1. Introduction
Metaphor is more than a rhetorical flourish. It is grounded in our everyday embodied and emotive experience, and drives much of the way we communicate (Gibbs, Reference Gibbs1994; Johnson, Reference Johnson1987; Kövecses, Reference Kövecses2015; Lakoff, Reference Lakoff1987; Lakoff & Johnson, Reference Lakoff and Johnson1980, Reference Lakoff and Johnson1999). We use metaphor to structure our understanding of one domain in terms of another, for instance, time as space (Radden, Reference Radden, Brdar, Omazić, Takač, Gradečak-Erdelijić and Buljan2011), anger as heat (Lakoff & Kövecses, Reference Lakoff, Kövecses and Quinn1987), desire as hunger (Gibbs, Lima, & Francozo, Reference Gibbs, Lima and Francozo2004), and politics as sports (Semino & Masci, Reference Semino and Masci1996). Metaphor can have a dramatic impact on how we understand important social matters, such as crime (Thibodeau & Boroditsky, Reference Thibodeau and Boroditsky2011, Reference Thibodeau and Boroditsky2013), climate change (Flusberg, Matlock, & Thibodeau, Reference Flusberg, Matlock and Thibodeau2017), and elections (Burnes, Reference Burnes2011).
Extensive research has examined metaphor across genres, especially in English. Manual methods, including qualitative analyses that use hand-annotation of texts, and corpus methods have been used to analyze when and how metaphor is used. The Metaphor Identification Procedure (Pragglejaz Group, 2007; Steen, Reference Steen, Gibbs and Steen1999; Steen, Biernacka, Dorst, Kaal, López-Rodríguez, & Pasma, Reference Steen, Biernacka, Dorst, Kaal, López-Rodríguez, Pasma, Low, Todd, Deignan and Cameron2010a; Steen, Dorst, Herrmann, Kaal, & Krennmayr, Reference Steen, Dorst, Herrmann, Kaal and Krennmayr2010b), is extensively used in the broader metaphor community (e.g., Demjén, Semino, & Koller, Reference Demjén, Semino and Koller2016; Demmen et al., Reference Demmen, Semino, Demjén, Koller, Hardie, Rayson and Payne2015; Steen, Dorst, Herrmann, Kaal, Krennmayr, & Pasma, Reference Steen, Dorst, Herrmann, Kaal, Krennmayr and Pasma2010c). The second approach falls within more traditional corpus linguistic methodologies (Deignan, Reference Deignan2005; Lederer, Reference Lederer, Noelle, Dale, Warlaumont, Yoshimi, Matlock, Jennings and Maglio2013; Martin, Reference Martin, Gries and Stefanowitsch2006; Philip, Reference Philip, MacArthur, Oncins-Martínez, Sánchez-García and Piquer-Píriz2004; Stefanowitsch & Gries, Reference Stefanowitsch and Gries2006), which use concordances, frequency counts, keyword analysis, and collocation patterns to identify potentially metaphoric uses of target words.
