Ensuring language coverage in dialog systems can be a challenge, since the language in a domain may drift over time, creating a mismatch between the original training data and current input. This in turn degrades performance by increasing misunderstanding and eventually leading to task failure. Without the capability of adapting the vocabulary and the language model based on certain domains or users, recognition errors may degrade the understanding performance, and even lead to a task failure, which incurs more time and effort to recover. This paper investigates how coverage can be maintained by automatically acquiring potential out-of-vocabulary (OOV) words by leveraging different types of relatedness between vocabulary items and words retrieved from web-based resources. Our experiments show that both recognition and semantic parsing accuracy can thereby be improved.