Elsevier

Neurocomputing

Volume 428, 7 March 2021, Pages 268-279
Neurocomputing

Metaphor identification: A contextual inconsistency based neural sequence labeling approach

https://doi.org/10.1016/j.neucom.2020.12.010Get rights and content

Abstract

Metaphor identification helps improve the performance of various natural language understanding tasks such as word sense disambiguation and sentiment analysis. Though many efforts have been made to deal with metaphor identification, most existing studies largely overlook a fact of contextual inconsistency of the metaphors in natural language. We observe that the greater the semantic inconsistency between current word and contextual words is, the more likely the word belongs to the metaphorical category. In this paper, we formulate the metaphor identification as a sequential tagging problem, and then develop a novel contextual inconsistency based neural sequence labeling approach, which can leverage the semantic contextual inconsistency among words of a sentence to address the problem. We propose to rely on distance metric to measure the contextual inconsistency, and evaluate four widely used distance functions in experiments. Experimental results on publicly available datasets validate the benefit of the proposed model over state-of-the-art baselines for metaphor identification.

Introduction

Metaphors are ubiquitous in language, on average in every third sentence of general-domain text [1]. A metaphor makes a sentence more vivid and poetic, as well as more obscure in linguistic research [2]. From the perspective of cognition, the essence of a metaphor is the understanding and experiencing of one kind of thing in terms of another [3]. A metaphor is not only a property of language, but also a tool for cognitive activities of humans that help to construct conceptual systems [1]. To alleviate the lack of appropriate words to express new concepts in a certain period of time, humans create new metaphorical meanings of words using active imagination [2]. Some metaphorical and unmetaphorical sentences are shown as follows:

Example 1: The painting won critical acclaim.

Example 2: This army won a battle.

Example 3: My sister’s memory is a camera that remembers everything we see.

Example 1 shows a metaphorical sentence, where painting belongs to the culture and art category, which is literally different from won from the battle category. Clearly, it would make the description of painting more powerful by using the word won. In contrast, in the unmetaphorical sentence of Example 2, both words army and won literally come from the same category battle. We observe that the literal difference between the source word or category (battle) and target word or category (culture and art) forms the semantic contextual inconsistency, and thus may result in the metaphor in the sentence. The essential feature of a metaphor, i.e., contextual inconsistency, is also observed in the metaphorical sentence of Example 3, as shown by the words memory and camera. Generally, the rich imagination and creativity of human minds may lead to the understanding of senses of natural language beyond the narrowly conceived conception [4]. Thus, it is often difficult to figure out the polysemous meanings and metaphorical uses of words in different contexts of language.

Metaphor identification is an imperative technology for semantic understanding tasks in natural language processing, such as machine translation, information retrieval, and sentiment analysis. Metaphorical meanings account for 20% of word interpretation tasks, which pose a great challenge to word sense disambiguation. Due to the obscurity of metaphor, 44% of metaphorical expressions are translated incorrectly in Google Translate [5]. The poetic nature and imagery of metaphorical expressions cause that they tend to express more implicit emotions than literal expressions [6], which is an obstacle for sentiment analysis.

By exploiting a range of metaphor properties, various efforts have been made to cope with metaphor identification, for example, selectional preference [7], [8], concreteness and imageability [9], [10], [11], and conceptual information [5], [12], [13]. Metaphor identification has been formulated as a classification problem in previous studies. Traditional methods focus on carefully designing various task-specific features by investigating the linguistic properties of metaphors [4], [14], [15], [16], [17], [18]. Recently, various kinds of neural network models have been proposed [19], [20], [21], [22], [23], [24], [25], [26], [27], and have achieved good performances. To our knowledge, the semantic contextual inconsistency, a strong indicator of metaphors, has not been studied in previous work.

In this work, following previous studies, we formulate the metaphor identification as a sequential tagging problem. To deal with the problem, we propose to learn from the semantic contextual inconsistency among words, in addition to sequential context semantics of given sentences. Generally, a metaphorical word is used to modify related words in an imaginative way, and is often literally inconsistent from the context in a sentence. We observe that the greater the semantic inconsistency between current word and contextual words is (e.g., from two different domains), the more likely the word would be categorized as metaphor. In order to measure contextual inconsistency, we represent certain semantic meaning of each word within a sentence by using an abstract distributed representation, and we propose to employ the distributional distance between the abstract distributed representations of each pair of words in the sentence. Then, we develop a novel neural sequence labeling model, named SEQ-CI, which can exploit the contextual inconsistency property of natural language to improve metaphor identification. One key benefit of SEQ-CI is that it exploits the contextual inconsistency as regularization, and may allow better learning of model parameters from real-life natural language data.

