Elsevier

Neurocomputing

Volume 352, 4 August 2019, Pages 33-41
Neurocomputing

Chinese metaphor sentiment computing via considering culture

https://doi.org/10.1016/j.neucom.2019.03.087Get rights and content

Abstract

Metaphors are frequently used to constitute strong emotional expressions within human communication. The sentiment of a metaphor is decided by many factors, including its context, target, cultural background, etc. Previous relevant work mainly focuses on modeling the context of metaphor, however, the importance of cultural factors is not fully discussed, even in cross-lingual systems. Our work builds a system to perform Chinese metaphor sentiment analysis. It considers both context and target of a metaphor, as well as Chinese culture-related knowledge. The system organizes cultural factors in the form of cultural attribute vectors. It models the influence of context and target by using an attention-based Long Short-Term Memory network (LSTM).

Introduction

Metaphors are a common phenomenon in human language. According to Richards [1], on an average, there will be a metaphor in every three sentences during daily communication. One of the critical reasons why metaphors are frequently used is that they can help us to express affects [2]. Tay et al. [3] indicate that metaphors are widely used in case study articles to describe a range of relevant topics including emotions. It has been proved that metaphors constitute stronger emotional expressions than their literal interpretations [4][5]. For instance, “She looked at him sweetly” is more emotional than “She looked at him kindly” [4]. These facts make the metaphor an unignorable object in affective computing.

Previous studies noticed that metaphors always display sentiment implicitly [6], [7], [8]. This means a metaphor tends to contain a few emotional words and hides the affective meaning [9] behind the literal explanation. These characteristics challenge the traditional dictionary-oriented methods of sentiment analysis. Facing this problem, Zhang et al. [10] manually summarizes the syntactic rules of metaphors and uses syntactic information to augment lexical dictionary knowledge. Rentoumi et al. [8], using contextual information, captures metaphorical semantics through a traditional machine learning method. Strzalkowski et al. [11] proposes a logic reasoning system based on the relationships between the source and target of the metaphor. Peng et al. [12] proposes a model that described the interaction of metaphors based on attention mechanism with a standard LSTM. They thought that the interaction of the target and the context decided the affective semantics of metaphor. Related work reveals the significance of context, the interaction of metaphor components, and manually built knowledge; however, the influence of cultural factors has not been fully detected. Our work holds that cultural background sometimes can even determine the computed results. Ma et al. [13] attached much importance to the impact of commonsense knowledge on sentiment analysis and incorporated external knowledge into their model by using extended LSTM cells. As a kind of commonsense knowledge [14] [15], cultural knowledge is important in the analysis of metaphor sentiment. For example, the “dragon”, which tends to be seen as evil and aggressive in western culture, is regarded as royal and holy in Chinese culture. When using dragon-related terms in a metaphor, such cultural difference can lead to analytical results that express opposite sentiment.

Analysis of metaphorical sentiment is culturally-related, largely because a metaphor, itself, has a deep cultural root. First, a metaphor is a kind of linguistic phenomenon; whereas, language is a carrier of culture [16]. Text, generated in a specific language environment, is inevitably encoded with cultural features. Furthermore, according to the widely accepted cognitive metaphor theory, conceptual metaphor theory [17], a metaphor is not merely a linguistic phenomenon, but also a cognitive mechanism of humans. The theory argues that a metaphor enables us to gain knowledge of a new concept (target domain) by referring to the experience of an old one (source domain), in which case the “experience” connects the two domains. Note that such experience is acquired within the cultural background involving almost every aspect of the life: religion, custom, geography, history, and society [18][19]. Thus, it is necessary to consider cultural factors when analyzing the sentiment of metaphors.

