Edinburgh Napier University

  When labels are organized into a meaningful taxonomy, the parent-child relationship between labels at different levels can give the classifier additional information not deducible from the data alone, especially with limited training data. As a case study, we illustrate this effect on the task of patent classification—the task of categorizing patent documents based on their technical content. Existing approaches do not take into consideration this additional information. Experiments on two patent classification datasets, WIPO-alpha and USPTO-2M, show that our regularized Gated Recurrent Unit (GRU) architecture already gives a performance improvement with a micro-averaged precision score using the top prediction of 0.5191 and 0.5740 on the two datasets, respectively. However, knowledge transfer along the label hierarchy gives further significant improvement on WIPO-alpha, raising the score to 0.5376, and a small improvement on USPTO-2M to 0.5743. Our analyses reveal that incorporating label information improves performance on classes with fewer examples and makes model robust to errors that result from predicting closely related labels.