A generalization of n-gram term modeling, namely sn,k -gram, has been recently proposed by allowing k-term skipping in the n-gram representation. This paper presents a so-called multi-modal sn,k -gram similarity which combines multiple similarity vectors resulting from computing similarity between several pairs of queries and documents each of which using s-grams with various n and k. Adjusting weights in the combination enables us to create a suitable approximate matching model between a relevant document and a query although such document does not include any exact terms as in the query or vice versa. To evaluate our proposed method, we analyzed two variants of a multimodal sn,k -gram model, called equal-weighting and performance-based-weighting over all queries on two collections of medical documents that are alike in context but different in written languages. The result shows that the multi-modal sn,k -gram similarity significantly outperforms the conventional unigrams and bigrams. © 2010 Springer-Verlag Berlin Heidelberg