How to optimize this model of semantic matching?

Question

I have a NLP matching task like this:

Input: short text(article title) + long text(article content)

Output: Semantic matching rate(Similarity)

E