Modeling Consistency Preference via Lexical Chains for Document-level Neural Machine Translation