Recent advances in neural metaphor processing: A linguistic, cognitive and social perspective

Metaphor is an indispensable part of human cognition and everyday communication. Much research has been conducted elucidating metaphor processing in the mind/brain and the role it plays in communication. in recent years, metaphor processing systems have benefited greatly from these studies, as well as the rapid advances in deep learning for natural language processing (NLP). This paper provides a comprehensive review and discussion of recent developments in automated metaphor processing, in light of the findings about metaphor in the mind, language, and communication, and from the perspective of downstream NLP tasks.


Introduction
Metaphor is a figurative device that allows us to portray one domain, often abstract, in terms of another that is typically more concrete (Lakoff and Johnson, 1980). In the sentences His eyes lightened at the news, or Her mood darkened, the abstract domain of emotion is described in terms of the more concrete domain of brightness. Through cross-domain associations and comparisons, metaphor enables us to communicate complex abstract concepts, to convey our affective states, to reinforce our arguments and, ultimately, to forge entirely novel word meanings. Metaphor is also ubiquitous in language, occuring on average in every three sentences, according to corpus-linguistic research (Steen et al., 2010b;Shutova, 2011). This makes accurate interpretation of metaphorical language essential for many NLP tasks and applications.
A systematic and comprehensive survey of metaphor processing systems was published five years ago (Shutova, 2015). Since then, automated metaphor processing has shifted focus towards neural approaches and there has been much activity in the area, including two shared tasks (Leong et al., 2018(Leong et al., , 2020. The development has kept pace with the development of deep learning techniques, and the state-of-the-art performance has been advanced very quickly, stressing the need for a new survey focusing on this recent progress. Moreover, the design and evaluation of metaphor processing systems should be informed by both the cognition of metaphor and its role in communication, as well as the potential applications of these systems. It is common practice in the NLP community to divide metaphor processing into two sub-tasks, metaphor identification and interpretation, and the two shared tasks encourage the use of a common task definition and particular metaphor corpora for the development of metaphor identification systems. However, there has been little discussion about whether or how well the existing systems address theoretical and cognitive considerations in the use of metaphor and how insights and models produced by this work could inform and benefit the wider natural language understanding (NLU) research.
This paper aims to provide a starting point for such a discussion, by reviewing recent advances in metaphor processing in the context of contemporary theories of metaphor, its sociolinguistic aspects and the needs of NLP applications. We begin with an overview of theories and findings about metaphor in the mind/brain, language, and communication. We then review metaphor processing approaches published in the last 5 years, as well as the datasets they used. We provide an analysis of these models from the linguistic and cognitive perspective and propose some future directions for modelling social aspects of metaphor use (e.g. the intent behind metaphor use, extended metaphor and metaphoric framing), as well as the underaddressed tasks of metaphor interpretation and generation. Finally, we discuss how elements of metaphor processing can be integrated in the real-world applications, such as machine translation, opinion mining, dialogue modelling, and modelling argumentative discourse (e.g., political discourse).

4674
2 Theoretical accounts of metaphor

Conceptual metaphors
According to the conceptual metaphor theory (CMT), linguistic metaphors have their roots in conceptual metaphors, cross-domain mappings in one's conceptual system (Lakoff and Johnson, 1980). Based on comparable properties and relations in the target domain and the source domain, conceptual metaphors invite one to conceptualise the former through the latter. For example, He attacked every point in her argument instantiates the conceptual metaphor ARGUMENT IS WAR, in which ARGUMENT is the target domain and WAR the source domain. It uses an event in the domain of WAR to describe an event about ARGUMENT. The target and source domains of a metaphor usually involve abstract and concrete concepts respectively, which has to do with concept representation in the brain. According to the theory of embodied cognition (Barsalou, 1999), concepts are represented within sensorimotor circuitry: GRASP-ING, for instance, is represented in areas that control hand movements (Gallese and Lakoff, 2005). When using expressions such as grasp a point, the same areas are involved for processing the metaphor (Aziz-Zadeh and Damasio, 2008).
Conceptual metaphors that are entrenched in one's conceptual system are termed conventional metaphors; those that are not entrenched are termed novel metaphors. ARGUMENT IS WAR is a typical conventional metaphor; the expression attack one's argument, for example, has entered contemporary dictionaries of English and is unlikely to be considered novel by the native speakers. On the contrary, Fear coiled around his heart instantiates a novel metaphor, FEAR IS A SNAKE. The verb has a single sense in the Macmillan dictionary: a particular movement of 'something long and thin'. Fear is an abstract concept, incapable of any movement to be seen. The usage is therefore novel.

