Evaluating the quality of natural language processing (NLP) systems is a complex task due to the nuances of language, making it challenging to determine when a text correctly responds to a given input. Developing NLP systems in pipelines increases the difficulty of measuring outcomes, as individual nodes' prediction quality can be evaluated separately from the entire system's performance. To address this challenge, quantitative evaluation methods such as accuracy or F1 scores are used, but these metrics have limitations, particularly when testing semantic NLP systems. Qualitative evaluation through real user feedback is essential to capture the complexities of human language and provide valuable insights for improving the system. Furthermore, a data-centric approach, focusing on the quality of training data rather than model architecture, is crucial for achieving accurate results and can lead to significant improvements in existing NLP systems.