| Title |
Comparison of Performance of Models Applying 1D-convolutional Layer to Automatic Patent Classification using Detailed Descriptions of Patent Documents |
| Authors |
김성훈(Sunghoon Kim) ; 김승천(Seungcheon Kim) |
| DOI |
https://doi.org/10.5573/ieie.2022.59.1.21 |
| Keywords |
1D convolutional layer; Patent classification; WaveNet; LSTM |
| Abstract |
In patent classification using user-defined classes, the detailed description of a patent sometimes plays an important role in the classification point of view depending on the patent classification point of view. If such patent classification is automatically classified using a recurrent neural network, good performance may not be achieved due to the long length of the detailed description. As a method of applying the data having such a long sequence to classification, there is a method of reducing the long sequence by using a 1D convolutional layer and then applying it to a recurrent neural network or using a model using only the 1D convolutional layer.In this study, we intend to measure the performance of patent classification using detailed descriptions of patent documents. The existing LSTM model, which is a recurrent neural network, a method applied to the LSTM model after going through 1D convolutional layer (1D Conv. LSTM), and a model using WaveNet using only 1D convolutional layer were defined. The classification performance of three patent datasets with a sequence length of 6,000 or more was applied to three models to compare the accuracy. WaveNet showed the highest accuracy for the three datasets, and LSTM showed the lowest accuracy. WaveNet showed a stable trend in verification loss compared to the other two models. In dataset #1, WaveNet uses 1D Conv. Accuracy was recorded 14.5% higher than LSTM and 17% higher than LSTM. In dataset #2, WaveNet uses 1D Conv. The accuracy was 3.1% higher than that of LSTM and 9.8% higher than that of LSTM. WaveNet and 1D Conv. LSTM recorded the same accuracy, and 7.3% higher accuracy than LSTM. When classifying user-defined classes of patents made using detailed descriptions of patents, WaveNet is judged to be a more suitable model compared to the existing recurrent neural network. |