The problem of retrieving a knowledge-from-text base, on a text level, is considered. A solution is given that is based on retrieving a text on a sentence level. It consists of converting a query whose qualification and target attributes are not present in the same sentence, into an equivalent set of subqueries, each of which having all of its attributes present in the same sentence. Logical conditions underlying this decomposition are based on functional and multivalued dependencies between qualification and target attributes of the subqueries. The dependencies are defined in the extended relational model with two types of null values, representing a text. These conditions are proved to provide for answers to the last subquery to coincide with the answers to the original query, i.e., to be correct.