Several researchers have attempted to confront the problems in quality attributes prediction using AI approaches consisting of knowledge-driven and data-driven techniques. However, due to the lack of a shared training dataset and standardized definition of quality attributes, inaccurate feature extraction may lead to the inconsistency and poor performance of a prediction model. Different from prior works, we have investigated the impact of software domain in quality attributes prediction using deep learning methods with different datasets. From the results, we conclude with two recommendations: (i) the existing secondary dataset such as PROMISE and Concordia are not enough to be used as a ground truth; (ii) a deep learning approach should be supported from the aspects of broader domains in order to capture a variety of natural language requirements. The contribution of this paper is to raise the awareness of identifying the quality attributes in requirements writing and help requirements providers understand what issues to focus. A prototype of requirements annotator is introduced to show how the requirements are processed.