Implementations of encryption methods generally provide a low level of security against brute-force attacks that try every possible key. The Honey Encryption (HE) method was created to overcome these problems. However, the latest development of HE implemented on two-pair conversations results in conversations that have no correlation between one chat and the other when the ciphertext is decrypted using the wrong key (decoy message). This study aims to improve the correlation or naturalness between chats in HE decoy messages. By implementing KeyBERT and Electra's Mask Language Model as seed space components, the correlation between chats in decoy messages can be improved.