The relationship between language and thought is often dealt with in terms of categorization, i.e. considering the way languages capture representations into meaningful units. However, in language activity, linguistic units never appear alone: they are always incorporated into, and at the same time operated upon by a process of linearization of thought in an utterance where meaning is constructed through a dynamic process. Formally speaking, morphemes are units which are ordered sequentially to make up an utterance; but semantically, words are not units of thought which can be added together to yield the meaning of the utterance, as the data show (here mostly from French). The complex interaction between two levels of meaning (meaning of the units and meaning of the utterance) sets up a dynamic process which proceeds throughout the speech act. This process is characterized by a constant retroaction of the units upon one other, thus assuming the existence of 'reentry-type' cognitive mechanisms, in language activity. This non-additive way of building up the meaning of the utterance is precisely what allows the dimensional reduction of thought to language. The reason why the construction of meaning is not additive is that first, words have their own meanings and their 'representational depth', and second, the utterance has structural features which shape meaning in non-linear ways.