Diverse Beam Search（ダイバースビームサーチ）

Beam Search出力例（翻訳）:
1. "The cat sat on the mat."
2. "The cat was sitting on the mat."
3. "The cat sat on the mat quickly."
4. "A cat sat on the mat."
5. "The cat sat upon the mat."
→ 5候補全て類似（sat/sitting/uponのみ違い）

# 概念的な実装
num_beams = 12
num_groups = 4  # 3ビームずつ4グループ
beams_per_group = num_beams // num_groups  # 3

for group_idx in range(num_groups):
    for t in range(max_length):
        # 標準スコア + 差異ペナルティ
        score = model_score(x)
        
        # 前グループの出力との類似度をペナルティ
        if group_idx > 0:
            prev_outputs = [group_beams[g] for g in range(group_idx)]
            diversity_term = hamming_distance_based_penalty(x, prev_outputs)
            score -= lambda_param * diversity_term
        
        select_top_beams(score, k=beams_per_group)

Diverse Beam Search出力例（翻訳）:
Group 1: "The cat sat on the mat."
Group 1: "The cat is sitting on the mat."
Group 1: "A cat sat on the mat."
Group 2: "On the mat, a cat was sitting."
Group 2: "Sitting on the mat was the cat."
Group 2: "A mat was under the sitting cat."
Group 3: "The feline rested upon the rug."
...
→ グループ間で構造・語彙が多様化

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("facebook/bart-large-cnn")
model = AutoModelForSeq2SeqLM.from_pretrained("facebook/bart-large-cnn")

inputs = tokenizer(article_text, return_tensors="pt")

# Diverse Beam Search
outputs = model.generate(
    **inputs,
    num_beams=12,          # 総ビーム数
    num_beam_groups=4,     # グループ数（4グループ×3ビーム）
    diversity_penalty=1.0, # 差異ペナルティの強さ
    num_return_sequences=4, # 返す候補数（グループ数以下）
    max_new_tokens=200,
)

for i, output in enumerate(outputs):
    print(f"候補{i+1}: {tokenizer.decode(output)}")