-es.en', 'English'),
    ('ru', 'en-ru.txt.zip', 'News-Commentary.en-ru.ru', 'Russian'),
    ('es', 'en-es.txt.zip', 'News-Commentary.en-es.es', 'Spanish'),
]

language_to_sentences = {}
language_to_news_path = {}
for language_code, zip_file, news_file, language_name in corpus_metadata:
  zip_path = tf.keras.utils.get_file(
      fname=zip_file,
      origin='http://opus.nlpl.eu/download.php?f=News-Commentary/v11/moses/' + zip_file,
      extract=True)
  news_path = os.path.join(os.path.dirname(zip_path), news_file)
  language_to_sentences[language_code] = pd.read_csv(news_path, sep='\t', header=None)[0][:1000]
  language_to_news_path[language_code] = news_path

  print('{:,} {} sentences'.format(len(language_to_sentences[language_code]), language_name))

Downloading data from http://opus.nlpl.eu/download.php?f=News-Commentary/v11/moses/ar-en.txt.zip
24715264/24714354 [==============================] - 2s 0us/step
1,000 Arabic sentences
Downloading data from http://opus.nlpl.eu/download.php?f=News-Commentary/v11/moses/en-zh.txt.zip
18104320/18101984 [==============================] - 2s 0us/step
1,000 Chinese sentences
Downloading data from http://opus.nlpl.eu/download.php?f=News-Commentary/v11/moses/en-es.txt.zip
28106752/28106064 [==============================] - 2s 0us/step
1,000 English sentences
Downloading data from http://opus.nlpl.eu/download.php?f=News-Commentary/v11/moses/en-ru.txt.zip
24854528/24849511 [==============================] - 2s 0us/step
1,000 Russian sentences
1,000 Spanish sentences

使用预训练的模型将句子转换为向量

我们分批次计算嵌入向量，使其适合 GPU 的 RAM。

# Takes about 3 minutes

batch_size = 2048
language_to_embeddings = {}
for language_code, zip_file, news_file, language_name in corpus_metadata:
  print('\nComputing {} embeddings'.format(language_name))
  with tqdm(total=len(language_to_sentences[language_code])) as pbar:
    for batch in pd.read_csv(language_to_news_path[language_code], sep='\t',header=None, chunksize=batch_size):
      language_to_embeddings.setdefault(language_code, []).extend(embed_text(batch[0]))
      pbar.update(len(batch))

0%|          | 0/1000 [00:00<?, ?it/s]
Computing Arabic embeddings
83178it [00:30, 2768.60it/s]
  0%|          | 0/1000 [00:00<?, ?it/s]
Computing Chinese embeddings
69206it [00:18, 3664.60it/s]
  0%|          | 0/1000 [00:00<?, ?it/s]
Computing English embeddings
238853it [00:37, 6319.00it/s]
  0%|          | 0/1000 [00:00<?, ?it/s]
Computing Russian embeddings
190092it [00:34, 5589.16it/s]
  0%|          | 0/1000 [00:00<?, ?it/s]
Computing Spanish embeddings
238819it [00:41, 5754.02it/s]

构建语义向量的索引

我们使用 SimpleNeighbors 库（这是 Annoy 库的封装容器）高效地从语料库中查找结果。

%%time

# Takes about 8 minutes

num_index_trees = 40
language_name_to_index = {}
embedding_dimensions = len(list(language_to_embeddings.values())[0][0])
for language_code, zip_file, news_file, language_name in corpus_metadata:
  print('\nAdding {} embeddings to index'.format(language_name))
  index = SimpleNeighbors(embedding_dimensions, metric='dot')

  for i in trange(len(language_to_sentences[language_code])):
    index.add_one(language_to_sentences[language_code][i], language_to_embeddings[language_code][i])

  print('Building {} index with {} trees...'.format(language_name, num_index_trees))
  index.build(n=num_index_trees)
  language_name_to_index[language_name] = index

0%|          | 1/1000 [00:00<02:21,  7.04it/s]
Adding Arabic embeddings to index
100%|██████████| 1000/1000 [02:06<00:00,  7.90it/s]
  0%|          | 1/1000 [00:00<01:53,  8.84it/s]
Building Arabic index with 40 trees...

