def compute_similarity()

in src/run_paraphrase.py [0:0]

• 17 lines of code
• 6 McCabe index (conditional complexity)

def compute_similarity(f1, f2, method):
    """Compute similarity between two contextualized representations.

    NOTE: Other methods besides bertscore aren't currently used but we did try
    them out.

    Args:
        f1: Tensor of shape (L_1, d)
        f2: Tensor of shape (L_2, d)
        method: Name of method
    Returns:
        Similarity score
    """
    if method == 'mean-cosine':
        r1 = torch.mean(f1, dim=0)  # d
        r2 = torch.mean(f2, dim=0)  # d
        return r1.dot(r2) / torch.sqrt(r1.dot(r1) * r2.dot(r2))
    elif method == 'bertscore' or method == 'bertscore-minmax':
        r1 = (f1.T / torch.norm(f1, dim=1)).T  # L_1, d
        r2 = (f2.T / torch.norm(f2, dim=1)).T  # L_2, d
        dots = torch.matmul(r1, r2.T)  # L_1, L_2
        if method == 'bertscore':
            s1 = torch.mean(torch.max(dots, dim=0)[0])
            s2 = torch.mean(torch.max(dots, dim=1)[0])
        elif method == 'bertscore-minmax':
            s1 = torch.min(torch.max(dots, dim=0)[0])
            s2 = torch.min(torch.max(dots, dim=1)[0])
        return 2 * s1 * s2 / (s1 + s2)
    raise NotImplementedError