With more and more attention on scalability to larger texts, a third approach is emerging from computational linguistics. It focuses on devising automated means of metaphor identification across larger or more diverse corpora. Metaphor is a burgeoning area in the field of computational linguistics, evident from the addition of a figurative language and metaphor processing workshop to the annual meeting of the North American Chapter of the Association for Computational Linguistics (NAACL <www.naacl.org>). While specific methodologies and goals vary, most NLP approaches aim for improvements in recall and precision mechanisms for automated linguistic metaphor identification based on internally defined retrieval definitions. The automation is usually exercised over unrestricted texts, texts that are sufficiently general to train a system to identify all types of metaphoric expressions. Some studies use statistical cluster methods (Birke & Sarkar, Reference Birke and Sarkar2006; Shutova & Sun, Reference Shutova and Sun2013; Shutova, Teufel & Korhonen, Reference Shutova, Teufel and Korhonen2012) for metaphor identification. Gutiérrez, Shutova, Marghetis, and Bergen (Reference Gutiérrez, Shutova, Marghetis and Bergen2016) employ compositional distributional semantic vector space models that operate on the composition of lexical representations across large corpora. CorMet (Mason, Reference Mason2004) uses the selectional preference of verbs, and clusters of nodes derived from WordNet senses. There are many other methods explored in computational metaphor detection: neural nets (Do Dinh & Gurevych, 2016), maximum entropy classification combined with hand-annotation of metaphoricity (Gedigian, Bryant, Naryanan, & Ciric, Reference Gedigian, Bryant, Narayanan and Ciric2006), selectional preferences of lexical items (Wilks, Reference Wilks1975, Reference Wilks1978), knowledge-representation models (Martin, Reference Martin1988, Reference Martin1994), and word sense disambiguation-based approaches (Krishnakumaran & Zhu, Reference Krishnakumaran and Zhu2007), among others. Dunn (Reference Dunn2013a, Reference Dunn2013b) describe four systems of metaphor identification and their comparison in performance, while Shutova (Reference Shutova2010) and Neuman et al. (Reference Neuman, Assaf, Cohen, Last, Argamon, Howard and Frieder2013) provide overviews of the current state of the art in metaphor NLP.
The above are just a few of the approaches and technologies available in this quickly expanding subfield of computational linguistics. Such automation approaches are useful because they are agnostic to specific research needs as well as specific textual genres, and are capable of yielding high scores for recall and precision in automated metaphor identification due to their rigorous implementation of lexical and semantic resources (e.g., SOMO ontology in Dunn, Reference Dunn2013a, and WordNet in Lönneker, Reference Lönneker2003). However, they may not be ideal when a metaphor researcher has specific questions about the functions of metaphor in a particular language context or cognitive and social domain, or seeks a bird’s-eye view of metaphor distribution within that domain.
Another area in which existing computer-aided metaphor research could be enhanced is in cross-linguistic studies. With some recent exceptions (Gordon, Hobbs, May, Morbini, & Vista, Reference Gordon, Hobbs, May, Morbini and Vista2015; Levin et al., Reference Levin, Mitamura, Fromm, MacWhinney, Carbonell, Feely, Frederking, Gershman and Ramirez2014; MacWhinney & Fromm, Reference MacWhinney and Fromm2014; Mohler, Tomlinson, & Rink, Reference Mohler, Tomlinson and Rink2015; Tsvetkov, Boytsov, Gershman, Nyberg, & Dyer, 2014), Footnote 2 metaphor NLP pipelines are created for detecting metaphor in one language, usually English. This constraint results from the nature of the lexical resources on which they are trained, as English is a language for which lexical resources, and NLP resources in general, are more developed relative to other languages.
MetaNet, the system of interest here, can search for cross-linguistic metaphor distribution in texts and narrow down the search to a particular domain of interest. It seeks to form a more symbiotic relationship between powerful computational models on the one hand, and on the other, methods of studying metaphor that have as their goal an understanding of cognitive and social realities. It is not intended to be a general NLP system like the ones described above, rather, a tool that helps metaphor analysts carry out large-scale studies, for instance, by narrowing the search according to target domain and providing cross-linguistic metaphor distribution. The ideal user is a researcher seeking an initial representation of metaphor distribution in specific domains (for instance, as shown here, in domains of social concern such as cancer and poverty) over large texts, and possibly over multilingual corpora. The distribution obtained would not paint an exhaustive picture of metaphor in those texts, as the tool cannot detect every possible metaphor, but it would be representative enough to enable the researcher to pursue particular hypotheses, or to focus on particular subdomains or metaphor families and text genres for further probing. Sweetser, David, and Stickles (in press) provide a thorough comparison of MetaNet with other automated metaphor detection systems, including those listed above.