Linguistic metaphors
Informed by CMT, Steen et al. (2010b) proposed the Metaphor Identification Procedure Vrije Universiteit (MIPVU) for manually identifying metaphorrelated words (MRWs) in a text. MRWs are word uses that can be attributed to underlying conceptual metaphors. What are usually called metaphorically used words are a subset of MRWs termed indirect metaphors. The source-domain word in attack one's argument is an indirect metaphor: it refers to a target-domain act, and is therefore indirectly related to the underlying conceptual metaphor, AR-GUMENT IS WAR.
Other types of MRWs are direct metaphors, implicit metaphors, and metaphor flags. Direct metaphors refer to source-domain entities or events. They often co-occur with metaphor flags, signals that a metaphor is used. In (1), the linking verb is signals a use of the BEAUTY AS FLOWER metaphor. The noun flower is a direct metaphor, referring directly to a source-domain concept. (1) Beauty is but a flower / which wrinkles will devour; [..] (Thomas Nashe) Implicit metaphors are substitutions or ellipses that are directly or indirectly related to underlying conceptual metaphors. In (2), for example, the pronoun that co-refers with the preceding metaphorically used antidote; the determiner this in the following sentence is also used metaphorically: it refers to what is talked about in the previous sentence, which is an abstraction of its basic, physical meaning (Steen et al., 2010b). Metaphorical usage of substitutions are important for rendering cohesive discourse (Steen et al., 2010b). (2) Fortunately, there is a single antidote effective against both these myths; and that is to start all over again . . . . This antidote is effective against the romantic-individualist myth . . . . (BNC) Discourse-level information is essential for identifying extended metaphors, sustained use of the same metaphors in a discourse fragment. A typical example is the 'All the World's a Stage' speech written by William Shakespeare. As is presented below, the speech begins with a WORLD AS STAGE metaphor, and proceeds with exploring various target-source pairs within the metaphor. Note that the last two lines could be mistaken as literal if presented as independent sentences. (3) All the world's a stage, / And all the men and women merely players: / They have their exits and their entrances; / And one man in his time plays many parts, [..]

Metaphor comprehension
There has been a debate about whether metaphor comprehension is a comparison or a categorisation process. According to the comparison view (e.g., Ortony, 1979), when one processes (1), for example, one searches for the shared properties and/or relational structures of the two domains, such as being pleasing to the eye, and the possibility of being fragile. Scholars advocating the categorisation view (e.g., Glucksberg and Keysar, 1990), however, would argue that the object, a flower, refers to a super-ordinate category that both BEAUTY and FLOWER belong to. The Career of Metaphor Theory (Bowdle and Gentner, 2005) suggests that both comparison and categorisation are possible paths of metaphor processing; which path is chosen depends on an interaction between the conventionality of the metaphor and its linguistic realisation. More specifically, novel metaphors are processed through comparison; as a metaphor becomes conventionalised, people tend to process it through categorisation, which is less cognitively demanding than a comparison process (Bowdle and Gentner, 2005).
An inevitable result of metaphor processing is the emergence of features not inherent in the target or the source domain (Tourangeau and Rips, 1991). This notion of emergent meaning corresponds to the connotations or inference of metaphorical language in linguistics literature. Consider the difference in meaning between attack one's argument and criticise one's argument. The first contains far richer shades of meaning than the second. It might be possible to express a similar meaning without using non-literal language, but the expression is unlikely to be as concise as the use of a single word, attack.

