Cross-Cultural Marketing Adaptation — 多语言 CAM 嵌入驱动的跨文化营销适配

"""
Cross-Cultural Marketing Adaptation via CAM-Inspired Embeddings
模拟 Class-Aware Masking 对比学习的核心效果
场景：10 个母婴产品 × 4 语言（中英泰越），跨语言语义匹配矩阵

关键设计：
- 无 CAM（语言捷径）：嵌入空间由大语言偏置主导，产品区分依赖语言标记而非语义
- 有 CAM：压制语言偏置，嵌入空间由语义核心主导，跨语言对齐更准确
"""
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

np.random.seed(42)

# ─── 1. 多语言产品嵌入模拟 ────────────────────────────────────────────────
PRODUCTS = [
    "breast_pump", "baby_formula", "diaper", "baby_stroller",
    "baby_carrier", "teething_toy", "bottle_sterilizer", "nursing_pad",
    "baby_monitor", "swaddle_blanket"
]
LANGUAGES = ["en", "zh", "th", "vi"]
n_products = len(PRODUCTS)
n_langs = len(LANGUAGES)
embed_dim = 16  # 低维使产品间区分度更难，更能体现捷径影响

def make_semantic_core(n_products, dim, seed=0):
    """每个产品的语义核心（跨语言共享的真实语义方向）"""
    rng = np.random.RandomState(seed)
    vecs = rng.randn(n_products, dim)
    # 归一化确保各产品语义方向不同
    return vecs / (np.linalg.norm(vecs, axis=1, keepdims=True) + 1e-8)

def build_no_cam_embeddings(semantic_core, lang_bias_scale=3.0, noise_scale=0.3):
    """
    无 CAM 嵌入：叠加强语言偏置（模拟语言捷径）
    同语言样本的嵌入被同一方向的大偏置拉近，跨语言对齐被破坏
    """
    embs = []
    rng = np.random.RandomState(99)
    for lang_idx in range(n_langs):
        # 每种语言有独特的强偏置方向
        lang_bias = rng.randn(dim) * lang_bias_scale
        noise = rng.randn(n_products, dim) * noise_scale
        e = semantic_core + noise + lang_bias  # 语言偏置远大于产品语义差异
        embs.append(e)
    return np.array(embs)

def build_cam_embeddings(semantic_core, lang_bias_scale=0.1, noise_scale=0.2):
    """
    CAM 嵌入：压制语言偏置（CAM 掩码强制跨语言对比，消除语言特征捷径）
    嵌入空间由语义核心主导
    """
    embs = []
    rng = np.random.RandomState(99)
    for lang_idx in range(n_langs):
        lang_bias = rng.randn(dim) * lang_bias_scale  # 偏置被压制
        noise = rng.randn(n_products, dim) * noise_scale
        e = semantic_core + noise + lang_bias
        embs.append(e)
    return np.array(embs)