Words interacting in a text may be compared, to a certain extent, to molecules interacting and building “complexes”, i.e. multiwords, named entities, or longer-range semantic or syntactic associations. We will call them “k-itemsets” in the sequel, k being their interaction level. We have shown (Cadot 06) than an adequately built subset of these k-itemsets is enough for describing the entirety of the relations at work in a corpus, whatever the level k of these relations. Experimental assessment: we have shown, on a subset of 120,000 abstracts of Web of Science database in the domain of genomics that a small proportion of these itemsets suffices for discriminating with a measurable precision, fifty sub-categories of genomic research. These ones are issued from an unsupervised categorization process involving the whole 230,000 1-itemsets, i.e. individuals words.