Metaphor use in communication
The linguistic expression of emotional states often employs metaphor (Fainsilber and Ortony, 1987;Fussell and Moss, 1998). As emotion is an abstract domain, it goes with CMT and embodied cognition that we employ more concrete domains, such as physical or bodily experience, to conceptualise it. Moreover, since metaphor gives rise to emergent meaning, metaphorical language has a stronger emotional effect than literal language, regardless of the source and target domains involved (Blanchette et al., 2001;Crawford, 2009;Mohammad et al., 2016). For instance, Citron and Goldberg (2014) found that metaphorical expressions involving taste (e.g., She looked at him sweetly), evoke a higher emotional response than their literal counterparts (e.g., She looked at him kindly).
Metaphor has also proved to be an effective persuasive device (Sopory and Dillard, 2002;van Stee, 2018). The persuasive power of metaphors is pronounced in metaphoric framing effect. Since metaphors encourage a particular way to conceptualise the target domain, repeated use of the same metaphors throughout discourses in mass media tends to affect how the public perceives and reacts to societal issues that belong to the target domain (Lakoff, 1991;Entman, 2003;Lakoff and Wehling, 2012). For instance, participants in a series of studies Boroditsky, 2011, 2013) favoured different social solutions to crime after reading articles that associate CRIME with different source domains. Moreover, while the participants could identify the implicitly advocated solutions given a metaphor, they were unaware of the influence of the metaphors on their own preference.
The Deliberate Metaphor Theory is an attempt to deal with the intention behind metaphor use systematically (Steen, 2008(Steen, , 2017. The theory defines deliberate metaphor use as the intentional introduction of a topic shift or perspective change to the discourse. Deliberate metaphor use is associated with online metaphor processing, the construction of conceptual metaphors during text comprehension (Steen, 2017). Examples of deliberate metaphors include conventional metaphors instantiated as copula metaphors and all novel metaphors. A systematic procedure for the identification of potential deliberate metaphors has also been proposed (Reijnierse et al., 2018).

Metaphor datasets
Metaphoricity annotations Tsvetkov et al. (2014) released a dataset (henceforth: TSV) consisting of an equal number (884) of metaphorical and non-metaphorical adjective-noun (AN) phrases collected from the web. The phrases were stated to be verified by multiple annotators, but the criteria for metaphor annotation were not provided.
The dataset released by Mohammad et al. (2016) (henceforth: MOH) consists of 1639 (1230 literal and 409 metaphorical) sentences extracted from WordNet, manifesting the use of 440 verbs. Metaphoricity annotation of the verb uses was obtained through crowdsourcing. Note that the Word-Net sentences are mainly instances of conventional metaphor. The specified association between word senses and metaphoricity makes it easier to determine the source and target domains involved.
The Language Computer Corporation metaphor datasets (LCC) (Mohler et al., 2016) contain linguistic metaphors extracted from web corpora for a small set of target domains in four languages: English, Spanish, Russian, and Farsi. Metaphoricity is annotated at sentence level, and the conceptual metaphors and affect information is specified.
VU Amsterdam Metaphor Corpus (VUA) (Steen et al., 2010b) is the fruit of the authors' application of MIPVU. The metaphoricity of each lexical unit (187,570 in total) in a subset of the British National Corpus (BNC Consortium, 2007) was annotated. VUA is the only metaphor corpus used in studies of automated metaphor identification that is built by cognitive linguists, and the only one that deals with the metaphoricity of function words. Mohammad et al. (2016) also obtained literal paraphrases of the sentences, in which the metaphorically used verbs are replaced by their synonyms (171 pairs of sentences in total), selected by the authors. Note that the literal paraphrases were considered to convey less emotion than the original metaphorical sentences in their experiment. It is therefore questionable to what extent the paraphrases capture the connotations of the metaphorical sentences.