Each automated identification run begins with a specific question issued by a researcher, either about how a target domain is talked about metaphorically (via target domain search, e.g., cancer), or how a source domain is used across several target domains (via a source domain search, e.g., Forward Motion or Physical Combat). Further, repeated iterations of MetaNet-aided metaphor extraction help augment its knowledge base, such that an increasing number of metaphors across an increasing number of domains (and languages) are accumulated and reused in future iterations. Gold standard annotations were used at multiple junctures to evaluate the system for precision and recall scores, and to reduce the detection of false positives over time (see Hong, Reference Hong2016). However, since the system is only trained to detect metaphors in a finite set of grammatical patterns (Dodge, Hong, & Stickles, Reference Dodge, Hong and Stickles2015; Stickles, David, Dodge, & Hong, Reference Stickles, David, Dodge and Hong2016) and for a finite set of metaphor families, some occurrences of domain-relevant metaphor may not be detected. For instance, give cancer the boot would not be detected because the ditransitive construction is not yet represented in the system. Nevertheless, this limitation does not encumber the system’s ability to give a good idea of the broad distribution of metaphors in the domain of interest by producing results on a large scale.
Here, we briefly summarize the architecture of metaphor detection automation in MetaNet. We focus on two implementations of this system for metaphor discovery. Our main interests are to showcase some results produced with this system, and to delve into some insights about cross-domain similarities, specifically in the domains of cancer and poverty. Prior work offers detailed discussion of the MetaNet architecture, so there is no need to focus on it here. Hong (Reference Hong2016) discusses the MetaNet pipeline, including how error (false positives and negatives) is handled, how metaphoricity scoring is determined, how gold standard annotations are used to train the system, and how precision and recall (f-score) are measured. The ontological structures and hierarchically organized knowledge base forming the metaphor and frame repository are detailed in Stickles et al. (Reference Stickles, David, Dodge and Hong2016). The latter provides a discussion of similarities and differences with other frame-representational lexicographic resources such as FrameNet (Ruppenhofer, Ellsworth, Petruck, Baker, & Scheffczyk, Reference Ruppenhofer, Ellsworth, Petruck, Baker and Scheffczyk2016) and WordNet (Fellbaum, Reference Fellbaum1998), as well as the role of grammatical constructions in determining the relationship between lexical items and source-to-target domain mappings. This paper also outlines taxonomies of frame-to-frame and metaphor-to-metaphor relations, and discusses how decisions were made on what counts as a conceptual metaphor, and how conceptual metaphors are assigned to linguistic instantiations. Additional discussion on the lexical and frame resources seeding MetaNet’s frame repository is found in Dodge et al. (Reference Dodge, Hong and Stickles2015), which includes examples of how recall and precision are evaluated, with special attention to poverty. David, Lakoff, and Stickles (Reference David, Lakoff and Stickles2016) provide a detailed sketch of how primary metaphor networks are designed to find data in the domain of gun rights, showing how existing primary structures are augmented (with new frames, new specific metaphors, and new lexical items) to enhance coverage. Similarly, David (Reference David and Dancygier2017) illustrates MetaNet’s metaphor network expandability with data from the target domain of democracy.
2. Theoretical grounding
Primary metaphor (Grady, Reference Grady1997) figures prominently in the network of inter-related metaphors in the metaphor repository. They serve as the basis for more specific metaphors. Primary metaphors are schematic, embodied, and likely to be universal (Kövecses, Reference Kövecses2005; Lakoff, Reference Lakoff2012). They are formed early in cognitive development, and result from predictable simultaneous co-experiences of two domains during interactions with entities, forces, and people in the world (Grady, Reference Grady1997; Johnson, Reference Johnson1987). Primary metaphors, such as states are locations, purposes are destinations, more is up/less is down, and communication is object exchange are an unchanging part of this core repository, while additional metaphors may be added as new linguistic metaphors about specific target domains are discovered via empirical methods. For instance, a phrase like tackle poverty instantiates the primary metaphor difficulty in action is physical combat, but it also fulfills a more specific metaphor, dealing with poverty is physical combat. The latter is a more specific subcase of the former and a closer match for the identified linguistic metaphor. The verb tackle evokes the source domain of Physical Combat, as do many other verbs such as combat, fight, attack, and defeat. By associating multiple lexical items with particular source domains, and setting those domains as the source domains of metaphors, the system is trained to find tokens on the basis of one type (e.g., tackle poverty) such that it finds multiple other types (using different verbs) and tokens while still fulfilling the same source domain. The association of multiple lexical units to a single frame is a procedure inherited from other frame-based approaches to lexicography and semantic computation, such as FrameNet (Ruppenhofer et al., Reference Ruppenhofer, Ellsworth, Petruck, Baker and Scheffczyk2016), but takes one step further by associating frames with conceptual metaphors (frame-to-frame mappings).