We have made the following main contributions in this paper:

  • We find that the contextual inconsistency is an essential feature of a metaphor. To the best our knowledge, this is the first work that leverages semantic contextual inconsistency to help identify metaphor.

  • A new enhanced sequence labeling approach to metaphor identification is presented, which can learn from the contextual relationships between pair-wise words in a sentence.

  • We evaluate the proposed model on publicly available data, and the results show the superiority of our model over the state-of-the-art baselines.

The rest of this paper is organized as follows. Section 2 presents the related work to metaphor identification. Section 3 describes the proposed contextual inconsistency based sequence labeling approach. Section 4 illustrates the experimental results and analysis. In Section 5, we present discussions on several main aspects of the proposed model, and conclude the paper in Section 6.

Section snippets

Related work

A variety of work has been done to cope with metaphor identification, and can be roughly grouped into following categories, i.e., selectional preference, concreteness and imageability, conceptual information, and classification formulation.

Overview

We aim to deal with the metaphor identification from natural language, and following previous studies [22], [26], [27], we formulate it as a sequential classification problem. Formally, let C be a labeled collection of sentences, C={(S1,Y1),,(Sl,Yl),,(SL,YL)}, where Sl is an input sentence with Nl words, Sl={w1,w2,,wNl}, while Yl={y1,y2,,yNl} denotes the labels of respective words in Sl. Note that yi{0,1}, and yi=1 means the word wi is labeled as metaphor (i=1,2,,Nl). The metaphor

Data

To evaluate the propose model SEQ-CI for metaphor identification, we adopted three commonly used English datasets, i.e., VUA [37], TroFi [38], and MOH-X [20], and one publicly available Chinese metaphor dataset called CHI.1 Table 1 lists the statistics of the datasets.

VUA: This is the largest dataset publicly available for metaphor detection, and each word of the dataset is annotated. VUA covers various types of text, including fiction, academic

Benefit of Exploiting Contextual Inconsistency

In this section, we designed the ablation study to evaluate the impact of contextual inconsistency of the proposed SEQ-CI model for metaphor identification. Specifically, we removed the contextual inconsistency term from loss function of our proposed model, which resulted in a reduced SEQ model. Table 8 shows the comparison results of SEQ-CI and SEQ models on VUA. Clearly, exploiting the semantic contextual inconsistency regularization term in learning improves the performance of metaphor

Conclusion

In this paper, we have developed a new sequence labeling model, SEQ-CI, which can learn from the semantic contextual inconsistency for improvement of metaphor identification. We conducted extensive experiments not only on publicly available English data but also on Chinese metaphor data, and the experimental results show the benefit of SEQ-CI over all the seven well-established baselines. The proposed metaphor identification model may provide support for downstream tasks in natural language

CRediT authorship contribution statement

Xin Chen: Conceptualization, Data curation, Formal analysis, Methodology, Software, Writing - original draft, Writing - review & editing. Zhen Hai: Conceptualization, Formal analysis, Methodology, Writing - review & editing. Suge Wang: Conceptualization, Formal analysis, Funding acquisition, Methodology, Supervision, Writing - review & editing. Deyu Li: Conceptualization, Formal analysis, Funding acquisition, Methodology, Supervision, Writing - review & editing. Chao Wang: Methodology. Huanbo

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

The authors would like to thank all anonymous reviewers for their valuable comments and suggestions which have significantly improved the quality of this work. The work is supported by the National Key Research and Development Project of China (2018YFB1005103), and the National Natural Science Foundation of China (61632011, 61772302).

Xin Chen received the M.S. degree in 2016 and is currently pursuing the Ph.D. degree with the School of Computer and Information Technology, Shanxi University, Shanxi, China since 2016. Her current research interests include natural language processing, text sentiment analysis.

References (41)

  • E. Shutova

    Design and evaluation of metaphor processing systems

    Comput. Linguist.

