基于知识增强的文本隐喻识别图编码方法

黄河燕; 刘啸; 刘茜

doi:10.7544/issn1000-1239.202110927

基于知识增强的文本隐喻识别图编码方法

北京理工大学计算机学院　北京　100081
北京海量语言信息处理与云计算工程研究中心　北京　100081
北京理工大学东南信息技术研究院　福建莆田　351100

基金项目: 国家重点研发计划项目（2018YFB1005100）；国家自然科学基金项目（61732005）

详细信息

作者简介:
黄河燕: 1963年生.博士，教授，博士生导师.CCF会员.主要研究方向为自然语言处理、信息抽取与信息检索

刘啸: 1994年生.博士研究生.主要研究方向为自然语言处理、信息抽取与事件抽取

刘茜: 1991年生.博士.主要研究方向为自然语言处理、信息检索与语义表示学习

中图分类号: TP391
计量
- 文章访问数: 271
- HTML全文浏览量: 33
- PDF下载量: 125
出版历程
- 收稿日期: 2021-09-08
- 修回日期: 2022-04-14
- 网络出版日期: 2023-02-10
- 刊出日期: 2022-12-31

Knowledge-Enhanced Graph Encoding Method for Metaphor Detection in Text

School of Computer Science and Technology, Beijing Institute of Technology, Beijing 100081
Beijing Engineering Research Center of High-Volume Language Information Processing and Cloud Computing Applications, Beijing 100081
Southeast Academy of Information Technology, Beijing Institute of Technology, Putian, Fujian 351100

Funds: This work was supported by the National Key Research and Development Program of China (2018YFB1005100) and the National Natural Science Foundation of China (61732005).

摘要

摘要:
隐喻识别是自然语言处理中语义理解的重要任务之一，目标为识别某一概念在使用时是否借用了其他概念的属性和特点. 由于单纯的神经网络方法受到数据集规模和标注稀疏性问题的制约，近年来，隐喻识别研究者开始探索如何利用其他任务中的知识和粗粒度句法知识结合神经网络模型，获得更有效的特征向量进行文本序列编码和建模.然而，现有方法忽略了词义项知识和细粒度句法知识，造成了外部知识利用率低的问题，难以建模复杂语境.针对上述问题，提出一种基于知识增强的图编码方法（knowledge-enhanced graph encoding method，KEG）来进行文本中的隐喻识别. 该方法分为3个部分：在文本编码层，利用词义项知识训练语义向量，与预训练模型产生的上下文向量结合，增强语义表示；在图网络层，利用细粒度句法知识构建信息图，进而计算细粒度上下文，结合图循环神经网络进行迭代式状态传递，获得表示词的节点向量和表示句子的全局向量，实现对复杂语境的高效建模；在解码层，按照序列标注架构，采用条件随机场对序列标签进行解码.实验结果表明，该方法的性能在4个国际公开数据集上均获得有效提升.
- 隐喻识别 /
- 图循环神经网络 /
- 知识增强方法 /
- 词义项知识 /
- 细类别句法知识 /
- 序列标注
Abstract:
Metaphor recognition is one of the essential tasks of semantic understanding in natural language processing, aiming to identify whether one concept is viewed in terms of the properties and characteristics of the other. Since pure neural network methods are restricted by the scale of datasets and the sparsity of human annotations, recent researchers working on metaphor recognition explore how to combine the knowledge in other tasks and coarse-grained syntactic knowledge with neural network models, obtaining more effective feature vectors for sequence coding and modeling in text. However, the existing methods ignore the word sense knowledge and fine-grained syntactic knowledge, resulting in the problem of low utilization of external knowledge and the difficulty to model complex context. Aiming at the above issues, a knowledge-enhanced graph encoding method (KEG) for metaphor detection in text is proposed. This method consists of three parts. In the encoding layer, the sense vector is trained using the word sense knowledge, combined with the context vector generated by the pre-training model to enhance the semantic representation. In the graph layer, the information graph is constructed using fine-grained syntactic knowledge, and then the fine-grained context is calculated. The layer is combined with the graph recurrent neural network, whose state transition is carried out iteratively to obtain the node vector and the global vector representing the word and the sentence, respectively, to realize the efficient modeling of the complex context. In the decoding layer, conditional random fields are used to decode the sequence tags following the sequence labeling architecture. Experimental results show that this method effectively improves the performance on four international public datasets.
- metaphor detection /
- graph recurrent neural network /
- knowledge-enhanced method /
- word sense knowledge /
- fine-grained syntactic knowledge /
- sequence labeling