Semantic frames are central to defining metaphor in MetaNet. A semantic frame is a knowledge schema whereby “a word’s meaning can be understood only with reference to a structured background of experience, beliefs, or practices, constituting a kind of conceptual prerequisite for understanding the meaning” (Fillmore & Atkins, Reference Fillmore, Atkins, Lehrer and Kittay1992, pp. 76–77). It is akin to scheme, script, or frame in psychology (Barsalou, Reference Barsalou1982), and emerged as a part of FrameNet and other computational and lexicographic resources (Fillmore, Johnson, & Petruck, Reference Fillmore, Johnson and Petruck2003; Ruppenhofer et al., Reference Ruppenhofer, Ellsworth, Petruck, Baker and Scheffczyk2016). The utility of frames for modeling metaphoric mappings has been considered (Sullivan, Reference Sullivan2006, Reference Sullivan2016), especially at the intersection of conceptual metaphors and lexico-grammatical structures. MetaNet models metaphor as frame-to-frame mappings, and therefore uses frames, frame elements, and lexeme-to-frame evoking patterns. It also uses grammatical constructions to determine how mapping in metaphoric words occurs, given their syntactic environment (David, Reference David2016; Stickles et al., Reference Stickles, David, Dodge and Hong2016).
3. Corpora and semantic resources
Of the two case studies presented here, one compares English and Spanish metaphors for poverty in the Gigaword corpora (Mendonça, Jaquette, Graff, & DiPersio, Reference Mendonça, Jaquette, Graff and DiPersio2011; Parker, Graff, Kong, Chen, & Maeda, Reference Parker, Graff, Kong, Chen and Maeda2011), and the other compares cancer metaphors in English in two corpora, a general corpus (the GLoWbE corpus: Davies, Reference Davies2013; Davies & Fuchs, Reference Davies and Fuchs2015) and a specialized corpus compiled by the MetaNet team, consisting only of cancer blog texts. The corpora are summarized in Table 1.
table 1. Summary of corpora used
In the first study, we test the system’s efficacy in a cross-linguistic comparison within one domain (poverty), and in the second, we perform a cross-corpus comparison in another domain (cancer). The cross-linguistic (poverty) comparison is between two corpora of the same size, and therefore the raw counts are comparable. The Gigaword corpora in English and Spanish were chosen because of their availability in both languages, their large sizes, and their representativeness in the domain of poverty, a topic that often appears in the type of newswire data in these corpora.
The cross-corpus (cancer) comparison is between a very large and a much smaller corpus, and therefore the results are presented as normalized per 100,000 tokens. Since the custom cancer corpus was collected from blog and forum entries, thus constituting a coherent genre, the GloWbE English corpus was selected for its ability to be filtered by the blog genre, providing a comparable subcorpus. The GloWbE corpus was mined only for a subset of data tagged as blogs because it is quite large, and because genres other than blogs are more general and not necessarily venues for writing about cancer.
The specialized cancer corpus was compiled to have a high concentration of cancer discourse, collected from online cancer discussion boards (the domains scraped are <https://community.breastcancer.org> and <www.cancerforums.net>). Genre-based comparisons in corpus linguistics help determine whether the observed trends are a feature of the language in general, or of a particular sampled speech community in particular. The discussion boards mined provide cancer patients, loved ones, and doctors with a venue for sharing information about treatment, venting frustrations with side effects and relapses, sharing good news, asking informational questions, and creating a sense of community where sensitive topics can be discussed without fear of judgment.
