1.College of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,Hubei,China
2.Hubei Province Key Laboratory of Intelligent Information Processing and Real-Time Industrial System,Wuhan 430065,Hubei,China
3.Hubei Center for Data and Analysis,Hubei University of Economics,Wuhan 430205,Hubei,China
Show less
文章历史+
Received
Published
2022-08-01
2023-04-24
Issue Date
2026-07-23
PDF (2144K)
摘要
多项选择作为机器阅读理解中的一项重要任务,在自然语言处理(natural language processing,NLP)领域受到了广泛关注。由于数据中需要处理的文本长度不断增长,长文本多项选择成为了一项新的挑战。然而,现有的长文本处理方法容易丢失文本中的有效信息,导致结果不准确。针对上述问题,提出了一种基于压缩与推理的长文本多项选择答题方法(Long Text Multiple Choice Answer Method Based on Compression and Reasoning,LTMCA),通过训练评判模型识别相关句子,将相关句拼接成短文本输入到推理模型进行推理。为了提高评判模型的精度,在评判模型中增加了文章与选项之间的交互以补充文章对选项的注意力,有针对性地进行相关语句识别,更加准确地完成多项选择答题任务。在本文构建的CLTMCA中文长文本多项选择数据集上进行了实验验证,结果表明本文方法能够有效地解决BERT在处理长文本多项选择任务时的限制问题,相比于其他方法,在各项评价指标上均取得了较高的提升。
Abstract
As a significant task in Machine Reading Comprehension, multiple choice has received widespread attention in Natural Language Processing(NLP). Since the length of text that needs to be processed in data continues to be longer, long text multiple choice becomes a new challenge. However, existing long text processing methods tend to lose useful information in the text, leading to inaccurate results. To solve the above problems, this paper proposes the Long Text Multiple Choice Answer Method Based on Compression and Reasoning (LTMCA), which identifies relevant sentences by training the judgment model, and combines the relevant sentences to form a short text to be input into the inference model for inference. In order to improve the accuracy of the evaluation model, the interaction between the essay and the options is added to the evaluation model to supplement the essay's attention to the options, and the relevant statements are identified in a targeted way to complete the multiple-choice answer task more accurately.Experimental verification is carried out on the CLTMCA Chinese long text multiple choice dataset constructed in this paper, and the results show that the proposed method can effectively solve the problems of the BERT in handling the long text multiple choice task. Compared with other methods, this method has greatly improved in various evaluation indicators.
自2018年以来,BERT(bidirectional encoder representation from transformer)[3]及其改造版本在多项自然语言处理任务中显示出了惊人的改进。在谷歌官方发布的BERT-base(Chinese)中,中文以字为粒度进行切分,并没有考虑中文需要分词的特点。而应用全词Mask,非字粒度的中文BERT模型有更好的表现,因此研究人员将全词Mask方法应用在了中文中,在中文维基百科、新闻、问答等数据上进行训练,构造出了新的模型。和原版中文BERT模型相比,全词覆盖的中文BERT预训练模型BERT-wwm*[4]在多个中文任务中取得了更好的表现。作为BERT-wwm*的改进版本,MacBERT[5]引入了纠错型掩码语言模型(MLM as correction,Mac)预训练任务,缓解了“预训练-下游任务”不一致的问题,在多项任务中取得了优于前作的成绩。此外,为了更好地将预训练模型应用于中文任务,研究人员还训练出了中文版本的XLNet[6]、ELECTRA[7]。
本文提出的LTMCA(Long Text Multiple Choice Answer method based on compression and reasoning)利用judger(评判器)与reasoner(推理器)两个预训练模型解决中文长文本多项选择答题任务。通过训练评判器识别相关句,将相关语句拼接后输入到推理器进行多项选择任务的推理。由于BERT的表示根植于字符,无法很好地表示句子的特征,本文更改了BERT的输入序列与嵌入,使句子识别成为可能。在识别过程中,增加了对选项信息的注意力以保证评判模型的准确率,在推理过程中又为评判模型提供更为准确的训练标签,形成上下游任务的联合训练。由于目前公开的中文多项选择数据集文本长度多在512之下,不能很好验证本文方法的有效性,本文提出了CLTMCA中文长文本多项选择数据集,并在此数据集上开展实验,进行对比分析,验证了本文方法的有效性。
TANH Y, LIX Y, LIUB. Reading comprehension based on external knowledge and hierarchical discourse representation[J]. Journal of Chinese Information Processing, 2020, 34(4): 85-91 (Ch). DOI: 10.3969/j.issn.1003-0077.2020.04.011 .
ZHANGH, ZHANGY, YANGZ Z, et al. Data augmentation based automatic answering of reading comprehension in college entrance examination[J]. Journal of Chinese Information Processing, 2021, 35(9): 132-140. DOI: 10.3969/j.issn.1003-0077.2021.09.013 (Ch ).
[5]
DEVLINJ, CHANGM W, LEEK, et al. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding[DB/OL].[2022-08-02].DOI: 10.18653/v1/n18-2 .
[6]
CUIY M, CHEW X, LIUT, et al. Pre-training with whole word masking for Chinese BERT[J]. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021, 29: 3504-3514. DOI: 10.1109/TASLP.2021.3124365 .
