Word Salad: Relating Food Prices and Descriptions Supplementary Material

We conjecture that systematic trends in errors can help detect patterns in restaurant prices. For example, if errors correlate with geography, we may identify regions with higheror lower-than-expected prices. To test this for New York, we retrained our price range predictor using the entirety of New York as the test set, Philadelphia as the dev set, and the remaining cities as training data. We only used features from the reviews, omitting geographic metadata features. We then compare the predictions of the model to the real price range values in geographic context. More precisely, for every restaurant represented by its projected coordinates, we compute the relative error ŷ−y y at this point and smooth the resulting two-dimensional price distribution using a triangular kernel. We obtain the map of Manhattan shown in Figure 1, revealing trends in our predictions. We first note the prevalence of blue in the plot; since we trained our models on cheaper cities and tested on Manhattan, we are systematically underpricing Manhattan restaurants. We see the most severe underprediction in Midtown East, an area known for its expensive cost of living. However, some areas differ from the overall trend; we observe moderate overprediction in Chinatown and parts of Greenwich Village and the East Village, areas which are generally known to be lower-priced than most of lower Manhattan. Figure 1: Smoothed distribution of relative differences between the value predicted from the reviews and the real price range for 2965 restaurants in Manhattan.