HTML全文

图 1 KEG整体网络架构图

Figure 1. Overall network framework of KEG

下载: 全尺寸图片幻灯片

图 2 文本编码层结构

Figure 2. Structure of the encoding layer

下载: 全尺寸图片幻灯片

图 3 图循环神经网络层结构

Figure 3. Layer structure of the graph recurrent neural network

下载: 全尺寸图片幻灯片

表 1 包含隐喻词汇的文本样例

Table 1 Examples of Sentences with Metaphor Words

编号	文本样例
1	It went to his head rather.
2	How to kill a process?
3	I invested myself fully in this relationship.
4	They forged contacts with those left behind in Germany.
注：黑体单词表示隐喻现象.

下载: 导出CSV

表 2 数据集统计信息

Table 2 Statistics of Datasets

数据集	单词总数	句子总数	隐喻总数	隐喻占比/%	平均单词数
VUA-VERB	239847	16189	6554	28.36	14.82
VUA-ALLPOS	239847	16189	15026	15.85	14.82
TroFi	105949	3737	1627	43.54	28.35
MOH-X	5178	647	315	48.68	8.00

下载: 导出CSV

表 3 本文实验中使用的超参数

Table 3 Hyper-Parameters Used in Our Experiment

超参数	取值
${{\boldsymbol{e}}^{\rm{c}}}$ 的向量维度	1024
${{\boldsymbol{e}}^{\rm{s}}}$ 的向量维度	200
${\boldsymbol{k}}$ 的向量维度	256
${\boldsymbol{m}}$ ， ${\boldsymbol{h}}$ ， ${\boldsymbol{g}}$ 的向量维度	512
状态传递次数 $K$	3
批大小（batch size）	16
BERT模型dropout率	0.5
模型其他部分dropout率	0.4
BERT模型学习率	1E−5
模型其他部分学习率	1E−3
BERT模型 ${L_2}$ 权重衰减	1E−5
模型其他部分 ${L_2}$ 权重衰减	1E−3
Warmup迭代轮次	5
最大迭代轮次	50
梯度裁剪	4.0

下载: 导出CSV

表 4 KEG与SOTA方法对比的实验结果

Table 4 Experimental Results of KEG and SOTA Methods %

组别	对比模型	VUA-VERB				VUA-ALLPOS				TroFi				MOH-X
组别	对比模型	P	R	F1	Acc	P	R	F1	Acc	P	R	F1	Acc	P	R	F1	Acc
第1组	RNN_ELMo^[7]	68.2	71.3	69.7	81.4	71.6	73.6	72.6	93.1	70.7	71.6	71.1	74.6	79.1	73.5	75.6	77.2
	RNN_BERT^[20]	66.7	71.5	69.0	80.7	71.5	71.9	71.7	92.9	70.3	67.1	68.7	73.4	75.1	81.8	78.2	78.1
	RNN_HG^[6]	69.3	72.3	70.8	82.1	71.8	76.3	74.0	93.6	67.4	77.8	72.2	74.9	79.7	79.8	79.8	79.7
	RNN_MHCA^[6]	66.3	75.2	70.5	81.8	73.0	75.7	74.3	93.8	68.6	76.8	72.4	75.2	77.5	83.1	80.0	79.8
第2组	Go Figure!^[9]	73.2	82.3	77.5		72.1	74.8	73.4
第2组	DeepMet-S^[10]	76.2	78.3	77.2	86.2	73.8	73.2	73.5	90.5
第3组	GCN_BERT^[11]									72.3	70.5	70.7	72.0	79.8	79.4	79.3	79.4
第3组	MWE^[11]									73.8	71.8	72.8	73.5	80.0	80.4	80.2	80.5
本文	KEG	80.5	77.4	78.9	86.7	74.6	76.5	75.5	92.8	75.6	77.8	76.7	76.2	82.1	81.6	81.8	81.6

