This paper analyzes the impact of sentiment from headlines in the Wall Street Journal on earnings surprises and stock returns of US equities. Negative word counts, lexical analyzers, customized dictionaries, and parts of speech analyzers are used together to determine the efficacy of context-specific sentiment analyzers. As headlines do not follow ordinary language rules and positive and negative words have different connotations in a financial context, five metrics are designed to test how different language analysis techniques capture different information. The results indicate that a combination of custom dictionaries, lexical analyzers, and part of speech analyzers captures sentiment relating to earnings surprise more accurately than simple word counts. All of the metrics are significantly related to next day returns but the variability of the prediction is too large to consider them as part of a profitable trading strategy. The results show there is potential for more complex natural language processing techniques for predicting returns.