Strength in Numbers: Using Big Data to Simplify Sentiment Classification.

Sentiment classification, the task of assigning a positive or negative label to a text segment, is a key component of mainstream applications such as reputation monitoring, sentiment summarization, and item recommendation. Even though the performance of sentiment classification methods has steadily improved over time, their ever-increasing complexity renders them comprehensible by only a shrinking minority of expert practitioners. For all others, such highly complex methods are black-box predictors that are hard to tune and even harder to justify to decision makers. Motivated by these shortcomings, we introduce BigCounter: a new algorithm for sentiment classification that substitutes algorithmic complexity with Big Data. Our algorithm combines standard data structures with statistical testing to deliver accurate and interpretable predictions. It is also parameter free and suitable for use virtually "out of the box," which makes it appealing for organizations wanting to leverage their troves of unstructured data without incurring the significant expense of creating in-house teams of data scientists. Finally, BigCounter's efficient and parallelizable design makes it applicable to very large data sets. We apply our method on such data sets toward a study on the limits of Big Data for sentiment classification. Our study finds that, after a certain point, predictive performance tends to converge and additional data have little benefit. Our algorithmic design and findings provide the foundations for future research on the data-over-computation paradigm for classification problems.