Fix graph Laplacian docstring and semantic weight scaling in KernelLanguageEntropy

[Kangwei-g]

Hi team!

I was reviewing the implementation of Kernel Language Entropy (KLE) and noticed a couple of details regarding the semantic graph construction that deviate slightly from the formulas in the original paper. I've made the following corrections:

• Fixed Laplacian formula in docstring: Corrected the comment from L = W - D to the standard definition L = D - W.
• Corrected weight matrix scaling: According to the paper, the semantic graph weight is defined as $W_{ij} = w \cdot NLI'(S_i, S_j) + w \cdot NLI'(S_j, S_i)$. The previous implementation divided this sum by 2. I removed the / 2 to strictly align with the paper's formula.
• Adjusted neutral matrix calculation: Since matrix_entail and matrix_contra are now the full sum (with a maximum possible value of 2 instead of 1), I updated the matrix_neutral calculation to use 2 * np.ones(...) to maintain the correct semantic proportions.

Let me know if you have any questions!