Access to well-organized, precise knowledge is critical for many practical applications, such as Semantic Search, Reasoning and Question Answering. Real-world knowledge is often unstructured, noisy and embedded in texts. This inspires us to connect text with knowledge. Text can be seen as both the source and the destination of knowledge. In one direction, knowledge can be distilled from text. While in the other direction, knowledge can be leveraged to understand text. Therefore, connecting text with knowledge can benefit both knowledge harvesting and text understanding, and ultimately facilitate many other tasks. Connecting text with knowledge is challenging for several reasons. First, knowledge is implicit in text. Various kinds of signals need to be leveraged to distill knowledge out of noises. Second, text is inherently ambiguous. The same textual mention can have different meanings, depending on the contexts of its appearance. Finally, there is a gap existing between knowledge representation and text understanding/reasoning.

In this thesis, we propose to tackle the challenges from two perspectives: (1) For linking text with knowledge, we study the entity linking problem. It links text mentions to the corresponding entries in a reference knowledge base, so that semantic information can be transferred from knowledge base to text, and then new knowledge can be harvested from text to complete the knowledge base. In this work we target for alleviating the limitations of using Wikipedia as the reference knowledge base. We design innovative models to mine evidences scattered in text corpus and leverage them to compensate the missing information in the reference knowledge base. (2) For leveraging knowledge for text understanding, we develop the "coherent scene extraction" algorithm to utilize background knowledge for filling in the implicit yet critical information in text and show its effectiveness in answering elementary science questions. All the methods proposed in this thesis are comprehensively evaluated on real-life data to demonstrate the power of connecting text with knowledge.