The study reported in this paper attempts to conduct a corpus-driven investigation of an electronic corpus with a view to spot possible metaphors. The corpus is composed of dissertations in Applied Linguistics. In Corpus Linguistics, an important distinction is made between corpus-based and corpus-driven investigations (Tognini-Bonelli, 2001). The former is considered to be restricted to using a corpus as a source of examples, generally to check intuitions and / or the frequency or plausibility of pre-selected samples of linguistic phenomena (words, grammatical structures, etc.). In a corpus-driven stance, on the other hand, the investigation typically proceeds in an inductive way, extracting evidence from the corpus and noting patterns as a means to express regularities in the data. The methodology used here consisted of identifying the collocates for a sample of the key words from the corpus, comparing collocates for word pairs and then finding word pairs which shared low levels of semantic congruity between them. After that, some of the potentially metaphorical word pairs flagged by the computer were then interpreted in detail. In our understanding, this methodology improves on previous corpus-based research in that the metaphorical expressions are not defined beforehand. Rather, it is the way in which words collocate with one another that provides the basis for the identification of metaphorical mappings. (Tony Berber Sardinha)