Widespread and intensive applications of large language models (LLMs) have raised ethical and legal concerns regarding undisclosed pre-training data. Previous research has shown the inclusion of problematic content in pre-training data for LLMs, such as copyrighted material, personal information, and even benchmark data, rendering pre-training data detection for LLMs indispensable. Existing methods are primarily built on capturing the discrepancy between seen and unseen texts solely from probability distributions of tokens derived by verbatim reproduction of the target text by the target LLM, while missing the textual difference between the target text and the verbatim reproduced text, which can also reflect the memorization of the target LLM regarding the target text. Moreover, these methods assume that a low token probability implies that the target text is an unseen text; however, relatively smooth probability distributions can also lead to a low token probability. To tackle these problems, this work proposes a novel method to distinguish dual-aspect discrepancies between seen and unseen texts for pre-training data detection (DAD-PDD). Specifically, DAD-PDD performs detection by modeling discrepancies in: 1) token-level scored rank distributions optimized from probability distributions of tokens and 2) sentence-level lexical matching between the target text and its verbatim reproduction. Our experimental evaluation confirms that DAD-PDD outperforms the SOTA methods, increasing the AUC scores on average by 12.67% and 16.79% on the WikiMIA and MIMIR datasets separately across various LLMs, demonstrating stable and effective detection performance. The source code is available at https://github.com/FANXIN-Gmail/DAD-PDD .