下载: 导出CSV

表 5 对比不同状态传递次数的实验结果

Table 5 Experimental Results With Different State Transition Times %

状态传递次数K	VUA-VERB				VUA-ALLPOS				TroFi				MOH-X
状态传递次数K	P	R	F1	Acc	P	R	F1	Acc	P	R	F1	Acc	P	R	F1	Acc
1	67.3	72.6	69.8	82.5	71.8	73.3	72.5	91.7	72.6	70.4	71.5	74.6	77.8	79.8	78.8	78.1
2	74.6	75.7	75.1	84.2	73.3	75.7	74.5	92.4	74.1	75.6	74.8	75.3	81.3	80.5	80.9	80.8
3	80.5	77.4	78.9	86.7	74.6	76.5	75.5	92.8	75.6	77.8	76.7	76.2	82.1	81.6	81.8	81.6
4	80.9	77.1	79.0	86.2	75.1	76.0	75.5	92.8	75.8	77.5	76.6	76.3	82.1	81.8	81.9	81.2
5	80.8	77.2	78.9	86.6	74.8	76.1	75.4	93.1	76.1	77.8	76.9	75.3	81.4	82.1	81.7	81.6

下载: 导出CSV

表 6 词向量和上下文向量构建方法的实验结果

Table 6 Experimental Results of Constructing Word Vectors and Context Vectors %

对比模型	VUA-VERB				VUA-ALLPOS				TroFi				MOH-X
对比模型	P	R	F1	Acc	P	R	F1	Acc	P	R	F1	Acc	P	R	F1	Acc
KEG-word2vec	74.2	73.2	73.7	82.3	73.5	74.7	74.1	90.6	74.1	73.5	73.8	74.4	78.5	76.6	77.5	78.6
KEG-GloVe	74.6	73.2	73.9	82.6	73.6	74.7	74.1	90.8	74.0	73.8	73.9	74.2	78.4	76.8	77.6	79.1
KEG-ELMo	76.4	75.5	75.9	84.8	74.2	76.1	75.1	91.7	74.6	75.8	75.2	75.6	80.4	80.3	80.3	80.7
KEG	80.5	77.4	78.9	86.7	74.6	76.5	75.5	92.8	75.6	77.8	76.7	76.2	82.1	81.6	81.8	81.6

下载: 导出CSV

表 7 KEG有无词级别义项知识的实验结果

Table 7 Experimental Results of KEG With and Without Lexicon Word Sense Knowledge %

对比模型	VUA-VERB				VUA-ALLPOS				TroFi				MOH-X
对比模型	P	R	F1	Acc	P	R	F1	Acc	P	R	F1	Acc	P	R	F1	Acc
KEG-NOS	68.2	72.9	70.5	82.7	72.6	73.1	72.8	92.1	73.2	71.5	72.3	75.4	79.4	78.2	78.8	78.8
KEG	80.5	77.4	78.9	86.7	74.6	76.5	75.5	92.8	75.6	77.8	76.7	76.2	82.1	81.6	81.8	81.6

下载: 导出CSV

表 8 KEG有无细粒度句法知识的实验结果

Table 8 Experimental Results of KEG With and Without Fine-Grained Syntactic Knowledge %

对比模型	VUA-VERB				VUA-ALLPOS				TroFi				MOH-X
对比模型	P	R	F1	Acc	P	R	F1	Acc	P	R	F1	Acc	P	R	F1	Acc
KEG-UCG	67.5	71.1	69.3	80.6	70.3	71.3	70.8	89.4	72.5	70.1	71.3	74.0	77.5	76.9	77.2	77.1
KEG-CSK	71.2	70.7	70.9	81.3	72.0	71.4	71.7	90.2	73.1	70.7	71.9	74.7	78.2	78.3	78.2	78.3
KEG	80.5	77.4	78.9	86.7	74.6	76.5	75.5	92.8	75.6	77.8	76.7	76.2	82.1	81.6	81.8	81.6

