基于深度学习模型的摘要结构功能识别方法研究

摘要：

【目的/意义】学术文献的摘要由目的、方法、结果等结构组成，这些结构具有特定的功能。目前，针对摘要

功能结构识别的研究不多，且存在识别效率不高的问题，本文引入双向循环神经网络（Bidirectional Recurrent Neu⁃

ral Network, BiRNN）、双向长短时记忆网络（Bidirectional Long Short Term Memory, BiLSTM）、BiLSTM-CRF、

BERT等深度学习模型，对1232篇情报类期刊论文进行摘要结构功能识别研究。【方法/过程】引入5折交叉验证法

进行多次实验，以避免一次实验的偶然性；实验结果用“均值±标准差”形式表示，同时考虑模型的平均性能和稳定

性；实验结果用F1值进行评价。【结果/结论】与BiRNN、BiLSTM、BiLSTM-CRF等模型相比，BERT模型具有最高

的均值和最低的标准差，这表明该模型不仅具有最优的结构功能识别能力，而且性能稳定，该模型特别适用于摘要

结构功能识别任务。【局限/创新】本文采用的实验语料规模较小且为人工标注，这限制了识别效率的提升。

Abstract:

【Purpose/significance】The academic-literature abstract is composed of several structures with specific functions, such as

purpose, method, result.【Method/process】There are few researches on the recognition methods of abstract structure function, and the

proposed methods performed poor. In view of this, bidirectional recurrent neural network (RNN), bidirectional long short-term memory

(BiLSTM), BiLSTM-CRF and bidirectional encoder representations from transformers (BERT) are introduced to summarize the journal

articles of 1232 CNKI databases. In our experiments, The 5-fold cross validation is used to avoid contingency, the experiment results

are represented by 'average ± standard deviation', which takes the average performance and stability into consideration, the experi⁃

ment results are evaluated by F1-value.【Result/conclusion】The comparative experiment results show that compared with BiRNN,

BiLSTM, BiLSTM-CRF, BERT performs best with highest average and lowest standard deviation, which indicates that this model is

quite fit for recognition of abstract structure function.【Innovation/limitation】The experimental corpus is small-scale and artificial-an⁃

notation, which limits the improvement of recognition performance.

刘忠宝, 王宇飞, 张志剑.

基于深度学习模型的摘要结构功能识别方法研究

[J]. 情报科学, 2021, 39(3): 107-112.

可视化