Japan Patent Office manually annotates submitted patents with F-terms (a patent classification scheme consisting of more than 300,000 labels) to aid search for prior patent applications. Keeping up the quality of F-term annotation is critical to patentability assessments, thus there is a demand for an automatic way to assist F-term annotation. One potential solution is to point out annotation mistakes by utilizing machine learning-based classification. However, the annotators cannot validate the predicted corrections because conventional classification methods do not give the rationales behind the corrections. Thus, the annotators may only adopt all or no corrections. The goal of this study was to assist F-term annotation by presenting annotators with corrections on the F-term annotation and the rationales behind the corrections.
We proposed a joint neural model for F-term annotation and rationale identification. The proposed method incorporates a large portion of data annotated only with F-terms and a small portion of data annotated with rationales. It was first trained for F-term annotation, and then fine-tuned using the ground-truth rationales to discriminate rationales from non-rationales.
We evaluated the proposed method on multiple F-terms from different technical domains. The proposed method outperformed baseline methods in terms of the rationale identification, implying that incorporating rationales in training is particularly useful in identifying rationales.