Multilingual Video Corpus Retrieval (mVCR) aims to localize the most relevant videos in a large collection of untrimmed bilingual instructional videos to a given natural language query. Existing methods often fall into the trap of learning spurious relationships as shortcuts. This retrieval bias mainly stems from two aspects: (1) Only a single query statement for each training sample. (2) The noise from irrelevant video features. To address these issues, we propose a Multi-Level QUery-Video Augmentation (MQuA) to facilitate the cross-modal interactions between query and video. Specifically, MQuA consists of Multilingual Query Paraphrase Generation (MQPG) and Multi-Level Video Moment Refinement (MVMR). MQPG prompts large language models (LLMs) to generate diverse paraphrases of queries while preserving their primary semantics, thereby guiding the retrieval model to align the core semantics of queries with video moments. MVMR uncovers the fine-grained video moments features that are highly related to the query. This module further helps the model to learn genuine relevance while eliminating spurious correlations due to noise. Extensive experimental results show that the proposed method achieves state-of-the-art (SOTA) performance compared to other competing methods, and we rank first in the NLPCC-2024-Shared-Task-2 mVCR competition.