The problem facing people who intend to work with data that does not yet exist becomes one of feature selection: what data matters and how do we use it? For NLP

The problem facing people who intend to work with data that does not yet exist becomes one of feature selection: what data matters and how do we use it? For NLP tasks, does stemming matter? What about part-of-speech tagging? Indeed, I worked on machine learning in NLP (fluency ranking, parse disambiguation). As a general rule, roughly 90% of the improvement of models is in clever feature engineering and exploiting the underlying system to get more interesting information that improves classification, 10% you get from using more clever machine learning techniques than, say a standard maxent learner with a gaussian prior (for linearly separable data). For instance, the last relatively large boosts of the accuracy of the parser developed by our research group came from feature engineering: http://www.let.rug.nl/vannoord/papers/iwpt2007.pdf http://www.let.rug.nl/vannoord/papers/iwpt2007.pdf