融合深度学习和链路预测的细粒度技术预测研究--以合成生物技术为例

Research on Fine-Grained Technology Prediction Based on Deep Learning and Link Prediction: Take Synthetic Biologyas an Example

[目的/意义]面向专利文本进行更细粒度的技术实体识别和技术预测,利于更详细地把握专利技术布局与趋势。[方法/过程]首先利用深度学习方法自动识别专利技术术语类实体,通过实验对比多组深度学习算法的优劣。其次,提出新的半监督标注和自定义标注方案,提高人工标注效率。最后,执行训练得到的最优模型,结合链路预测方法,对合成生物技术进行细粒度的技术预测。[结果/结论]实证结果表明RoBERTa-BiLSTM-CRF模型更适用于语义复杂的专利技术实体识别,F1值可达到86.8%,技术识别结果比传统IPC分析方法更精细。同时,细粒度的技术预测结果表明,合成生物学的合成方法在不断改进创新,合成物研究向合成燃料发展。

[Purpose/Significance]It is beneficial to grasp the layout and trend of patent technology by identif-ying technical entities and predicting technology with finer granularity for patent texts.[Method/Process]The deep learning method was used to automatically identify patent technology terms entities,and the advantages and disadvan-tages of several groups of deep learning algorithms were compared by empirical analysis.At the same time,new semi-supervised labeling and self-defined labeling schemes were proposed to improve the efficiency of manual labeling.Fi-nally,the optimal model obtained by training was implemented,and the fine-grained technical prediction of synthetic biotechnology was made by combining the link prediction method.[Result/Conclusion]The empirical results show that RoBERTa-BiLSTM-CRF model is more suitable for the recognition of patent technical terms with complex seman-tics,and the F1 value reaches 86.8%.The technical recognition result is more detailed than the traditional IPC a-nalysis method.The fine-grained technical prediction shows that the synthetic methods of synthetic biology are con-stantly improving and innovating,and the synthetic research is developing towards synthetic fuels.