【作者】 冯晓硕； 吕泽宇； 陈东明；

【Author】 Xiaoshuo Feng;Zeyu Lv;Dongming Chen;Naval Research Institute;Software college of Northeastern University;

【机构】 海军研究院； 东北大学软件学院；

【摘要】 在生成式文本摘要领域中,当前一般使用预训练模型与摘要数据集训练摘要模型。当需要为不同领域的文本生成摘要,且该领域文本数据量较少并无监督信息时,为该领域文本生成摘要面临巨大挑战。利用少量的无对应摘要的特殊领域的新闻文本,针对特殊领域新闻的特点,使用新闻前两句高概括性的文本对模型进行训练。改进了在特殊领域新闻上的生成效果。算法中在使用特殊领域的新闻文本训练后,冻结模型编码器的参数,对解码器进行训练,解决了训练文本与训练目标多次更换后摘要生成错误的问题。通过将编码器输出的嵌入应用于自行收集的特殊领域新闻上的分类任务中,将预测准确率从79.3%提升到82.1%,有效证明了训练任务的有效性。并通过人工评估,证明了在特殊领域文本上生成的摘要与训练之前有显著区别并且更加准确。更多还原

【Abstract】 In the field of generative text summarization,pre-trained models and summarization datasets are generally used to train summarization models.When needed to generate summarizations for texts in different domain,and the amount of text data in the domain is small and the data lacks supervision information,generating summarizations for texts in the domain faces a huge challenge.Using a small number of news texts in special domain without corresponding summarizations,and aiming at the characteristics of news in special fields,continue training the model by using the first two sentences of news with high generality text which is similar to summarization generation task.Finally,freeze the parameters of the model encoder in stages when training the model,which improves the effect of generating news in special fields,and solves the problem of incorrect summarization generation after changing the domain of text and training objectives.By applying the embedding calculated by the encoder to the classification task of self-collected special domain news,the accuracy increases from 79.3% to 82.1%,which effectively proves the effectiveness of the training task.And through manual evaluation,it is proved that the summarizations generated on the special domain news is significantly different from the summarization generated by models which trained on general domain texts and is more accurate.更多还原