[7]
CUIY M, CHEW X, LIUT, et al. Revisiting pre-trained models for Chinese natural language processing[C]//Findings of the Association for Computational Linguistics: EMNLP 2020. Stroudsburg: Association for Computational Linguistics, 2020: 657-668. DOI: 10.18653/v1/2020.findings-emnlp.58 .
[8]
YANGZ L, DAIZ H, YANGY M, et al. XLNet: Generalized autoregressive pretraining for language understanding[EB/OL]. 2019: arXiv: 1906.08237.
[9]
CLARKK, LUONGM T, LEQ V, et al. ELECTRA: Pre-training text encoders as discriminators rather than generators[EB/OL]. 2020: arXiv: 2003.10555. DOI: 10.18653/v1/2020.emnlp-main.20 .
[10]
WANGW, YANM, WUC. Multi-granularity hierarchical attention fusion networks for reading comprehension and question answering[C]//Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Stroudsburg: Association for Computational Linguistics, 2018: 1705-1714. DOI: 10.18653/v1/p18-1158 .
[11]
PAPPAGARIR, ZELASKOP, VILLALBAJ, et al. Hierarchical transformers for long document classification[C]//2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). New York: IEEE Press, 2020: 838-844. DOI: 10.1109/ASRU46091.2019.9003958 .
[12]
VASWANIA, SHAZEERN, PARMARN, et al. Attention is all you need[EB/OL]. 2017: arXiv: 1706.03762.
[13]
DAIZ H, YANGZ L, YANGY M, et al. Transformer-XL: Attentive language models beyond a fixed-length context[C]//Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Stroudsburg: Association for Computational Linguistics, 2019: 2978-2988. DOI: 10.18653/v1/p19-1285 .
[14]
RAEJ W, POTAPENKOA, JAYAKUMARS M, et al. Compressive transformers for long-range sequence modelling[EB/OL]. 2019: arXiv: 1911.05507.
[15]
KITAEVN, KAISERŁ, LEVSKAYAA. Reformer: The efficient transformer[EB/OL]. 2020: arXiv: 2001.04451.
[16]
QIUJ Z, MAH, LEVYO, et al. Blockwise self-attention for long document understanding[C]//Findings of the Association for Computational Linguistics: EMNLP 2020. Stroudsburg: Association for Computational Linguistics, 2020: 2555-2565. DOI: 10.18653/v1/2020.findings-emnlp.232 .
DINGM, ZHOUC, YANGH X, et al. CogLTX: Applying BERT to long texts[C]//Proceedings of the 34th International Conference on Neural Information Processing Systems. New York: ACM, 2020: 12792-12804. DOI: 10.5555/3495724.3496797 .
[19]
INOUEN, TRIVEDIH, SINHAS, et al. Summarize-then-answer: Generating concise explanations for multi-hop reading comprehension[C]//Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. Stroudsburg: Association for Computational Linguistics, 2021: 6064-6080. DOI: 10.18653/v1/2021.emnlp-main.490 .
CHENGJ P, LAPATAM. Neural summarization by extracting sentences and words[C]//Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Stroudsburg: Association for Computational Linguistics, 2016: 484-494. DOI: 10.18653/v1/p16-1046 .
[22]
NALLAPATIR, ZHAIF F, ZHOUB W. SummaRuNNer: A recurrent neural network based sequence model for extractive summarization of documents[C]//Proceedings of the 31st AAAI Conference on Artificial Intelligence. New York: ACM, 2017: 3075-3081. DOI: 10.1609/aaai.v31i1.10958 .
[23]
CHENY C, BANSALM. Fast abstractive summarization with reinforce-selected sentence rewriting[C]//Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Stroudsburg: Association for Computational Linguistics, 2018: 675-686. DOI: 10.18653/v1/p18-1063 .
[24]
PAULUSR, XIONGC M, SOCHERR. A deep reinforced model for abstractive summarization[EB/OL]. 2017: arXiv: 1705.04304.
ZHONGM, LIUP F, CHENY R, et al. Extractive summarization as text matching[C]//Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Stroudsburg: Association for Computational Linguistics, 2020: 6197-6208. DOI: 10.18653/v1/2020.acl-main.552 .
[27]
LAIG K, XIEQ Z, LIUH X, et al. RACE: large-scale ReAding comprehension dataset from examinations[C]//Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. Stroudsburg: Association for Computational Linguistics, 2017: 785-794. DOI: 10.18653/v1/d17-1082 .
[28]
SUNK, YUD, CHENJ S, et al. DREAM: A challenge data set and models for dialogue-based reading comprehension[J]. Transactions of the Association for Computational Linguistics, 2019, 7: 217-231. DOI: 10.1162/tacl_a_00264 .
[29]
LIUY H, OTT M, GOYALN, et al. RoBERTa: A robustly optimized BERT pretraining approach[EB/OL]. 2019: arXiv: 1907.11692.
[30]
LANZ Z, CHENM D, GOODMANS, et al. ALBERT: A lite BERT for self-supervised learning of language representations[EB/OL]. 2019: arXiv: 1909.11942.