基于神经网络的汽车说明书问答系统
|
摘要
为了简化用户查阅汽车说明书的流程,设计了针对中文汽车说明书的问答系统(CM-QA),包括以下3个问题:1)如何充分利用文档信息表示文档;2)领域词汇的分词和复述问题;3)正负样本不均衡。为了解决上述问题,结合卷积神经网络和双向长短时记忆网络对文本建模,手工构建领域词的复述词典,并使用字向量替代词向量。最后,尝试将模型转换为基于Pairwise思想的排序模型和扩展正例两种训练策略来解决正负样本不均衡的问题。在800条人工标注的问题上对系统进行了测试,其准确率达到了93.07%。
In order to simply the process for users to read the car manual,we construct a new QA system on the Chinese car manual.We call it Car Manual Question Answering(CM-QA)System.The goal of this task is to find the document with the relevant answers in the manual when given a question.This system includes three difficulties:(1)How to make use of all the information in the document.(2)This task contains a large number of domain words.Each domain word has not only one segmentation results and paraphrases.(3)The proportion of positive and negative cases in the corpus is extremely uneven.To solve these problems,we model the question and document by Convolution Neural Network(CNN)and bidirectional Long Short Time Memory(Bi-LSTM).and we replace the word vector representations with character vector representations and construct a paraphrase dictionary of domain words by hand.Finally,we design two different training strategies,namely,transforming the model into a ranking model based on pairwise method and expanding the positive case,so as to improve the imbalance between positive and negative cases.We test our system on 800hand-build test samples,with the accuracy of 93.07%.
In order to simply the process for users to read the car manual,we construct a new QA system on the Chinese car manual.We call it Car Manual Question Answering(CM-QA)System.The goal of this task is to find the document with the relevant answers in the manual when given a question.This system includes three difficulties:(1)How to make use of all the information in the document.(2)This task contains a large number of domain words.Each domain word has not only one segmentation results and paraphrases.(3)The proportion of positive and negative cases in the corpus is extremely uneven.To solve these problems,we model the question and document by Convolution Neural Network(CNN)and bidirectional Long Short Time Memory(Bi-LSTM).and we replace the word vector representations with character vector representations and construct a paraphrase dictionary of domain words by hand.Finally,we design two different training strategies,namely,transforming the model into a ranking model based on pairwise method and expanding the positive case,so as to improve the imbalance between positive and negative cases.We test our system on 800hand-build test samples,with the accuracy of 93.07%.
引文
[1] Robertson S E,Walker S,Beaulieu M,et al.Okapi at Trec-7:Automatic Ad hoc,Filtering,VLC and Interactive Track[J].Nist Special Publication,1999:253-264.
[2] Ponte J M,Croft W B.A Language Modeling Approach to Information Retrieval[C]∥Proc of the 21st annual Int ACM SIGIR Conf on Research and Development in Information Retrieval.New York:ACM,1998:275-281.
[3] Lafferty J,Zhai C.Document Language Models,Query Mmodels,and Risk Minimization for Information Retrieval[C]∥Proc of the 24th Annual Int ACM SIGIR Conf on Research and Development in Information Retrieval.New York:ACM,2001:111-119.
[4] Nallapati R.Discriminative Models for Information Retrieval[C]∥Proc of the 27th Annual Int ACM SIGIR Conf on Research and Development in Information Retrieval.New York:ACM,2004:64-71.
[5] Gao J,Qi H,Xia X,et al.Linear Discriminant Model for Information Retrieval[C]∥Proc of the 25th Annual Int ACM SIGIR Conf on Research and Development in Information Retrieval.New York:ACM,2005:290-297.
[6] Burges C,Shaked T,Renshaw E,et al.Learning to Rank Using Gradient Descent[C]∥Proc of the 22nd Int Conf on Machine learning.New York:ACM,2005:89-96.
[7] Yang Y,Yih W,Meek C.Wikiqa:A Challenge Dataset for Open-domain Question Answering[C]∥Proc of the 2015Conf on Empirical Methods in Natural Language Processing.Stroudsburg,PA,USA,ACL,2015:2013-2018.
[8] Bordes A,Chopra S,Weston J.Question Answering with Subgraph Embeddings[C]∥Proc of the 2014Conf on Empirical Methods in Natural Language Processing.Stroudsburg,PA,USA,ACL,2014:615-620.
[9] Yu L,Hermann K M,Blunsom P,et al.Deep Learning for Answer Sentence Selection[Z/OL].arXivpreprint arxiv:1412.1632,2014.
[10] Yin W,Schütze H,Xiang B,et al.ABCNN:Attention-Based Convolutional Neural Network for Modeling Sentence Pairs[J].Transactions of the Association of Computational Linguistics,2016,4(1):259-272.
[11] Wang B,Liu K,Zhao J.Inner Attention based Recurrent Neural Networks for Answer Selection[C]∥Proc of the 54th Annual Meeting of the Association for Computational Linguistics(Volume 1:Long Papers).Stroudsburg,PA,USA,ACL,2016,1:1288-1297.
[12] Rajpurkar P,Zhang J,Lopyrev K,et al.SQuAD:100,000+Questions for Machine Comprehension of Text[C]∥Proc of the 2016Conf on Empirical Methods in Natural Language Processing.Stroudsburg,PA,USA,ACL,2016:2383-2392.
