文章嘗試將卷積神經網絡用于數字人文古籍漢字的元數據加工,將古籍漢字識別問題轉換為卷積神經網絡的分類問題,在缺乏訓練集的情況下通過數據生成技術構建訓練集進行模型訓練,并用于古籍漢字的識別。通過TensorFlow平臺,對773個漢字生成約24萬個訓練樣本,網絡模型可自行判定不可識別的圖片;在提高精確率同時,對這部分數據可直接轉由人工識別,系統更為可靠,作為數字人文古籍元數據加工的半自動化工具,旨在提高古籍資源在數字人文應用研究中的效率。

Convolutional neural network(CNN) is used to index the metadata of Chinese characters in ancientbooks in the field of digital humanities, so that the recognition of Chinese characters in ancient books istransformed into the classification of CNN. As a result of the absence of training sets,data generation technology isused for model training, and then for the recognition of Chinese characters in ancient books. In detail, theTensorFlow platform is used to generate about 240,000 training samples for 773 Chinese characters, and theadopted network model can be used to pick out those unrecognizable character pictures automatically. Then,theunrecognizable character pictures would be transferred for manual recognition,which would be more reliable. Inshort,though still a semi-automatic tool,it can save the manpower cost to a certain extent in the indexing of digital humanistic metadata.