Abstract. This paper describes two procedures for generating very short summaries for documents from the DUC-2003 competition: a chunk extraction method based on syntactic dependences, and a simple keyword-based extraction. We explore different techniques for extraction and weighting chunks from the texts, and we draw conclusions on the evaluation metric used and the kind of features that are more useful. Two preliminary versions of this procedure ranked in the 12th and 13th positions with respect to unigram recall (ROUGE-1) at DUC-2004 (out of 39 runs submitted).