After the development of rule-based, example-based, and statistical translation research, the advances in neural machine translation have qualitatively refined machine translation and attracted unprecedented attention after the advent of ChatGPT. On the grounds of cognitive metaphor and integrating the framing theory, which features the same cognitive function, this paper explores the fundamental mapping relations between the source frame of human neural networks and the target frame of artificial neural networks, profound connotations, and metaphorical coherence. On that basis, the attention mechanism and training mechanism in neural machine translation are explored by combining the metaphorical reasoning and the generic level and specific level within the scope of metaphor. By enquiring into the metaphorical elements contained in neural machine translation, this paper provides a cognitive perspective for deconstructing neural machine translation.