基于主动学习和持续学习的同义词挖掘模型

Synonym mining model based on active learning and continual learning

要提升同义词挖掘的效果通常需要现成的相关领域同义词库的支持。由于相关领域同义词库极其稀缺,给模型优化带来了阻碍。针对缺少相关领域同义词库而导致模型在相关领域的同义词挖掘效果难以持续提升的问题,提出了基于主动学习和持续学习的同义词挖掘模型(SYN-AC)。首先,基于主动学习的方法获取专家标记数据,设计了一个新的损失函数并利用标记后的数据去微调模型;其次,为了减少时间和空间消耗,采用了持续学习的方法,使模型在只使用当前组标记的数据进行训练的情况下,也能不断提高同义词挖掘效果,而不需要每次都使用所有标记数据对模型重新微调。使用了3个数据集模拟专家标记的过程,实验结果表明,在其中2个数据集上比效果最好的BERT(Bidirectional Encoder Representations from Transformers)模型F1值分别提升了9.34个百分点和2.75个百分点。验证了SYN-AC能够有效提高同义词挖掘的效果。

To improve the effectiveness of synonym mining,the support of existing synonym databases in relevant fields is usually required,but the extremely scarcity of synonym databases in related fields hinders model optimization.A SYNonym mining model based on Active learning and Continuous learning(SYN-AC)was proposed to address the issue of the lack of a synonym database in related fields,which makes it difficult to continuously improve the models synonym mining performance in related fields.Firstly,the active learning method was used to obtain expert labeled data,a new loss function was designed and the labeled data was used to fine-tune the model;secondly,in order to reduce time and space consumption,the continuous learning method was adopted,enabling the model to continuously improve synonym mining performance even when only using the labeled data by the current group for training,without the need to use all labeled data to fine-tune the model every time.In the article,three datasets were used to simulate the process of expert labeling.The experiment results show that on two of these datasets,SYN-AC improves the F1 value by 9.34 percentage points and 2.75 percentage points respectively,compared to BERT(Bidirectional Encoder Representations from Transformers)model.So SYN-AC was validated to effectively improve the effectiveness of synonym mining.