一种利用语义相似度改进问答摘要的方法

doi:10.13209/j.0479-8023.2017.028

摘要/Abstract

摘要：

在搜索引擎中对用户问题直接给出简要的答案(即答案摘要)可以帮助用户更快捷的获取信息。针对这一任务, 设计一种基于特征的答案摘要抽取方法。为了进行句子相似性的计算, 提出通过使用卷积神经网络表示句子语义和计算相似性, 同时给出基于最大间隔学习的网络训练方法。在百度知道问答语料上的实验结果表明, 所提出的答案摘要抽取方法能够生成质量良好的简短回答。与基于词袋的相似性计算相比, 使用卷积神经网络能够更好地描述句子语义, 计算问题和句子之间的相似性, 有效地改善答案摘要的质量。

关键词: 问答摘要, 语义相似度计算, 最大间隔学习, 卷积神经网络

Abstract:

In search services, users can get information more conveniently by reading the succinct answers to their questions. This paper introduces a feature-based method for the query-focused summarization to extract the answer summary of a user query. A convolutional neural network (CNN) is used to learn the semantic representation of a sentence, by which the similarity between a candidate answer sentence and a user query is evaluated. The neural network is trained under the framework of max-margin learning. Experiments in Baidu Knows verify that the proposed method can generate the concise answer of a user query.

Key words: query-focused summarization, semantic similarity, max-margin learning, convolutional neural network

中图分类号:

TP391

应文豪, 肖欣延, 李素建, 吕雅娟, 穗志方. 一种利用语义相似度改进问答摘要的方法[J]. 北京大学学报自然科学版, 2017, 53(2): 197-203.

Wenhao YING, Xinyan XIAO, Sujian LI, Yajuan LÜ, Zhifang SUI. Improving Query-Focused Summarization with CNN-Based Similarity[J]. Acta Scientiarum Naturalium Universitatis Pekinensis, 2017, 53(2): 197-203.

导出引用管理器 EndNote|Ris|BibTeX

链接本文: https://xbna.pku.edu.cn/CN/10.13209/j.0479-8023.2017.028

https://xbna.pku.edu.cn/CN/Y2017/V53/I2/197

图/表 10

图1 百度知道的问题及回答示例1

Fig. 1 Example 1 of a question and its answer in Baidu Knows

图2 百度知道的问题及回答示例2

Fig. 2 Example 2 of a question and its answer in Baidu Knows

图3 问答摘要系统的整体流程

Fig. 3 Architecture of the proposed method

图4 学习句子语义表示的卷积神经网络结构

Fig. 4 Convolutional neural network for the semantic representation of a sentence

图5 基于最大间隔学习的神经网络训练框架

Fig. 5 Training framework of proposed convolutional neural network based on max-margin learning

表1 句子选择使用的特征

Table 1 Features for sentence selection

特征	说明
位置	句子在原始答案的位置
中心度	句子的图排序得分
词汇级特征	指示词、命名实体、tf-idf和高频词等
句子相似性	基于BOW或CNN的相似性得分

表2 训练数据的句子等级

Table 2 Rankings for a training sentence

N-gram准确率	等级	N-gram准确率	等级
>0.9	4	>0.2	1
>0.6	3	≤0.2	0
>0.4	2

表3 交叉验证数据集上的实验结果

Table 3 Experiment results in 3-fold cross-validation

模型	Rouge-1			Rouge-2
模型	R	P	F	R	P	F
Random	0.4511	0.4093	0.4041	0.3255	0.2972	0.2956
GraphRank	0.5079	0.4729	0.4726	0.3997	0.3753	0.3737
From Head	0.6704	0.6019	0.6154	0.6152	0.5402	0.5544
Sum_Model	0.6628	0.6003	0.6080	0.5966	0.5363	0.5452
BowSim_Model	0.6541	0.5960	0.6019	0.5821	0.5278	0.5345
CNNSim_Model	0.6759	0.6102	0.6188	0.6119	0.5482	0.5577

表4 交叉验证数据集上的答案平均长度和压缩率

Table 4 Statistics of the original answers in 3-fold cross-validation dataset and their summaries produced by proposed model

数据集	答案摘要平均长度/字	原始答案平均长度/字	平均压缩率/%
Dataset-1	133.05	669.73	34.99
Dataset-2	129.71	858.86	30.79
Dataset-3	137.04	766.08	32.67

表5 百度知道答案摘要示例

Table 5 Example of the simplified answer to a question in Baidu Knows

知道问题	答案摘要
PN结是什么?	采用不同的掺杂工艺, 通过扩散作用, 将P型半导体与N型半导体制作在同一块半导体(通常是硅或锗)基片上, 在它们的交界面就形成空间电荷区称PN结。PN结具有单向导电性。P是positive的缩写, N是negative的缩写, 表明正荷子与负荷子起作用的特点。

参考文献 15

[1]	Cao Z, Wei F, Li S, et al.Learning summary prior representation for extractive summarization // Proceedings of ACL-2015. Beijing, 2015: 829-833
[2]	Cao Z, Wei F, Dong L, et al.Ranking with recursive neural networks and its application to multi-document summarization // Proceedings of AAAI-2015. Austin, 2015: 2153-2159
[3]	Ouyang Y, Li W, Li S, et al.Applying regression models to query-focused multi-document summari-zation. Information Processing & Management, 2011, 47(2): 227-237
[4]	Svore K M, Vanderwende L, Burges A C.Enhancing single-document summarization by combining ranknet and third-party sources // Proceedings of EMNLP-2007. Prague, 2007: 448-457
[5]	Ye S, Qiu L, Chua T S, et al.NUS at DUC 2005: understanding documents via concept links // Proceedings of the Fifth Document Understanding Conference. Vancouver, 2005: 1-6
[6]	Li S, Ouyang Y, Wang W, et al.Multi-document summarization using support vector regression // Proceedings of the 7th Document Understanding Conference. Rochester, 2007: 1-5
[7]	Kim Y.Convolutional neural networks for sentence classification // Proceedings of EMNLP-2014. Doha, 2014: 1746-1751
[8]	Kalchbrenner N, Grefenstette E, Blunsom P.A con-volutional neural network for modelling sentences // Proceedings of ACL-2014. Baltimore, 2014: 655-665
[9]	He H, Gimpel K, Lin J.Multi-perspective sentence similarity modeling with convolutional zeural net-works // Proceedings of EMNLP-2015. Lispon, 2015: 1576-1586
[10]	Herbrich R, Graepel T, Obermayer K.Support vector learning for ordinal regression // Proceedings of the 9th International Conference on Aritifical Neural Networks (ICANN). Edinburgh, 1999: 97-102
[11]	Burges C, Shaked T, Renshaw E, et al.Learning to rank using gradient descent // Proceedings of ICML-2005. Boon, 2005: 89-96
[12]	Norouzi M.Minimal loss hashing for compact binary codes // Proceedings of ICML-2011. Bellevue, 2011: 353-360
[13]	Collobert R, Weston J, Bottou L, et al.Natural language processing (almost) from scratch. Journal of Machine Learning Research, 2011(12): 2493-2537
[14]	Lin C Y, Hovy E. Automatic evaluation of summaries using n-gram co-occurrence statistics // Proceedings of the Association for Computaional Linguistics on Human Language Technology. Stroudsburg: Association for Computational Lingus-tics, 2003: 71-78
[15]	Erkan G, Radev D R.LexRank: graph-based lexical centrality as salience in text summarization. Journal of Aritifical Intelligence, 2004, 22: 457-479