The resources discussed here are available online. The MetaNet source code is available in a GitHub repository (https://github.com/TheMetaNetProject), and a publicly searchable MetaNet Wiki, containing all the metaphors and frames (English only), is located at <https://metaphor.icsi.berkeley.edu/pub/en/>.
4. Cross-domain comparability in metaphors for social issues
MetaNet is a large-scale automated metaphor identification system that can be used to better understand non-literal language use in two different domains, poverty and cancer. Before discussing our analysis, we consider why it is useful to study these domains together, and why we might expect them to have anything in common.
Both cancer and poverty are social problems, and thus, the metaphors used to communicate about them often lead to negative inferences about their effects on individuals and on society. For instance, dealing with poverty and cancer is frequently seen as a physical struggle (combat poverty, fight against cancer), and as movement (get out of poverty, my cancer journey). Studying them together gives a better sense of how high-level primary metaphors (such as difficulties in action are physical struggles with an opponent and progress is forward motion) behave across multiple discourse domains and genres. Rather than arguing that for two separate metaphors, cancer is war and poverty is war, we can state a generalized metaphor such as social/personal difficulties are physical combat. In so doing, we can use the same concrete domain inferences to reason about multiple target domains, possibly opening the door to others in the future (e.g., terrorism, drugs, addiction).
That poverty should be metaphorically construed as a personal and social struggle is unsurprising – poverty is indeed a major problem. How is metaphor used in discourse about poverty, and how does it shape policy, ultimately? First, in descriptions of poverty in the media we often observe a preponderance of violence metaphors. President Lyndon B. Johnson’s famous ‘war on poverty’ slogan resulted in a national-wide effort to target many aspects of poverty. It led to the creation of long-lasting national policies (see Cahn & Cahn, Reference Cahn and Cahn1964). In this case, poverty is framed as the enemy, and the resisting protagonists are individuals or agencies who have to fight poverty.
Recent work on metaphors for poverty using the MetaNet system (Dodge, Reference Dodge2016) also shows how it is often framed as some type of harm (including disease) or movement, usually to and from a physically low location, e.g., fall into poverty, get out of poverty. As we illustrate here, these metaphors appear in both English and Spanish, with greater or lesser degrees of violence evoked by the lexical items used. Some metaphors, as in (1) and (2), suggest general physical antagonism, while others, like the Spanish and English examples in (3) and (4), are more strongly militaristic and evoke either weaponized combat or hand-to-hand combat. Footnote 3
(1) The agenda for an upcoming gathering of conservative Anglican clerics includes discussions about dialogue with Islam and fighting poverty. (engw_72353)
(2) Feminists also used the occasion to draw attention to the estimated 100,000 Filipinas forced by poverty to work as prostitutes, some of them exported to Japan and other countries. (engw_40)
(3) La miseria invade a los libaneses desplazados de Trípoli.
‘Poverty invades Lebanese displaced out of Tripoli.’ (esgw_501479958)
(4) A job is the best weapon against poverty. (engw_16010)
Nevertheless, though metaphors of fighting, war, and violence are common, inspection of the data reveal that other metaphors are used as well. In (5) and (6), for instance, poverty is construed as a low location into which one can sink, and from which one hopes to emerge.
(5) … ha enfrentado graves conflictos políticos, militares y sociales, en un país sumido cada vez más en la pobreza. (esgw_502037711)
‘… (he) has faced serious political conflict, both military and social, in a country immersed (deeper and deeper) in poverty.’
(6) He also estimated massive aid from outside was needed to help the African nations emerge from poverty and support political and economic reforms. (engw_145)
Another common source domain for construal of poverty is disease, one that can infect individuals and societies, and one that must be addressed with metaphorical treatments and medicine, usually in the form of strong socioeconomic reform. As shown in (7), speakers often mix metaphors. In mixing war metaphors with disease metaphors they construe the social condition simultaneously as one that needs to be combatted and one that is a pathology.
