The Transformer model continues to serve as the standard encoder-decoder architecture in neural machine translation (NMT). It effectively captures the essential features of sentences in the source language through its encoder. However, this representation often encodes source language information in a general manner, over-looking crucial details within the sentence. Therefore, this method may hinder the model from accurately understanding the deep semantics of the source language, ultimately impact the translation quality and may lead to mistranslation. To tackle this issue, this paper presents a novel model for integrating sentence trunk information, aiming to enhance the representational capacity of source language sentences. Specifically, we employed a sentence trunk extraction algorithm that accurately identifies the trunk information within a sentence. The trunk information contains key elements necessary to determine the meaning of the sentence. Subsequently, we smoothly integrate the extracted sentence trunk information into the encoder of the NMT model, thereby enhancing and refining the source language representation. Experimental findings from numerous translation tasks reveal that our pro-posed sentence trunk fusion NMT model considerably enhances the translation performance of NMT models.