NoisyEQA
收藏资源简介:
NoisyEQA数据集由南洋理工大学MARS实验室创建,旨在评估具身问答(EQA)系统在处理含噪声问题时的鲁棒性。该数据集包含500个含噪声的问题,涵盖四种常见的噪声类型:潜在幻觉噪声、记忆噪声、感知噪声和语义噪声。数据集通过自动化框架生成,模拟了真实世界中人类提问时可能出现的错误和偏差。创建过程中,研究人员对噪声进行了细粒度分类,并生成了对应的问题。NoisyEQA数据集主要应用于提升EQA系统在复杂和现实场景中的语言理解和推理能力,旨在解决系统在面对含噪声问题时的准确性和鲁棒性问题。
The NoisyEQA dataset was created by the MARS Lab at Nanyang Technological University, with the objective of evaluating the robustness of embodied question answering (EQA) systems when processing noisy questions. This dataset includes 500 noisy questions spanning four common noise categories: latent hallucination noise, memory noise, perceptual noise, and semantic noise. It was generated via an automated framework that simulates the realistic errors and biases that arise during human questioning in real-world settings. During the dataset construction process, researchers performed fine-grained categorization of the noise types and generated matching questions. The NoisyEQA dataset is primarily utilized to improve the language understanding and reasoning abilities of EQA systems in complex and realistic scenarios, targeting the resolution of accuracy and robustness issues of such systems when confronted with noisy questions.

- 1NoisyEQA: Benchmarking Embodied Question Answering Against Noisy Queries南洋理工大学 · 2024年