(7) We must combat such social pathologies as widespread poverty, the breakdown of family life, crime, alcohol and drug abuse. (engw_56)
Indeed, framing crime as a disease or virus can cause people to think about systemic, reform-oriented solutions, and framing it as a beast or monster can cause them to believe harsher punishments are needed (Thibodeau & Boroditsky, Reference Thibodeau and Boroditsky2013). Since the corpus work presented here reveals that poverty is also widely metaphorically framed as a disease, this could inspire new empirical work on whether similar patterns of reasoning will emerge with respect to this socioeconomic issue.
Like poverty, we have reason to believe that cancer, as a personal and a social problem, is metaphorically construed using two prominent metaphors, cancer is a war (or physical combat) and cancer is a journey (Demmen et al., Reference Demmen, Semino, Demjén, Koller, Hardie, Rayson and Payne2015, Semino et al., Reference Semino, Demjén, Demmen, Koller, Payne, Hardie and Rayson2015). Indeed, in a similar spirit to President Johnson, President Nixon declared a ‘war on cancer’ only seven years later (National Cancer Institute, 1971), and many other ‘wars’ on many other perceived social ills have been waged since then, from the war on drugs to the war on terror (Elwood, Reference Elwood1995).
As is clear from the following excerpt from our specialized cancer blog corpus, war language is not unique to leaders talking about policy goals as a nation. It also appears in individuals’ discussions of their subjective, personal experience with the disease, as in (8).
(8) I describe my year of cancer as sleeping with an enemy, unrelenting and omnipresent. I couldn’t shake its presence, even in those brief moments of relaxation. […] To the world, I presented as unsinkable. This is what my dear family and friends required. They needed their Amy back, the Amy who always marches on, through the best and the worst. Probably, I needed it, too. Everyone felt better telling me how strong I am and how I will beat it, but there, in the dark, I was terrified and helpless. […] Each time I went for chemo, I felt like a lamb going in for slaughter. I tried to visualize tiny resistance fighters living in my breast, my own Polish forest, beating away the Nazis inside of me. (Small-McKinney, Reference Small-McKinney2014)
In this excerpt, the writer introduces cancer as an enemy, a persistent undesirable, evil companion to fight. She refers to expectations for her to be a soldier in the battle, one who will be strong and continue marching on despite feelings of fear and helplessness. She refers to her immune system as small soldiers who fight the cancer inside her breast, which she depicts as a Polish forest, referring back to a WWII location that saw heavy losses. Invoking military history allows her to vivify her experience with cancer, including her feelings of helplessness. Often, as the lamb for the slaughter idiom above shows, the metaphor shifts to other, non-militaristic yet still violent images, sustaining the struggle dynamics introduced at the beginning.
As (8) illustrates, cancer treatment is often discussed in terms of battles in the popular media, among friends and family members, and in doctor–patient interactions (Magaña & Matlock, Reference Magaña and Matlock2018; Olweny, Reference Olweny1997; Semino et al., Reference Semino, Demjén, Demmen, Koller, Payne, Hardie and Rayson2015). Common in such discourse is language such as fighting the disease, knocking the cancer down, army of oncologists, and winning the battle. Indeed, war language is predominant in metaphors for all types of diseases (Casarett et al., Reference Casarett, Pickard, Fishman, Alexander, Arnold, Pollak and Tulsky2010; Weiss, Reference Weiss1997), as well as for dealing with pain in general (Semino, Reference Semino2010; Stewart, Reference Stewart2014).
There are other types of metaphors for both poverty and cancer as well as for other social issues. Common metaphoric source domains include measurable objects with size and area (cancer/poverty grows, cancer/poverty spreads), vision-related domains (live in the shadow of poverty/cancer), and living entities (the cancer pest, the jaws of poverty). Nevertheless, while some diversity exists, experimental and corpus research has shown that social, psychological, and other non-tangible phenomena perceived as detrimental in some way are often metaphorically characterized using one or two of a very limited set of source domains. For instance, experiments show that people can naturally be nudged towards policy change decisions when crime is construed as either a beast or a disease (Thibodeau & Boroditsky, Reference Thibodeau and Boroditsky2011), and the conceptualization of climate change has similar salient dual construal patterns, i.e., dealing with climate change is seen either as war or as a race (Flusberg et al., Reference Flusberg, Matlock and Thibodeau2017).