Metaphor paraphrase datasets
Bizzoni and Lappin (2018) built a metaphor paraphrase dataset containing 200 sets of 5 sentences; 4 paraphrases at varying levels of aptness are provided for each metaphorical sentence. Apart from verbs, the dataset also includes metaphorical uses of adjectives, copula metaphors, and multi-word metaphors. The dataset takes into account the connotations of the metaphorical sentences to some extent. For instance, candidate paraphrases for the copula metaphor My job is a dream include I love my job and I hate my job, which indicate opposite sentiment poles. A metaphor processing system will need to infer the sentiment to select the apt paraphrase. Most neural models treat metaphor identification as a sequence labelling task, outputing a sequence of metaphoricity labels for a sequence of input words (usually a sentence) (Bizzoni and Ghanimifard, 2018;Dankers et al., 2019;Gao et al., 2018;Gong et al., 2020;Mao et al., 2019;Mykowiecka et al., 2018;Pramanick et al., 2018;Su et al., 2020;Wu et al., 2018). The first sequence labelling systems typically represented an input sentence as a sequence of pre-trained word embeddings and produced a task-and context-specific sentence representation through bidirectional long short-term memory (BiLSTM) (Dankers et al., 2019;Gao et al., 2018;Mykowiecka et al., 2018;Pramanick et al., 2018). Bizzoni and Ghanimifard (2018) experimented with separating long sentences into smaller chunks, which led to a 6% increase in F-score when using a BiLSTM architecture. Their BiLSTM system outperformed their compositional system, which employs a sequence of fully-connected neural networks (NNs) and essentially performs bigram phrase composition to modulate the representation of input words with respect to their neighbours. BiLSTM models also outperformed bidirectional gated recurrent unit (Bi-GRU) models in the study of Mykowiecka et al. (2018). From Gao et al. (2018) (Gong et al., 2020). A summary of the best performing recent metaphor identification systems is presented in Table 1 (a summary of all systems can be found in Appendix A).
Several BiLSTM-based systems consider both contextualised and pre-trained representations in the classification layers (Mao et al., 2019;Swarnkar and Singh, 2018). The Di-LSTM Contrast system (Swarnkar and Singh, 2018) encodes the left-and right-side context of a target word using forward and backward LSTMs. The classification is based on a concatenation of the target-word vector and its difference with the encoded context. Mao et al. (2019) combined GloVe and BiLSTM hidden states for sequence labelling, which outperformed the best model in the 2018 VUA All POS track. Wu et al. (2018) and Su et al. (2020) employed separate encoding of local and long-range context. Wu et al. (2018) used a convolutional neural network (CNN) and a BiLSTM to extract local and sentence context respectively. Su et al. (2020) used separate Transformer encoder layers to encode global and local text features for each word. The two systems achieved the best performance on the VUA All POS tracks in their respective shared tasks (Leong et al., 2018(Leong et al., , 2020. Modelling metaphor in discourse Several recent approaches have also incorporated wider discourse properties in their models. Mu et al. (2019) focused on the metaphoricity of verbs. They used general-purpose word, sentence and document embedding methods (e.g. GloVe, ELMo, doc2vec (Le and Mikolov, 2014), skip-thought (Kiros et al., 2015)) to represent the surrounding paragraphs. Their system feeds into a gradient boosting decision tree classifier (Chen and Guestrin, 2016) a concatenation of three feature vectors, representing 1) the lemma of the target verb, 2) its subject and direct object, and 3) its surrounding paragraph. All the representations are learned from embedding methods. Representing the features with ELMo led to the highest F-score (0.668), using the VUA Verbs 2018 shared task data.
Dankers et al. (2020) fine-tuned a BERT model that receives a discourse fragment as input. Using hierarchical attention (which computes both token-and sentence-level attention) after the encoded layers achieved better performance than applying general attention to all tokens. Both Dankers et al. (2020) and Mu et al. (2019) thus demonstrated the importance of context beyond sentence for word-level metaphor identification. Their qualitative analysis shows that co-reference resolution is one of the driving factors in the performance increase.
Note that the above systems use discourse to aid in detecting linguistic metaphors whose metaphoricity is otherwise ambiguous, but do not monitor whether a metaphor is sustained throughout a fragment of discourse. Thus, they are unlikely to be directly applicable to the identification of extended metaphors or metaphoric frames.

Cognitively-inspired approaches
Categorial features Metaphor processing is concerned with how concepts are organised in the brain/mind, and is closely related to categorisation. It therefore makes sense to employ categorial features for metaphor identification. Tekiroglu et al. (2015) tested the use of sensorial categories (the five human senses) for identifying AN synaes-thetic metaphors (e.g., sweet music, soft light). Using sensorial categories in addition to Word-Net supersenses, concreteness, and imageability led to improved performance (accuracy 0.890 vs 0.845 on TSV). Mykowiecka et al. (2018) used another general-purpose resource, the Harvard IV psychosocial dictionary 1 , which includes categories of emotions, people and animals, objects, places, etc. However, it did not lead to consistent improvement in model performance. Bulat et al. (2017) compared property-based and linguistic embeddings for input word representation. They obtained property-based word embeddings by mapping linguistic word embeddings onto a conceptual space, using as training data a frequency-based human property-norm dataset (McRae et al., 2005). Using property-based word embeddings led to a 4% increase in F-score on TSV.
Sensory features An important function of conceptual metaphors, according to CMT, is to use bodily experience to understand abstract concepts; concreteness features have therefore also proved useful for automated metaphor identification (Bizzoni and Ghanimifard, 2018;Turney et al., 2011).  tested combination of visual and linguistic embeddings on MOH and TSV. The multimodal system outperformed the monomodal systems in both tests. Gong et al. (2020) also included both categorial and sensory features. In addition to RoBERTa, the system employs concreteness features, topic distributions, WordNet classes, VerbNet classes, verb clusters, and part of speech (POS), which led to improvements in performance on VUA All POS.
Word-context incongruity The neural model of Swarnkar and Singh (2018) computes the difference between a target word and its context. This operation can be associated with the comparison view of metaphor, including CMT, and is reflected in MIPVU.  used cosine similarity between word or phrase embeddings to predict the metaphoricity of verb-noun (VN) or AN pairs; a word pair is marked metaphorical if the cosine similarity is below a trained threshold. The systems reached an F-score of 0.71 on MOH and 0.76 on TSV using linguistic embeddings alone. Rei et al. (2017) proposed a supervised similarity network, which learns to calculate weighted co-    (Mehrabian, 1996), and each of these was considered separately in a MTL setup. The best performance was achieved with the BERT architecture. Dankers et al. (2019) found that while predicting dominance was the most challenging task on the emotion side, it also provided the greatest and most consistent improvements to metaphor identification, and vice versa.