[13]Wang S,Jiang J.Machine Comprehension Using Match-lstm and Answer Pointer[Z/OL].arXiv preprint arXiv:1608.07905,2016.
[14] Wang W,Yang N,Wei F,et al.Gated Self-matching Networks for Reading Comprehension and Question Answering[C]∥Proc of the 55th Annual Meeting of the Association for Computational Linguistics(Volume 1:Long Papers).Stroudsburg,PA,USA,ACL,2017,1:1288-1297 189-198.
[15] Yu A W,Dohan D,Luong M T,et al.QANet:Combining Local Convolution with Global Self-Attention for Reading Conprehension[Z/OL].arXiv Preprint arXiv.1804.09541.2018.
[16] Song X M,Feng F L,Liu J H,et al.NeuroStylist:Neural Compatibility Modeling for Clothing Matching[C]∥Proc of the 2017ACM on Multimedia Conf.New York:ACM,2017:753-761.
[17] Mikolov T,Sutskever I,Chen K,et al.Distributed Representations of Words and Phrases and Their Compositionality[C]∥Int Conf on Neural Information Processing Systems.Curran Associates Inc.2013:3111-3119.
[18] Kingma D P,Ba J.Adam:a Method for Stochastic Optimization[Z/OL].arXiv preprint arXiv:1412.6980,2014.
[2] Ponte J M,Croft W B.A Language Modeling Approach to Information Retrieval[C]∥Proc of the 21st annual Int ACM SIGIR Conf on Research and Development in Information Retrieval.New York:ACM,1998:275-281.
[3] Lafferty J,Zhai C.Document Language Models,Query Mmodels,and Risk Minimization for Information Retrieval[C]∥Proc of the 24th Annual Int ACM SIGIR Conf on Research and Development in Information Retrieval.New York:ACM,2001:111-119.
[4] Nallapati R.Discriminative Models for Information Retrieval[C]∥Proc of the 27th Annual Int ACM SIGIR Conf on Research and Development in Information Retrieval.New York:ACM,2004:64-71.
[5] Gao J,Qi H,Xia X,et al.Linear Discriminant Model for Information Retrieval[C]∥Proc of the 25th Annual Int ACM SIGIR Conf on Research and Development in Information Retrieval.New York:ACM,2005:290-297.
[6] Burges C,Shaked T,Renshaw E,et al.Learning to Rank Using Gradient Descent[C]∥Proc of the 22nd Int Conf on Machine learning.New York:ACM,2005:89-96.
[7] Yang Y,Yih W,Meek C.Wikiqa:A Challenge Dataset for Open-domain Question Answering[C]∥Proc of the 2015Conf on Empirical Methods in Natural Language Processing.Stroudsburg,PA,USA,ACL,2015:2013-2018.
[8] Bordes A,Chopra S,Weston J.Question Answering with Subgraph Embeddings[C]∥Proc of the 2014Conf on Empirical Methods in Natural Language Processing.Stroudsburg,PA,USA,ACL,2014:615-620.
[9] Yu L,Hermann K M,Blunsom P,et al.Deep Learning for Answer Sentence Selection[Z/OL].arXivpreprint arxiv:1412.1632,2014.
[10] Yin W,Schütze H,Xiang B,et al.ABCNN:Attention-Based Convolutional Neural Network for Modeling Sentence Pairs[J].Transactions of the Association of Computational Linguistics,2016,4(1):259-272.
[11] Wang B,Liu K,Zhao J.Inner Attention based Recurrent Neural Networks for Answer Selection[C]∥Proc of the 54th Annual Meeting of the Association for Computational Linguistics(Volume 1:Long Papers).Stroudsburg,PA,USA,ACL,2016,1:1288-1297.
[12] Rajpurkar P,Zhang J,Lopyrev K,et al.SQuAD:100,000+Questions for Machine Comprehension of Text[C]∥Proc of the 2016Conf on Empirical Methods in Natural Language Processing.Stroudsburg,PA,USA,ACL,2016:2383-2392.
[13]Wang S,Jiang J.Machine Comprehension Using Match-lstm and Answer Pointer[Z/OL].arXiv preprint arXiv:1608.07905,2016.
[14] Wang W,Yang N,Wei F,et al.Gated Self-matching Networks for Reading Comprehension and Question Answering[C]∥Proc of the 55th Annual Meeting of the Association for Computational Linguistics(Volume 1:Long Papers).Stroudsburg,PA,USA,ACL,2017,1:1288-1297 189-198.
[15] Yu A W,Dohan D,Luong M T,et al.QANet:Combining Local Convolution with Global Self-Attention for Reading Conprehension[Z/OL].arXiv Preprint arXiv.1804.09541.2018.
[16] Song X M,Feng F L,Liu J H,et al.NeuroStylist:Neural Compatibility Modeling for Clothing Matching[C]∥Proc of the 2017ACM on Multimedia Conf.New York:ACM,2017:753-761.
[17] Mikolov T,Sutskever I,Chen K,et al.Distributed Representations of Words and Phrases and Their Compositionality[C]∥Int Conf on Neural Information Processing Systems.Curran Associates Inc.2013:3111-3119.
[18] Kingma D P,Ba J.Adam:a Method for Stochastic Optimization[Z/OL].arXiv preprint arXiv:1412.6980,2014.