Classifying Homographs in Japanese Social Media Texts Using a User Interest Model

The analysis of text data from social media is hampered by irrelevant noisy data, such as homographs. Noisy data is not usable and makes analysis, such as counting estimates, of the target data diffcult, which adversely affects the quality of the analysis results. We focus on this issue and propose a method to classify homographs that are contained in social media texts (i.e. Twitter) using topic models. We also report the results of an evaluation experiment. In the evaluation experiment, the proposed method showed an accuracy improvement of 8.5% and a reduction of 16.5% in the misidentification rate compared with conventional methods.