Discovering pattern of onomatopoeia used in food reviews

In this paper, we investigated onomatopoeia usage pattern in food reviews by proposing LDA (Latent Dirichlet Allocation) based onomatopoeia usage pattern analysis model. We collected total 685 numbers of onomatopoeias which are distributed to 208 food categories from 3,581,808 food reviews of Japanese food review site Tabelog. From the experimental result, we found several patterns how the onomatopoeias are chosen. The onomatopoeia is chosen based on user´s interest on the combination of {location of food, material of the food, cooking method} and {the texture of food, sound when eating, and looks of food people´s status when eating the food}. In addition, we investigate how the precision of the clustering result changes depending on the N (number of onomatopoeia of each food categories). We found that the results of N=30 is better than one of N=100 as large number of onomatopoeia for each food categories like 100 is likely to include onomatopoeias that are irrelevant to food.