Graph-like structure of language has been an integral part of natural language processing from the start: context-free and dependency parse trees are in the core of syntactic processing, and semantic networks were proposed as early as 1950s for semantic processing and they are still very relevant in the Semantic Web context. I will present some of our approaches to graph-based language mining: First, at the lexical level, Eulerian graphs derived from character n-gram corpus profiles are used to visualize collocations. Second, we extract biological events from biomedical scientific literature, using rules induces from dependency graphs and subgraph isomorphism. Third, we demonstrate how typed feature-structure graphs can be used to answer natural-language questions using a textual knowledge-base.