Jina AI「テキスト埋め込みによる語順の捕捉の失敗とその修正方法」

import requests
from google.colab import userdata
import numpy as np


def cosine_similarity(vec_list):
    vec1 = np.array(vec_list[0])
    vec2 = np.array(vec_list[1])
    
    dot_product = np.dot(vec1, vec2)
    norm_vec1 = np.linalg.norm(vec1)
    norm_vec2 = np.linalg.norm(vec2)
    
    if norm_vec1 == 0 or norm_vec2 == 0:
        raise ValueError("ゼロベクトルが含まれているため、コサイン類似度は計算できません。")
    
    return dot_product / (norm_vec1 * norm_vec2)


def get_embedding_pair(sentence_list):
    url = 'https://api.jina.ai/v1/embeddings'
    JINA_API_KEY = userdata.get('JINA_API_KEY')
    headers = {
        'Content-Type': 'application/json',
        'Authorization': f'Bearer {JINA_API_KEY}'
    }

    data = {
        "model": "jina-embeddings-v3",
        "task": "text-matching",
        "late_chunking": False,
        "dimensions": 1024,
        "embedding_type": "float",
        "input": sentence_list
    }

    response = requests.post(url, json=data, headers=headers)
    return [d["embedding"] for d in response.json()["data"]]

sentence_pairs_en = {
    "direction": [
        "She flew from Paris to Tokyo.",
        "She drove from Tokyo to Paris.",
    ],
    "temporal": [
        "She ate dinner before watching the movie.",
        "She watched the movie before eating dinner.",
    ],
    "causal": [
        "The rising temperature melted the snow.",
        "The melting snow cooled the temperature.",
    ],
    "comparative": [
        "Coffee tastes better than tea.",
        "Tea tastes better than coffee.",

    ],
    "negation": [
        "He is standing by the table.",
        "He is standing far from the table.",
    ],
}

for k, sentences in sentence_pairs_en.items():
    print(f"#### {k} ####")
    print("\nsentences:")
    print("\n".join([f"- {s}" for s in sentences]))
    print("\nscore:", cosine_similarity(get_embedding_pair(sentences)))
    print()

#### direction ####

sentences:
- She flew from Paris to Tokyo.
- She drove from Tokyo to Paris.

score: 0.8810929055764013

#### temporal ####

sentences:
- She ate dinner before watching the movie.
- She watched the movie before eating dinner.

score: 0.9581018576771294

#### causal ####

sentences:
- The rising temperature melted the snow.
- The melting snow cooled the temperature.

score: 0.7663670043382221

#### comparative ####

sentences:
- Coffee tastes better than tea.
- Tea tastes better than coffee.

score: 0.9538634701296309

#### negation ####

sentences:
- He is standing by the table.
- He is standing far from the table.

score: 0.8089181733708365

sentence_pairs_ja = {
    "direction": [
        "彼女はパリから東京まで行った",
        "彼女は東京からパリまで行った",
    ],
    "temporal": [
        "彼女は映画を見る前に夕食を食べた",
        "彼女は夕食を食べる前に映画を見た",
    ],
    "causal": [
        "気温が上がって雪が溶けた",
        "雪が溶けて気温が下がった",
    ],
    "comparative": [
        "コーヒーは紅茶よりおいしい",
        "紅茶はコーヒーよりおいしい",
    ],
    "negation": [
        "彼はテーブルのそばに立っている",
        "彼はテーブルから離れたところに立っています",
    ],
}

for k, sentences in sentence_pairs_ja.items():
    print(f"#### {k} ####")
    print("\nsentences:")
    print("\n".join([f"- {s}" for s in sentences]))
    print("\nscore:", cosine_similarity(get_embedding_pair(sentences)))
    print()

#### direction ####

sentences:
- 彼女はパリから東京まで行った
- 彼女は東京からパリまで行った

score: 0.9849495181142561

#### temporal ####

sentences:
- 彼女は映画を見る前に夕食を食べた
- 彼女は夕食を食べる前に映画を見た

score: 0.9429784007932432

#### causal ####

sentences:
- 気温が上がって雪が溶けた
- 雪が溶けて気温が下がった

score: 0.8876550504181947

#### comparative ####

sentences:
- コーヒーは紅茶よりおいしい
- 紅茶はコーヒーよりおいしい

score: 0.9245611137311809

#### negation ####

sentences:
- 彼はテーブルのそばに立っている
- 彼はテーブルから離れたところに立っています

score: 0.8468101146022733

sentence_pairs_negation = {
    "negation_en": [
        "He is standing by the table.",
        "He is not standing by the table.",
    ],
    "negation_ja": [
        "彼はテーブルのそばに立っている",
        "彼はテーブルのそばに立っていない",
    ],
}

for k, sentences in sentence_pairs_negation.items():
    print(f"#### {k} ####")
    print("\nsentences:")
    print("\n".join([f"- {s}" for s in sentences]))
    print("\nscore:", cosine_similarity(get_embedding_pair(sentences)))
    print()

#### negation_en ####

sentences:
- He is standing by the table.
- He is not standing by the table.

score: 0.8280262605748225

#### negation_ja ####

sentences:
- 彼はテーブルのそばに立っている
- 彼はテーブルのそばに立っていない

score: 0.8206575665424007