【目的】探究不同深度學習模型的科技論文摘要語步識別效果,并分析識別效果差異原因。【方法】構建大規模的科技論文結構化摘要語料庫,選擇10 000和50 000兩種樣本量的訓練集,以傳統機器學習方法 SVM作為對比基準,引入多種深度學習方法(包括DNN、LSTM、Attention-BiLSTM等神經網絡模型),開展語步識別實驗,并對實驗結果進行對比分析。【結果】Attention-BiLSTM方法在兩種樣本量下的實驗中都取得最好的識別效果, 50 000樣本量下F1值達0.9375; SVM方法的識別效果意外好于DNN、LSTM兩種深度學習方法;但是,樣本量從10 000增加到50 000時, SVM方法的識別效果提升最小(F1值提升0.0125), LSTM方法效果提升最大(F1值提升0.1125)。【局限】由于該領域尚未有公開的通用語料,主要以筆者收集的結構化論文摘要作為訓練和測試語料,因此本文的研究結果在與他人比較時有一定的局限性。【結論】雙向LSTM網絡結構和注意力機制能夠顯著提升深度學習模型的語步識別效果;深度學習方法在大規模訓練集下更能體現其優越性。

[Objective] This paper compares the performance of move recognition methods with different deep learning algorithms. [Methods] Firstly, we built a large training corpus. Then, we used the traditional machine learning method SVM as a benchmark, and developed four moves recognition models based on DNN, LSTM, Attention-Bi LSTM and LSTM. Finally, we conducted two rounds of experiments with sample size of 10,000 and 50,000. [Results] Attention-Bi LSTM method achieved the best results in both experiments over the four methods(F1=0.9375 with the larger sample). SVM method outperformed DNN and LSTM in both experiments. While changing sample size from 10,000 to 50,000, SVM received the least increase of F1 score(0.0125), and LSTM had the largest increase of F1 score(0.1125). [Limitations] There is no universal test corpus for similar research. Therefore, our results could not be compared with the results of other studies. [Conclusions] The bi-directional LSTM network structure and attention mechanism can significantly improve the performance of move recognition. The deep learning methods work better with larger sample size.