下载: 导出CSV

表 9 2种解码方式的实验结果

Table 9 Experimental Results of Two Decoding Methods %

对比模型	VUA-VERB				VUA-ALLPOS				TroFi				MOH-X
对比模型	P	R	F1	Acc	P	R	F1	Acc	P	R	F1	Acc	P	R	F1	Acc
KEG-C	79.5	77.3	78.4	86.5	74.2	75.6	74.9	92.2	75.8	76.3	76.0	75.9	82.0	81.2	81.6	81.5
KEG	80.5	77.4	78.9	86.7	74.6	76.5	75.5	92.8	75.6	77.8	76.7	76.2	82.1	81.6	81.8	81.6

下载: 导出CSV

参考文献(42)

[1]	Shutova E. Models of metaphor in NLP[C] //Proc of the 48th Annual Meeting of the Association for Computational Linguistics. Stroudsburg, PA: ACL, 2010: 688−697
[2]	Jang H, Moon S, Jo Y, et al. Metaphor detection in discourse[C] //Proc of the 16th Annual Meeting of the Special Interest Group on Discourse and Dialogue. Stroudsburg, PA: ACL, 2015: 384−392
[3]	Klebanov B, Leong C, Gutiérrez E, et al. Semantic classifications for detection of verb metaphors[C] //Proc of the 54th Annual Meeting of the Association for Computational Linguistics. Stroudsburg, PA: ACL, 2016: 101−106
[4]	Bulat L, Clark S, Shutova E. Modelling metaphor with attribute-based semantics[C] //Proc of the 15th Conf of the European Chapter of the Association for Computational Linguistics. Stroudsburg, PA: ACL, 2017: 523−528
[5]	Shutova E, Kiela D, Maillard J. Black holes and white rabbits: Metaphor identification with visual features[C] //Proc of the 2016 Conf of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Stroudsburg, PA: ACL, 2016: 160−170
[6]	Mao Rui, Lin Chenghua, Guerin F. End-to-end sequential metaphor identification inspired by linguistic theories[C] //Proc of the 57th Annual Meeting of the Association for Computational Linguistics. Stroudsburg, PA: ACL, 2019: 3888−3898
[7]	Gao Ge, Choi E, Choi Y, et al. Neural metaphor detection in context[C] //Proc of the 2018 Conf on Empirical Methods in Natural Language Processing. Stroudsburg, PA: ACL, 2018: 607−613
[8]	Rei M, Bulat L, Kiela D, et al. Grasping the finer point: A supervised similarity network for metaphor detection[C] //Proc of the 2017 Conf on Empirical Methods in Natural Language Processing. Stroudsburg, PA: ACL, 2017: 1537−1546
[9]	Chen Xianyang, Leong C, Flor M, et al. Go Figure! Multi-task transformer-based architecture for metaphor detection using idioms: ETS team in 2020 metaphor shared task[C] //Proc of the 2nd Workshop on Figurative Language Processing, Fig-Lang@ACL 2020. Stroudsburg, PA: ACL, 2020: 235−243
[10]	Su Chuandong, Fukumoto F, Huang Xiaoxi, et al. DeepMet: A reading comprehension paradigm for token-level metaphor detection[C] //Proc of the 2nd Workshop on Figurative Language Processing, Fig-Lang@ACL 2020. Stroudsburg, PA: ACL, 2020: 30−39
[11]	Rohanian O, Rei M, Taslimipoor S, et al. Verbal multiword expressions for identification of metaphor[C] //Proc of the 58th Annual Meeting of the Association for Computational Linguistics. Stroudsburg, PA: ACL, 2020: 2890−2895
[12]	Leong C, Klebanov B, Shutova E. A report on the 2018 VUA metaphor detection shared task[C] //Proc of the 1st Workshop on Figurative Language Processing, Fig-Lang@NAACL-HLT 2018. Stroudsburg, PA: ACL, 2018: 56−66
[13]	Birke J, Sarkar A. A clustering approach for nearly unsupervised recognition of nonliteral language[C] //Proc of the 11th Conf of the European Chapter of the Association for Computational Linguistics. Stroudsburg, PA: ACL, 2017: 523−528
