The focus of the experiments reported in this paper was techniques for combining evidence for cross-language retrieval, searching Arabic documents using English queries. Evidence from multiple sources of translation knowledge was combined to estimate translation probabilities, and four techniques for estimating query-language term weights from document-language evidence were tried. A new technique that exploits translation probability information was found to outperform a comparable technique in which that information was not used. Comparative results for three variants of Arabic ^\light^] stemming are also presented. A simple variant of an existing stemming algorithm was found to result in significantly better retrieval effectiveness. UMIACS-TR-2003-26 LAMP-TR-101