To organize cultural factors into the metaphor sentiment analysis system, we propose the idea of “cultural attribute mapping” based on the conceptual metaphor theory (see Section 3). The “cultural attribute mapping” mechanism describes how culture-related life experiences connect the target and source domains of a culture-related conceptual metaphor. In this paper, the knowledge of a culture-related life experience is revealed as culture-related attributes and then transformed into culture-related attribute embedding, which constitutes a part of the input vector of an attention-based LSTM neural network. We use the attention mechanism and the LSTM to model the effects of the target and the context of the metaphor. Our proposed method is thus named “Culture-Related Attention-Based Long Short-Term Memory” network (CRAB-LSTM). Our proposed CRAB-LSTM model performs binary (positive and negative) sentiment classification and is evaluated on a Chinese metaphor corpus. Experimental results justify the effectiveness of our proposal.

The main contributions of this work are as follows:

  • 1.

    This article proposes the CRAB-LSTM to detect binary sentiment of metaphors. The significance of cultural factors in metaphor sentiment analysis is emphasized. Our method presents a possible way to organize culture-related knowledge into a metaphor sentiment analysis system.

  • 2.

    Our experimental results show that the sentiment of a metaphor is influenced by the interaction of the target, context, and cultural factors, which is a finding that has not been discussed in previous studies.

The rest of the article is organized as follows: Section 2 reviews related work in metaphor sentiment analysis. Section 3 outlines our idea and introduces related concepts. Section 4 describes the CRAB-LSTM model in detail. Section 5 presents experimental results and an analysis. Section 6 concludes the article and outlines future work.

Section snippets

Related work

Basically, there are two dominant methods in sentiment analysis, i.e. knowledge-based method and machine learning method. Besides, method that combines knowledge with neural network has also gained popularity. In this section, related works of the three methods are introduced. Particularly, we pay additional attention to metaphor sentiment analysis.

Theoretical background

This section describes our idea of combining context, target, and culture-related information into a metaphor sentiment analysis system. To explain our idea, some concepts are defined.

The CRAB-LSTM for metaphor sentiment analysis

Word embedding combined with culture-related attribute embedding is used as the input of our system, which uses an attention mechanism to model the interaction between the target and the context and relies on LSTM to produce semantic representations. The structure of our model is shown in Fig. 4.

Experiment

The proposed CRAB-LSTM model was applied to Chinese metaphor sentiment analysis. Chinese metaphor sentiment analysis is still a relatively new area. Furthermore, there are few study in metaphor sentiment analysis considering culture. We built up several LSTM-based methods and used different combinations of context, target, and cultural factors as baselines. Besides, we introduce some methods of previous study as baselines, including TRAT-LSTM [12], TD-LSTM [30] and RAM [33]. The F1 metric and

Conclusion

This work proposed the idea of combining culture-related knowledge into metaphor sentiment analysis system and used the interaction of the target and the context to calculate the semantic representation of a metaphorical context. We constructed a culture-related attribute dictionary to serve as the source of the cultural factors. The culture-related attributes were transformed into vectors and combined with word2vec word embedding as the system input. The interaction between the target and the

Conflict of interest

There is no conflict of interest to this work.

Acknowledgment

This work was supported by National Natural Science Foundation of China (Project 61075058). The authors wish to thank Mr. Michael McAllister for his valuable assistance in proofreading this paper.

Chang Su received her Ph.D. degree from Xiamen University, Xiamen, China, in 2008. She is now an associate professor in the Cognitive Science department of Xiamen University. Her research interests include natural language understanding, metaphor computation and Affect computing.

References (56)

  • TangD. et al.

    Document modeling with gated recurrent neural network for sentiment classification

    Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing

    (2015)
  • I. Richards

    The philosophy of rhetoric

    Philos. Rev.

    (1937)
  • K. Ahmad et al.

    What is new? news media, general elections, sentiment, and named entities

    Proceedings of the Workshop on Sentiment Analysis where AI meets Psychology (SAAIP 2011)

    (2011)
  • D. Tay

    Metaphor in case study articles on chinese university counseling service websites

    Chin. Lang. Discourse

    (2015)
  • F.M. Citron et al.

    Metaphorical sentences are more emotionally engaging than their literal counterparts.