    (2015)
  • Y. Wang

    Cognitive Linguistics

    (2006)
  • G. Lakoff et al.

    Metaphors We Live by

    (2008)
  • B.B. Klebanov, C.W. Leong, E.D. Gutierrez, E. Shutova, M. Flor, Semantic classifications for detection of verb...
  • E. Shutova et al.

    Multilingual metaphor processing: experiments with semi-supervised and unsupervised learning

    Comput. Linguist.

    (2017)
  • S. Mohammad et al.

    Metaphor as a medium for emotion: an empirical study

  • H. Haagsma et al.

    Detecting novel metaphor using selectional preference information

  • J. Lederer

    Finding metaphorical triggers through source (not target) domain lexicalization patterns

  • P.D. Turney et al.

    Literal and metaphorical sense identification through concrete and abstract context

  • L. Bulat, S. Clark, E. Shutova, Modelling metaphor with attribute-based semantics, in: Proceedings of the 15th...
  • M. Köper, S.S. im Walde, Improving verb metaphor detection by propagating abstractness to words, phrases and individual...
  • I. Heintz et al.

    Automatic extraction of linguistic metaphors with LDA topic modeling

  • E. Shutova et al.

    Statistical metaphor processing

    Comput. Linguist.

    (2013)
  • I.-H. Chen et al.

    Leveraging eventive information for better metaphor detection and classification

  • S. Rai et al.

    Supervised metaphor detection using conditional random fields

  • H. Jang, Y. Jo, Q. Shen, M. Miller, S. Moon, C. Rose, Metaphor detection with topic transition, emotion and cognition...
  • A. Mosolova, I. Bondarenko, V. Fomin, Conditional random fields for metaphor detection, in: Proceedings of the Workshop...
  • F.U. Jianhui et al.

    Chinese metaphor phrase recognition via combining the clustering and classification

    J. Chin. Inf. Process.

    (2018)
  • S. Sun, Z. Xie, BiLSTM-based models for metaphor detection, in: Proceedings of the National CCF Conference on Natural...
  • E. Shutova, D. Kiela, J. Maillard, Black holes and white rabbits: Metaphor identification with visual features, in:...
  • Cited by (11)

    View all citing articles on Scopus

    Xin Chen received the M.S. degree in 2016 and is currently pursuing the Ph.D. degree with the School of Computer and Information Technology, Shanxi University, Shanxi, China since 2016. Her current research interests include natural language processing, text sentiment analysis.

    Zhen Hai received the Ph.D. degree in computer science and engineering from Nanyang Technological University, Singapore. Zhen is with the DAMO Academy, Alibaba Group. His research interests include natural language processing, text mining, and machine learning.

    Suge Wang received the Ph.D. degree from Shanghai University. She is currently a Professor with the Key Laboratory of Computational Intelligence and Chinese Information Processing of Ministry of Education, School of Computer and Information Technology, Shanxi University. She has published more than 40 papers in international journals. Her research interests include natural language processing, text sentiment analysis, and machine learning.

    Deyu Li received the Ph.D. degree from Xi’an Jiaotong University. He is currently a Professor with the Key Laboratory of Computational Intelligence and Chinese Information Processing of Ministry of Education, School of Computer and Information Technology, Shanxi University. He has published more than 60 papers in international journals. His research interests include artificial intelligence, granular computing, data mining, and machine learning.

    Chao Wang is the CTO of 6Estates. Previously, he was a senior research scientist at Baidu. He holds a PhD in Computer Science from Tsinghua University. His work has appeared in major journals and conferences such as SIGIR, CIKM, TOIS, and IRJ. His main achievements include 2015 SIGIR Best Paper Honorable Mention Award, 2015 Scientific Technology Advance Award of Beijing City (First Prize), and 2016 CIPS Excellent PHD thesis Award. In 6Estates, he led the team to achieve 3rd Position in 2018 Chinese Machine Reading Comprehension competition and 3rd Position in 2019 Chinese Machine Reading Comprehension competition.

    Huanbo Luan is the deputy director and senior research scientist of NExT Search Center at both Tsinghua University and National University of Singapore. He received his B.S. degree in computer science from Shandong University in 2003 and Ph.D. degree in computer science from Institute of Computing Technology, Chinese Academy of Sciences in 2008. His research interests include NLP, multimedia information retrieval, social media and big data analysis.

    View full text