This dissertation proposes an account for cross-linguistic variation at the lexical level. The proposal stems from a strong version of Hale and Keyser's hypothesis according to which lexicalization processes are syntactically driven. I claim that variation at the lexical level should be analyzed as variation at the sentential level, by means of the unification of the l(exical) and s(entential)-syntax through the theory of ciclycity provided by Phase Theory. The phase is interpreted as a point of access of the interfaces to the derivation and as the domain for lexicalization. From this perspective, the difference in the specification of f -features in functional heads determines what is a phase and what is not in a particular language, so that the emergence of variation is derived despite the uniformity of syntax. With that framework in mind, I propose a unified solution to the variation across languages of certain lexicalization patterns, namely, the expression of path, manner and possession. The theory I develop builds on a syntactic theory of argument structure that takes into account the role that non-relational elements, namely, roots, have in the model. I propose that syntax operates with root positions that are semantically and phonologically underspecified and that are defined as non-projecting heads, i. e. , non-labeling heads, which by general syntactic principles can only be present at the bottom, first merge position of every (sub)derivation. I propose that phase heads establish domains for lexicalization that allow us to consider morphological operations such as feature percolation or vocabulary insertion as two distinct ways for phonologically interpreting syntactic representations. As for the case studies, I first study the properties of path expressions in Romance and Germanic languages. I propose a simplified structure for PPs in which features such as boundedness arise configurationally, not through a specific functional projection. Then, I argue that case in the adpositional system is structural and emerges from the agreement relation between a DP and the f -features of the functional projection or phase head, p. I argue that spatial expressions in Romance are always locative and that there is defectivity in the content of f -features of path heads that makes them to be non-phase heads and to belong to the vP domain. Second, I deal with some cases of verbal elasticity attested in Romance languages: the existence of cognate objects, that is, unergative intransitive verbs that can take under certain restricted conditions a direct object; and the existence of resultative constructions in Romance where a secondary predicate is said to denote a resultative change of state. In order to account for these constructions in a unified way I propose that they all involve a preposition of central coincidence that establishes a predicative relation between the complement and the verbal root. Finally, I study Measure Verbs that are characterized by having a complement, a Measure Phrase, that behaves sometimes as an adjunct and sometimes as an argument. I propose that this is due mainly to the referential properties of this quantificational element. However, I add into the picture another property that has never been discussed before: the variable behavior of Measure Verbs with respect to unaccusativity. I propose that Measure Verbs have an underlying possessive structure that is headed by a BE predicate that selects a central coincidence preposition, p. The non-defective phase p head in Romance languages can incorporate into BE extending the phase and transforming BE, a defective unaccusative phase, into a non-defective transitive phase, HAVE, which is able to license accusative case and to allow for an external argument position.