Adding Chinese embeddings to index
100%|██████████| 1000/1000 [02:05<00:00,  7.99it/s]
  0%|          | 1/1000 [00:00<01:59,  8.39it/s]
Building Chinese index with 40 trees...

Adding English embeddings to index
100%|██████████| 1000/1000 [02:07<00:00,  7.86it/s]
  0%|          | 1/1000 [00:00<02:17,  7.26it/s]
Building English index with 40 trees...

Adding Russian embeddings to index
100%|██████████| 1000/1000 [02:06<00:00,  7.91it/s]
  0%|          | 1/1000 [00:00<02:03,  8.06it/s]
Building Russian index with 40 trees...

Adding Spanish embeddings to index
100%|██████████| 1000/1000 [02:07<00:00,  7.84it/s]
Building Spanish index with 40 trees...
CPU times: user 11min 21s, sys: 2min 14s, total: 13min 35s
Wall time: 10min 33s


%%time

# Takes about 13 minutes

num_index_trees = 60
print('Computing mixed-language index')
combined_index = SimpleNeighbors(embedding_dimensions, metric='dot')
for language_code, zip_file, news_file, language_name in corpus_metadata:
  print('Adding {} embeddings to mixed-language index'.format(language_name))
  for i in trange(len(language_to_sentences[language_code])):
    annotated_sentence = '({}) {}'.format(language_name, language_to_sentences[language_code][i])
    combined_index.add_one(annotated_sentence, language_to_embeddings[language_code][i])

print('Building mixed-language index with {} trees...'.format(num_index_trees))
combined_index.build(n=num_index_trees)

0%|          | 1/1000 [00:00<02:00,  8.29it/s]
Computing mixed-language index
Adding Arabic embeddings to mixed-language index
100%|██████████| 1000/1000 [02:06<00:00,  7.92it/s]
  0%|          | 1/1000 [00:00<02:24,  6.89it/s]
Adding Chinese embeddings to mixed-language index
100%|██████████| 1000/1000 [02:05<00:00,  7.95it/s]
  0%|          | 1/1000 [00:00<02:05,  7.98it/s]
Adding English embeddings to mixed-language index
100%|██████████| 1000/1000 [02:06<00:00,  7.88it/s]
  0%|          | 1/1000 [00:00<02:18,  7.20it/s]
Adding Russian embeddings to mixed-language index
100%|██████████| 1000/1000 [02:04<00:00,  8.03it/s]
  0%|          | 1/1000 [00:00<02:17,  7.28it/s]
Adding Spanish embeddings to mixed-language index
100%|██████████| 1000/1000 [02:06<00:00,  7.90it/s]
Building mixed-language index with 60 trees...
CPU times: user 11min 18s, sys: 2min 13s, total: 13min 32s
Wall time: 10min 30s

验证语义相似度搜索引擎是否有效

我们将在本部分演示以下内容：

  1. 语义搜索功能：从语料库中检索与给定查询语义相似的句子。
  2. 多语言功能：查询语言和索引语言匹配时以多种语言显示
  3. 跨语言功能：使用与索引语料库不同的语言发起查询
  4. 混合语言语料库：以上所有内容均在一个索引中，包含来自所有语言的条目

语义搜索跨语言功能

在本部分中，我们将展示如何检索与一组英语例句相关的句子。待尝试的内容如下：

  • 尝试一些不同的例句
  • 尝试更改返回结果的数量（它们会按照相似度的顺序返回）
  • 通过返回不同语言的结果来尝试跨语言功能（可以使用 Google 翻译将部分结果翻译成您的母语来进行健全性检查）

English sentences similar to: "The stock market fell four points."
['Nobel laureate Amartya Sen attributed the European crisis to four failures – political, economic, social, and intellectual.',
 'Just last December, fellow economists Martin Feldstein and Nouriel Roubini each penned op-eds bravely questioning bullish market sentiment, sensibly pointing out gold’s risks.',
 'His ratings have dipped below 50% for the first time.',
 'As a result, markets were deregulated, making it easier to trade assets that were perceived to be safe, but were in fact not.',
 'Consider the advanced economies.',
 'But the agreement has three major flaws.',
 'This “predetermined equilibrium” thinking – reflected in the view that markets always self-correct – led to policy paralysis until the Great Depression, when John Maynard Keynes’s argument for government intervention to address unemployment and output gaps gained traction.',
 'Officials underestimated tail risks.',
 'Consider a couple of notorious examples.',
 'Stalin was content to settle for an empire in Eastern Europe.']