Automated metaphor interpretation
Recent research on automated metaphor interpretation mainly followed Shutova (2010) in treating the problem as a paraphrasing task. Su et al. (2017) proposed a property transfer process for the interpretation of copula metaphors in Chinese and English. Given a target-source pair, the system extracts source-domain properties from hand-crafted databases 2 , represented as adjectives. It then selects the property that contributes the most to the semantic relatedness of the target-source pair. The resultant pair of target-domain word and property is taken as interpretation of the copula metaphor. The metaphor LOVE IS TIDE, for instance, was interpreted as The love is unstoppable. Su et al. (2017) thus took into account the emergent meaning of metaphors. Note that the literal explanations they obtained can be regarded as explanations of conceptual metaphors as well.
The above-mentioned system of Mao et al. (2018) performs both metaphor interpretation and identification: the synonym or hypernym of the metaphoric target word that matches the context can be considered the literal counterpart of the target word. The output interpretations were evaluated through a English-Chinese machine translation task: words classified as metaphorically used were paraphrased prior to translation. The system improved the accuracy of both Google Translation (0.60 vs 0.34) and Bing Translation (0.66 vs 0.42) on the metaphorical sentences in MOH. The experiment thus demonstrated the value of metaphor interpretation for machine translation.
Bizzoni and Lappin (2018) presented a neural model that detects paraphrases of sentences containing metaphor use. Given a metaphorical sentence and a candidate paraphrase, the system uses parallel CNN-LSTM blocks to encode the two sentences separately. The sentence representations are then merged and passed through fully-connected layers to produce a prediction. The system reached an F-score of 0.746 in a binary classification task and a Pearson correlation of 0.553 in a paraphrase ordering task, on the authors' own dataset.