[14]	Dunn J. Measuring metaphoricity[C] //Proc of the 52nd Annual Meeting of the Association for Computational Linguistics. Stroudsburg, PA: ACL, 2014: 745−751
[15]	Gutiérrez E, Shutova E, Marghetis T, et al. Literal and metaphorical senses in compositional distributional semantic models[C] //Proc of the 54th Annual Meeting of the Association for Computational Linguistics. Stroudsburg, PA: ACL, 2016: 183−193
[16]	Mohammad S, Shutova E, Turney P. Metaphor as a medium for emotion: An empirical study[C] //Proc of the 5th Joint Conf on Lexical and Computational Semantics, *SEM@ACL 2016. Stroudsburg, PA: ACL, 2016: 23−33
[17]	Tsvetkov Y, Boytsov L, Gershman A, et al. Metaphor detection with cross-lingual model transfer[C] //Proc of the 52nd Annual Meeting of the Association for Computational Linguistics. Stroudsburg, PA: ACL, 2014: 248−258
[18]	Wilks Y. A preferential, pattern-seeking, semantics for natural language inference[J]. Artificial Intelligence, 1975, 6(1): 53−74 doi: 10.1016/0004-3702(75)90016-8
[19]	Pragglejaz Group. MIP: A method for identifying metaphorically used words in discourse[J]. Metaphor and Symbol, 2007, 22(1): 1−39 doi: 10.1080/10926480709336752
[20]	Devlin J, Chang M, Lee K, et al. BERT: Pre-training of deep bidirectional transformers for language understanding[C] //Proc of the 2019 Conf of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Stroudsburg, PA: ACL, 2019: 4171−4186
[21]	Vaswani A, Shazzer N, Parmar N, et al. Attention is all you need[C] //Proc of the Advances in Neural Information Processing Systems 30: Annual Conf on Neural Information Processing Systems 2017. San Diego, CA: NIPS, 2017: 5998−6008
[22]	Liu Yinhan, Ott M, Goyal N, et al. RoBERTa: A robustly optimized BERT pretraining approach [J]. arXiv preprint, arXiv: 1907.11692, 2019
[23]	Leong C, Klebanov B, Hamill C, et al. A report on the 2020 VUA and TOEFL metaphor detection shared task[C] //Proc of the 2nd Workshop on Figurative Language Processing, Fig-Lang@ACL 2020. Stroudsburg, PA: ACL, 2020: 18−29
[24]	Kipf T, Welling M. Semi-supervised classification with graph convolutional networks[C/OL] //Proc of the 5th Int Conf on Learning Representations. La Jolla, CA: ICLR, 2017 [2018-05-20]. https://openreview.net/pdf?id=SJU4ayYgl
[25]	Steen G, Dorst A, Herrmann J, et al. A method for linguistic metaphor identification: From MIP to MIPVU[M] //Converging Evidence in Language and Communication Research, Volume 14. Amsterdam: John Benjamins Publishing Company, 2010: 25−42
[26]	Klebanov B, Leong C, Flor M. A corpus of non-native written english annotated for metaphor[C] //Proc of the 2018 Conf of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Stroudsburg, PA: ACL, 2018: 86−91
[27]	Levin L, Mitamura T, MacWhinney B, et al. Resources for the detection of conventionalized metaphors in four languages[C] //Proc of the 9th Int Conf on Language Resources and Evaluation. Paris: LREC, 2014: 498−501
[28]	Mohler M, Brunson M, Rink B, et al. Introducing the LCC metaphor datasets[C] //Proc of the 10th Int Conf on Language Resources and Evaluation. Paris: LREC, 2016: 4221−4227