混合语料库功能

现在，我们将以英语发起查询，但结果将来自任何一种已建立索引的语言。

English sentences similar to: "The stock market fell four points."
['Nobel laureate Amartya Sen attributed the European crisis to four failures – political, economic, social, and intellectual.',
 'It was part of the 1945 consensus.',
 'The end of the East-West ideological divide and the end of absolute faith in markets are historical turning points.',
 'Just last December, fellow economists Martin Feldstein and Nouriel Roubini each penned op-eds bravely questioning bullish market sentiment, sensibly pointing out gold’s risks.',
 'His ratings have dipped below 50% for the first time.',
 'As a result, markets were deregulated, making it easier to trade assets that were perceived to be safe, but were in fact not.',
 'Consider the advanced economies.',
 'Since their articles appeared, the price of gold has moved up still further.',
 'But the agreement has three major flaws.',
 'Gold prices even hit a record-high $1,300 recently.',
 'This “predetermined equilibrium” thinking – reflected in the view that markets always self-correct – led to policy paralysis until the Great Depression, when John Maynard Keynes’s argument for government intervention to address unemployment and output gaps gained traction.',
 'What Failed in 2008?',
 'Officials underestimated tail risks.',
 'Consider a couple of notorious examples.',
 'One of these species, orange roughy, has been caught commercially for only around a quarter-century, but already is being fished to the point of collapse.',
 'Meanwhile, policymakers were lulled into complacency by the widespread acceptance of economic theories such as the “efficient-market hypothesis,” which assumes that investors act rationally and use all available information when making their decisions.',
 'Stalin was content to settle for an empire in Eastern Europe.',
 'Intelligence assets have been redirected.',
 'A new wave of what the economist Joseph Schumpeter famously called “creative destruction” is under way: even as central banks struggle to maintain stability by flooding markets with liquidity, credit to business and households is shrinking.',
 'It all came about in a number of ways.',
 'The UN, like the dream of European unity, was also part of the 1945 consensus.',
 'The End of 1945',
 'The Global Economy’s New Path',
 'But this scenario failed to materialize.',
 'Gold prices are extremely sensitive to global interest-rate movements.',
 'Fukushima has presented the world with a far-reaching, fundamental choice.',
 'It was Japan, the high-tech country par excellence (not the latter-day Soviet Union) that proved unable to take adequate precautions to avert disaster in four reactor blocks.',
 'Some European academics tried to argue that there was no need for US-like fiscal transfers, because any desired degree of risk sharing can, in theory, be achieved through financial markets.',
 '$10,000 Gold?',
 'One answer, of course, is a complete collapse of the US dollar.',
 '1929 or 1989?',
 'The goods we made were what economists call “rival" and “excludible" commodities.',
 'This dream quickly faded when the Cold War divided the world into two hostile blocs. But in some ways the 1945 consensus, in the West, was strengthened by Cold War politics.',
 'The first flaw is that the spending reductions are badly timed: coming as they do when the US economy is weak, they risk triggering another recession.',
 'One successful gold investor recently explained to me that stock prices languished for a more than a decade before the Dow Jones index crossed the 1,000 mark in the early 1980’s.',
 'Eichengreen traces our tepid response to the crisis to the triumph of monetarist economists, the disciples of Milton Friedman, over their Keynesian and Minskyite peers – at least when it comes to interpretations of the causes and consequences of the Great Depression.',
 "However, America's unilateral options are limited.",
 'Once it was dark, a screen was set up and Mark showed home videos from space.',
 'These aspirations were often voiced in the United Nations, founded in 1945.',
 'Then I got distracted for about 40 years.']