Discussion
Metaphor identification Word-level metaphor identification is the most studied topic in automated metaphor processing. The systems utilise the latest NLP techniques and findings about metaphor in cognitive sciences; the state-of-the-art performance has been advanced quickly. A limitation in these studies, however, is that the systems do not deal with different types of metaphors. The high percentage of conventional metaphors in VUA (Steen et al., 2010a,b) implies that systems trained on the corpus may not capture conventional and novel metaphor uses equally well. It is the processing of novel metaphors that is a particular challenge for NLU systems, and the current experimental setup does not allow us to explicitly benchmark metaphor identification systems in this ability. The same goes for deliberate metaphor use, which is essential for modelling metaphor in communication and applies to both conventional and novel metaphors.
Whilst there have been huge advances in perfor-mance in word-level metaphor identification, the broader tasks of identifying conceptual metaphors, extended metaphors, and metaphoric framing, have been largely ignored. Conceptual metaphors form the framework for both extended metaphors and metaphoric framing, from a cognitive perspective, and these latter two are a key part of the intentional use of metaphor in communication.
Metaphor interpretation Recent studies on automated metaphor interpretation largely focused on literal paraphrases of metaphorical sentences. The inference of emergent meaning is, however, an integral part of metaphor comprehension and should be captured by metaphor interpretation systems. To capture emergent meaning is to reveal the properties and relations of the target domain emphasised by the metaphor and any additional connotations and emotional effects that arise in the interpretation process. In this respect, the existing paraphrase datasets (Shutova, 2010;Mohammad et al., 2016) are unlikely to reflect the emergent meaning of the metaphors. Bizzoni and Lappin (2018) consider the sentiment poles of metaphorical expressions, but it is questionable to what extent the apt paraphrases also reflect other dimensions, such as the intensity of an emotion.
One way to deal with the inference of linguistic metaphors is to convert it to the task of paraphrasing copula metaphors. As is the case with Su et al. (2017) and Bizzoni and Lappin (2018), the paraphrasing of copula metaphors has been dealt with as obtaining the property of the target domain that is indicated by the metaphor. We can thus deal with the inference of any linguistic metaphor by first identifying its conceptual metaphor and then treating the conceptual metaphor as a copula metaphor. To tackle the task, we will need datasets of the emergent meaning of a wide range of conceptual metaphors. Such datasets can be built from data obtained in psycholinguistic studies about emergent meaning, in which participants were asked to provide concepts or properties that were related to metaphors as well as their target and source domains alone (e.g., Gineste et al., 2000). A more comprehensive model of metaphor interpretation would also consider the complexity of linguistic metaphors, which essentially arises from the social aspects of metaphor use, such as the various roles that metaphor plays in communication.
Metaphor generation To the best of our knowledge, automated metaphor generation has largely remained an unexplored area. As metaphor is an important tool for reasoning and linguistic expression, the task is a crucial step toward the generation of human-like language. Note, however, that despite the scarcity of explicit attempts at metaphor generation, it is likely to be an inherent part of existing natural language generation (NLG) systems already. Metaphorical use of function words, for instance, are unavoidable for linguistic expression. Highly conventionalised metaphorical use of some content words may also be captured by current NLG techniques. For instance, the use of bright for the description of colours (e.g., bright blue eyes) is considered metaphorical in VUA whereas it is one of the most frequent usages of the adjective. It is likely that data-driven NLG (including neural models) instantiate this usage of the word without 'knowing' that it is a metaphor.
It is thus in the intentional use of metaphors that the difficulty of metaphor generation lies. The development of metaphor generation systems therefore relies on the modelling of metaphor use in communication, which we discuss below.
Implications for other NLP tasks That linguistic metaphors have to do with concepts in two different domains implies that their automated processing is potentially indispensable for resolving referents in text. This is important for a number of NLP applications, such as document summarisation, question answering, and machine translation. The resolution of the actual referents of MRWs is mainly concerned with indirect metaphors, which use source-domain words to refer to target-domain concepts. Note that implicit metaphors can be indirect metaphors as well. It is therefore necessary to take the metaphoricity of function words such as pronouns and determiners into account when it comes to resolving the referent of metaphorically used words. We suggest that the processing of implicit metaphors, especially substitutions, can be facilitated by combining metaphor identification with coreference resolution, as the metaphorically used pronouns essentially refer to the same referents as a preceding metaphorical expression.
Furthermore, there has not been much discussion of the form of metaphor interpretation that would be most beneficial for downstream NLP tasks. Most research on automated metaphor interpretation focuses on producing human-readable literal paraphrases. However, it is likely that internal model representations that capture the rich meaning of metaphorical language can be more elegantly and successfully integrated in the task models. Consider, for instance, machine translation. After obtaining an internal representation of a metaphorical sentence in the source language, the system can then search for a literal or metaphorical expression in the target language that best matches the representation. We suggest that the development of such a system can begin with investigating the hidden states of existing deep learning models of metaphor identification: the vectors used to output metaphoricity labels may already capture the emergent meaning of the input sentences.
Metaphor in communication In addition to the semantic aspects, metaphor use often also has social implications. Metaphor processing is an essential component of opinion mining: since metaphorical language tends to express stronger emotions than literal language, the presence of metaphors and their inferences would contribute to the sentiment conveyed by a text. It is also worth considering distinguishing deliberate and non-deliberate metaphor uses: focusing on the former may achieve a better approximation of the speaker/writer's sentiment and communicative goals.
Metaphor processing should thus also be incorporated into dialogue modelling. Metaphors could be extended or even evolve with the turn taking in a conversation, as speakers seek mutual understanding, or the resolution of a difference of opinion in argumentative discourse such as debates. Systems that are concerned with speaker's intent will therefore benefit from the identification and interpretation of metaphorical language.
Given the relationship between metaphor and persuasiveness, it may be beneficial to model metaphor use in other types of argumentative discourse, such as political speeches and news articles. A recent development in this regard is a MTL approach to modelling metaphor, emotion, and framing in political texts (Huguet Cabot et al., 2020).

Conclusion
In this paper, we reviewed recent advances in automated metaphor processing and brought to light its less explored aspects, pertaining to cognitive processing of metaphor, its role in communication and its connection with downstream NLP applications. We hope that our work will encourage further in-terdisciplinary research on metaphor and facilitate progress in this important and fascinating area.    WordClass Conc CosSim - Table 3: Cognitively-inspired approaches for metaphor identification. Effect notation: positive + ; negative -; not tested. Conc: concreteness. Image: imageability. Emb: embeddings. CosSim: cosine similarity. TopicDiff: cosine similarity between topic distributions.