This paper proposes a learning-based methodology to deal with Natural Language. Our system tries to acquire a knowledge model from an analyzed text. After that, a new text can be analyzed by using that model. The methodology introduces a unified representation mechanism for every kind of information present in Natural Language treatment (corresponding to Natural Language phases in other systems). In this way, the acquired knowledge model can contain any kind of language information (lexical, syntactic, semantic, conceptual or even world-knowledge). With this approach we are not trying to give an overall solution to Natural Language treatment; but to solve Natural Language problems using any kind of relevant information. We also provide an inference mechanism that puts asside the idea of Natural Language phase.