ベクトル空間の探究

OpenAI API で学ぶ埋め込み入門

Emmanuel Pire

Senior Software Engineer, DataCamp

例：見出しを埋め込む

articles = [
    {"headline": "Economic Growth Continues Amid Global Uncertainty", "topic": "Business"},
    {"headline": "Interest rates fall to historic lows", "topic": "Business"},
    {"headline": "Scientists Make Breakthrough Discovery in Renewable Energy", "topic": "Science"},
    {"headline": "India Successfully Lands Near Moon's South Pole", "topic": "Science"},
    {"headline": "New Particle Discovered at CERN", "topic": "Science"},
    {"headline": "Tech Company Launches Innovative Product to Improve Online Accessibility", "topic": "Tech"},
    {"headline": "Tech Giant Buys 49% Stake In AI Startup", "topic": "Tech"},
    {"headline": "New Social Media Platform Has Everyone Talking!", "topic": "Tech"},
    {"headline": "The Blues get promoted on the final day of the season!", "topic": "Sport"},
    {"headline": "1.5 Billion Tune-in to the World Cup Final", "topic": "Sport"}
]

例：見出しの埋め込み

A list of dictionaries containing the headlines and their embeddings.

複数入力の埋め込み

headline_text = [article['headline'] for article in articles]
headline_text

["Economic Growth Continues Amid Global Uncertainty",
 ...,
 "1.5 Billion Tune-in to the World Cup Final"]

response = client.embeddings.create(
  model="text-embedding-3-small",
  input=headline_text
)
response_dict = response.model_dump()

バッチ処理を行うと、複数のAPI呼び出しを使用するよりも効率的

[...]

'data': [
    {
      "embedding": [-0.017142612487077713, ..., -0.0012911480152979493],
      "index": 0,
      "object": "embedding"
    },
    {
      "embedding": [-0.032995883375406265, ..., -0.0028605300467461348],
      "index": 1,
      "object": "embedding"
    },
    ...
  ]

[...]

複数入力の埋め込み

articles = [
    {"headline": "Economic Growth Continues Amid Global Uncertainty", "topic": "Business"},
     ...
]

for i, article in enumerate(articles):

    article['embedding'] = response_dict['data'][i]['embedding']


print(articles[:2])

[{'headline': 'Economic Growth Continues Amid Global Uncertainty',
  'topic': 'Business',
  'embedding': [-0.017142612487077713, ..., -0.0012911480152979493]}
 {'headline': 'Interest rates fall to historic lows',
  'topic': 'Business',
  'embedding': [-0.032995883375406265, ..., -0.0028605300467461348]}]

埋め込みベクトルの長さはどれくらいか？

「不確実な世界でも経済成長は継続」

len(articles[0]['embedding'])

「テック企業がアクセシビリティを向上させる革新的な製品を発表」

len(articles[5]['embedding'])

常に1536個の数値を返す！

次元削減とt-SNE

次元数を減らすためのさまざまな手法
t-SNE （t-distributed Stochastic Neighbor Embedding、t分布型確率的近傍埋め込み）

¹ https://www.datacamp.com/tutorial/introduction-t-sne

t-SNEの実装

from sklearn.manifold import TSNE
import numpy as np


embeddings = [article['embedding'] for article in articles]


tsne = TSNE(n_components=2, perplexity=5)

embeddings_2d = tsne.fit_transform(np.array(embeddings))

n_components: 結果として得られる次元数
perplexity: アルゴリズムで使用され、データ数より小さくなければならない
一部の情報が失われる

¹ https://www.datacamp.com/tutorial/introduction-t-sne

埋め込みの可視化

import matplotlib.pyplot as plt

plt.scatter(embeddings_2d[:, 0], embeddings_2d[:, 1])


topics = [article['topic'] for article in articles]
for i, topic in enumerate(topics):
    plt.annotate(topic, (embeddings_2d[i, 0], embeddings_2d[i, 1]))

plt.show()

埋め込みの可視化

類似した記事はまとめられる！
モデルはセマンティックな意味を捉える

次回: 類似度の計算

A plot of the 2D vector space showing that reviews with the same sentiment and topic are mapped more closely together in the vector space.

練習しましょう！

OpenAI API で学ぶ埋め込み入門

ベクトル空間の探究

例： 見出しを埋め込む

例： 見出しの埋め込み

複数入力の埋め込み

複数入力の埋め込み

埋め込みベクトルの長さはどれくらいか？

次元削減とt-SNE

t-SNEの実装

埋め込みの可視化

埋め込みの可視化

練習しましょう！

例：見出しを埋め込む

例：見出しの埋め込み