尝试您自己的查询：

English sentences similar to: "The stock market fell four points."
['(Chinese) 新兴市场的号角',
 '(English) It was part of the 1945 consensus.',
 '(Russian) Брюссель. Цунами, пронёсшееся по финансовым рынкам, является глобальной катастрофой.',
 '(Arabic) هناك أربعة شروط مسبقة لتحقيق النجاح الأوروبي في أفغانستان:',
 '(Spanish) Su índice de popularidad ha caído por primera vez por debajo del 50 por ciento.',
 '(English) His ratings have dipped below 50% for the first time.',
 '(Russian) Впервые его рейтинг опустился ниже 50%.',
 '(English) As a result, markets were deregulated, making it easier to trade assets that were perceived to be safe, but were in fact not.',
 '(Arabic) وكانت التطورات التي شهدتها سوق العمل أكثر تشجيعا، فهي على النقيض من أسواق الأصول تعكس النتائج وليس التوقعات. وهنا أيضاً كانت الأخبار طيبة. فقد أصبحت سوق العمل أكثر إحكاما، حيث ظلت البطالة عند مستوى 3.5% وكانت نسبة الوظائف إلى الطلبات المقدمة فوق مستوى التعادل.',
 '(Russian) Это было частью консенсуса 1945 года.',
 '(English) Consider the advanced economies.',
 '(English) Since their articles appeared, the price of gold has moved up still further.',
 '(Russian) Тогда они не только смогут накормить свои семьи, но и начать получать рыночную прибыль и откладывать деньги на будущее.',
 '(English) Gold prices even hit a record-high $1,300 recently.',
 '(Chinese) 另一种金融危机',
 '(Russian) Европейская мечта находится в кризисе.',
 '(English) What Failed in 2008?',
 '(Spanish) Pero el acuerdo alcanzado tiene tres grandes defectos.',
 '(English) Officials underestimated tail risks.',
 '(English) Consider a couple of notorious examples.',
 '(Spanish) Los mercados financieros pueden ser frágiles y ofrecen muy poca capacidad de compartir los riesgos relacionados con el ingreso de los trabajadores, que constituye la mayor parte de la renta de cualquier economía avanzada.',
 '(Chinese) 2008年败在何处？',
 '(Spanish) Consideremos las economías avanzadas.',
 '(Spanish) Los bienes producidos se caracterizaron por ser, como señalaron algunos economistas, mercancías “rivales” y “excluyentes”.',
 '(Arabic) إغلاق الفجوة الاستراتيجية في أوروبا',
 '(English) Stalin was content to settle for an empire in Eastern Europe.',
 '(English) Intelligence assets have been redirected.',
 '(Spanish) Hoy, envalentonados por la apreciación continua, algunos están sugiriendo que el oro podría llegar incluso a superar esa cifra.',
 '(Russian) Цены на золото чрезвычайно чувствительны к мировым движениям процентных ставок.',
 '(Russian) Однако у достигнутой договоренности есть три основных недостатка.']

其他主题

多语言

最后，我们建议您尝试使用任何一种支持的语言进行查询：英语、阿拉伯语、中文、荷兰语、法语、德语、意大利语、日语、韩语、波兰语、葡萄牙语、俄语、西班牙语、泰语和土耳其语。

另外，即使我们仅对部分语言建立了索引，您也可以使用任何支持的语言为内容建立索引。

模型变体

我们提供已针对各个方面（如内存、延迟和/或质量）进行了优化的 Universal Sentence Encoder 模型变体。您可以随意试验，找出最合适的一个。

最近邻库

我们使用 Annoy 来高效查找最近邻。请参阅权衡部分，以了解要搜索的树的数量（取决于内存）和项目的数量（取决于延迟）。SimpleNeighbors 仅允许控制树的数量，但对代码进行重构以直接使用 Annoy 应该会比较简单，我们只想为普通用户尽可能简化此代码。

如果 Annoy 不适合您的应用，请试试 FAISS。

祝您顺利构建多语言语义应用！

[1] J. Tiedemann, 2012, Parallel Data, Tools and Interfaces in OPUS. 第八届语言资源与评估国际会议论文集 (LREC 2012)