    J. Cognit. Neurosci.

    (2014)
  • S. Mohammad et al.

    Metaphor as a medium for emotion: an empirical study

    Joint Conference on Lexical and Computational Semantics

    (2016)
  • E. Riloff et al.

    Learning extraction patterns for subjective expressions

    Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing

    (2003)
  • J. Wiebe et al.

    Annotating expressions of opinions and emotions in language

    Lang. Resour. Eval.

    (2005)
  • V. Rentoumi et al.

    Investigating metaphorical language in sentiment analysis: a sense-to-sentiment perspective

    Acm Trans. Speech Lang. Process.

    (2012)
  • G.N. Leech

    Semantics

    [M]. Harmomdsworth: Penguin

    (1974)
  • ZhangL.

    Metaphor interpretation and context-based affect detection

    International Conference on Computational Linguistics: Posters

    (2010)
  • T. Strzalkowski et al.

    Computing affect in metaphors

    The Workshop on Nlp in Metaphor

    (2014)
  • PengY. et al.

    Chinese metaphor sentiment analysis based on attention-based lstm

    2018 Tenth International Conference on Advanced Computational Intelligence (ICACI),

    (2018)
  • MaY. et al.

    Targeted aspect-based sentiment analysis via embedding commonsense knowledge into an attentive lstm

    Thirty-Second AAAI Conference on Artificial Intelligence

    (2018)
  • ZouX. et al.

    Culture as common sense: perceived consensus versus personal beliefs as mechanisms of cultural influence.

    J. Personal. Soc. Psychol.

    (2009)
  • N. Popovic

    The Synthesis

    (2008)
  • BaiZ.k. et al.

    Chinese verb metaphor recognition based on topic model

    Comput. Technol. Dev.

    (2016)
  • G. Lakoff et al.

    Metaphors we live by

    Ethics

    (1983)
  • B. Shatenstein et al.

    Influences on diet, health behaviours and their outcome in select ethnocultural and religious groups.

    Nutrition

    (1998)
  • G. Gyal

    On social characteristics of Tibetan dance and its social functions

    Tibet. Stud.

    (2011)
  • ZhuoS. et al.

    Chinese text sentiment analysis based on fuzzy semantic model

    2014 IEEE 13th International Conference on Cognitive Informatics & Cognitive Computing (ICCI* CC)

    (2014)
  • PengH. et al.

    A review of sentiment analysis research in chinese language

    Cognit. Comput.

    (2017)
  • WuF. et al.

    Microblog sentiment classification with contextual knowledge regularization.

    AAAI

    (2015)
  • E. Cambria

    Affective computing and sentiment analysis

    IEEE Intell. Syst.

    (2016)
  • WangX. et al.

    Dependency-attention-based lstm for target-dependent sentiment analysis

    Chinese National Conference on Social Media Processing

    (2017)
  • YangM. et al.

    Attention based lstm for target dependent sentiment classification.

    AAAI

    (2017)
  • ZhangZ. et al.

    Textual sentiment analysis via three different attention convolutional neural networks and cross-modality consistent regression

    Neurocomputing

    (2018)
  • WangY. et al.

    Attention-based lstm for aspect-level sentiment classification

    Conference on Empirical Methods in Natural Language Processing

    (2016)
  • Chang Su received her Ph.D. degree from Xiamen University, Xiamen, China, in 2008. She is now an associate professor in the Cognitive Science department of Xiamen University. Her research interests include natural language understanding, metaphor computation and Affect computing.

    Yijiang Chen received his Ph.D. degree from Xiamen University, Xiamen, China. He is now an associate professor in the Computer Science Department of Xiamen University. His research interests include natural language processing and machine learning.

    Junchao Li is currently a master candidate at the School of Information Science and Engineering, Xiamen University. His research interests include natural language processing and metaphor computation.

    Ying Peng received her Master degree from Xiamen University, Xiamen, China, in 2018. Her research interests include natural language processing, metaphor computation and affect computing.

    View full text