[29]	Parde N, Nielsen R. A corpus of metaphor novelty scores for syntactically-related word pairs[C] //Proc of the 11th Int Conf on Language Resources and Evaluation. Paris: LREC, 2018: 1535−1540
[30]	Tredici M, Bel N. Assessing the potential of metaphoricity of verbs using corpus data[C] //Proc of the 10th Int Conf on Language Resources and Evaluation. Paris: LREC, 2016: 4573−4577
[31]	Chen Xinxiong, Liu Zhiyuan, Sun Maosong. A unified model for word sense representation and disambiguation[C] //Proc of the 2014 Conf on Empirical Methods in Natural Language Processing. Stroudsburg, PA: ACL, 2014: 1025−1035
[32]	Mikolov T, Sutskever I, Chen Kai, et al. Distributed representations of words and phrases and their compositionality[C] //Proc of the Advances in Neural Information Processing Systems 26: Annual Conf on Neural Information Processing Systems 2013. San Diego, CA: NIPS, 2013: 3111−3119
[33]	Marcheggiani D, Titov I. Encoding sentences with graph convolutional networks for semantic role labeling[C] //Proc of the 2017 Conf on Empirical Methods in Natural Language Processing. Stroudsburg, PA: ACL, 2017: 1506−1515
[34]	Liu Xiao, Luo Zhunchen, Huang Heyan. Jointly multiple events extraction via attention-based graph information aggregation[C] //Proc of the 2018 Conf on Empirical Methods in Natural Language Processing. Stroudsburg, PA: ACL, 2018: 1247−1256
[35]	Zhang Yue, Liu Qi, Song Linfeng. Sentence-state LSTM for text representation[C] //Proc of the 56th Annual Meeting of the Association for Computational Linguistics. Stroudsburg, PA: ACL, 2018: 317−327
[36]	Song Linfeng, Zhang Yue, Wang Zhiguo, et al. A graph-to-sequence model for AMR-to-text generation[C] //Proc of the 56th Annual Meeting of the Association for Computational Linguistics. Stroudsburg, PA: ACL, 2018: 1616−1626
[37]	Cho K, Merrienboer B, Gülçehre Ç, et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation[C] //Proc of the 2014 Conf on Empirical Methods in Natural Language Processing. Stroudsburg, PA: ACL, 2014: 1724−1734
[38]	Lafferty J, McCallum A, Pereira F. Conditional random fields: Probabilistic models for segmenting and labeling sequence data[C] //Proc of the 18th Int Conf on Machine Learning. San Diego, CA: ICML, 2001: 282−289
[39]	Linguistic Data Consortium. BLLIP 1987-89 WSJ Corpus Release 1[DB/OL]. [2020-12-28]. https://catalog.ldc.upenn.edu/LDC2000T43
[40]	Loshchilov I, Hutter F. Decoupled weight decay regularization[C/OL] //Proc of the 7th Int Conf on Learning Representations. La Jolla, CA: ICLR, 2019 [2020-01-10]. https://openreview.net/pdf?id=Bkg6RiCqY7
[41]	Pennington J, Socher R, Manning C. GloVe: Global vectors for word representation[C] //Proc of the 2014 Conf on Empirical Methods in Natural Language Processing. Stroudsburg, PA: ACL, 2014: 1532−1543
[42]	Peters M, Neumann M, Iyyer M, et al. Deep contextualized word representations[C] //Proc of the 2018 Conf of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Stroudsburg, PA: ACL, 2018: 2227−2237

施引文献(3)

期刊类型引用(2)

1.	刘阳，鲁圆圆，郭成城. 基于优先级的数据中心任务优化调度算法设计. 计算机仿真. 2025(01): 497-500+507 . 百度学术
2.	骆海霞. 基于递推估计的Web前端偶发任务能耗感知方法. 黑龙江工业学院学报(综合版). 2023(10): 115-120 . 百度学术