In this talk we present a new approach to high dimensionality Machine Learning (ML) in general and to text analysis and knowledge representation in particular. This model yields the cognitive plausibility of the representation at multiple levels. we introduce a novel representation method which is initially based on the different type of closeness between words in text passages (inter-relationship) over the entire corpus. In our multi-level lightweight ontological representation method (TOR-FUSE), corpus entries are being represented based on their contexts and the goal of the learning task simultaneously. The multi-level model will allow data interpretation in a more conceptual space rather than just containing separate words appearing in the text. In all the levels, the model considers the cognitive plausibility of interpretation as one of its bases. It aims to perform the extraction of the knowledge beneficial for the classification task by automatically creation of a lightweight ontological hierarchy of representations. In the last step, the data representation will train a tailored ensemble of learners over a stack of representations in different conceptual granularity. The modeling performs mapping/weighting of the targeted concept of classification over a stack of representations and granular attributes in each level at the same time. (A hierarchical mapping instead of linear mapping over a vector)