This paper addresses the problem of detecting name errors in automatic speech recognition (ASR) output, when available ASR output training data is sparse and covers only a limited subset of scenarios from a broad-domain task. We reduce the dimensionality of the lexical feature set using a related task (sentence-level name detection) and improve feature coverage by expanding single words to classes using semantic similarity measures derived from both WordNet and neural network word embeddings. Phrase patterns are learned over the selected word features using a frequency-based pattern selection algorithm. In experiments on English dialogs, we find that adding sentence-level features performs better than using local n-gram context. Automatically-learned seed features perform better than handcrafted patterns, and their class expansions generalize better. Finally, the embedding-based class expansions yield better results than the corresponding WordNet-based configurations.