Long text matching refers to the process of matching two pieces of text at the document level. Current methods struggle to effectively capture the key information scattered throughout long texts, and are insensitive to the semantic differences of identical words and the semantic similarities of different words. Therefore, this paper proposes the integrating hierarchical key information and semantic difference features for long text matching (HKIDF). Firstly, at the sentence level, we use information entropy to extract key sentences with a large amount of information, and then at the word level, we use an improved TextRank algorithm integrated with Transformer for word-level filtering of key sentences; Secondly, we use the improved Diff Transformer in BERT to learn difference features from the filtered key information; Finally, through a gate mechanism, we selectively fuse the semantic representations of key information and difference features to determine the matching relationship between long texts. Experimental results on the Chinese datasets CNSE and CNSS and the English dataset PAN demonstrate that our model outperforms existing baseline models.