The field of natural language processing (NLP) aims to develop computer systems that are capable of understanding and responding to human language.  To process language robustly, NLP systems need a huge repository of real-world knowledge.  In this talk, I describe how I mine interesting knowledge from three kinds of unstructured web data sources: (1) raw English text, extracted from the entire web, (2) bilingual text (i.e., sentences paired with their translations), and (3) visual data found in web images.  I describe new and effective ways to apply these knowledge sources to difficult NLP problems.  This work can be used to improve a range of important applications, including automated question answering systems (like IBM's Watson), Internet search engines, and tools for automatic machine translation.