Anomaly detection in machine learning is a recurring challenge studied in various domains, which has a blossoming interest and present unique challenges. One issue of existing approaches is that they overlook a crucial aspect: the type of targeted anomaly. Furthermore, experimental protocols often do not differentiate existence of several textual anomalies, limiting the scope of the work. After formalizing two types of anomalies (independent and contextual), we propose a novel approach using robust ensemble autoencoders, which are randomly connected. Our approach successfully detects both simple and complex anomalies. It is noteworthy that our approach excels not only in providing competitive results compared to existing methods but also in handling contextual anomalies.