To deduce, using traditional corpus methods, what the metaphors for poverty, climate change, or other domains might be, we would have to perform individual studies on each domain and compare findings. This would be challenging from a standpoint of methodological comparability, as each study could be carried out according to different criteria, using different linguistic triggers, and based on varying decisions regarding determination of metaphoricity (see Gibbs, Reference Gibbs2015, and Semino, Heywood, & Short, Reference Semino, Heywood and Short2004, for commentary on problems inherent in metaphor analysis).
We present a multidomain, multilingual study using the automated metaphor identification method provided by MetaNet. In the next section we detail some of the inner workings of this system, and illustrate how it goes about finding metaphor in texts.
5. The MetaNet architecture and procedure
MetaNet consists of two parts: the metaphor repository and the automated metaphor identification processor. The repository is a network of hierarchically organized metaphors. It contains relationships between lexical items and the source- and target-domain frames linked to those metaphors. For instance, the metaphor dealing with cancer is physical combat is a subcase of the metaphor difficulties in action are physical combat. The lexical items and phrases cancer, cancerous, cancer treatment, etc. are associated with the Cancer frame, the target-domain frame of the first metaphor. The lexical items fight, attack, punch, etc. are associated with the Physical Combat source domain frame of both metaphors. All of these entities and their relationships are stored in the repository, which later acts as a knowledge base for the automated metaphor identification processor to perform its function.
The scripts in the automated metaphor identification processor use information from the repository to crawl over corpora and find metaphoric expressions within a limited set of domains (Hong, Reference Hong2016). The domains currently covered by MetaNet are include poverty, gun control, democracy, taxation, governance, and bureaucracy and other domains of social concern. Because these problems receive much coverage in the media, there is ample opportunity to scrape data from relevant news reports. Together, the two parts yield a third component, the annotation database, which stores the automatically identified metaphoric sentences, assigning them internal ID numbers (reported along with example sentences in the current work), and provides information about what source domain word (e.g., crushing) and what target domain word (e.g., poverty) were involved.
The architecture is set up so the source and target domain frames of metaphors are arranged in an inheritance lattice relative to each other, in which specific frames are subcases of more general ones. For instance, as shown in Figure 1, Poverty is a specific type of Social Problem, and Disease, a more specific type of Physical Affliction. Metaphors are also hierarchically organized. For instance, poverty is a disease is a subcase of social problems are physical afflictions.
Fig. 1. Metaphor inheritance diagram for poverty is a disease with relations to lexical items.
When a metaphoric collocation is encountered in text, e.g., cure poverty, at least one of two metaphors is automatically realized. The metaphor poverty is a disease, if present in the metaphor repository, would be recognized via link number 2 in Figure 1 because Poverty and Disease would be directly linked to that metaphor as target and source domain frames, respectively. However, if such a specific metaphor is unavailable, we at least get the more general social problems are physical afflictions as a candidate metaphor, via the links labeled 1 in the diagram. This way, because Poverty is a subcase of another higher-level frame (which itself is the target domain frame of a higher-level metaphor), we achieve accurate metaphor detection from the lexical inputs of cure and poverty at varying levels of specificity.
This mechanism enables queriers to use few metaphor networks to identify many metaphoric expressions across diverse texts. The system is multilingual, and thus able to dispatch a core set of conceptual metaphors over multiple languages (e.g., English and Spanish). Data presented below were identified as linguistic metaphors with metaphoricity scores of 0.7 or higher. Each automatically detected linguistic metaphor receives a metaphoricity score from 0 to 1, based on the best path through the metaphor and frame inheritance networks triggered by the lexical items. Footnote 4 Therefore, when literal uses of source domain-evoking terms are picked up, they are filtered out. This reduces the occurrence of false positives.
The methodology presented here differs from existing corpus-based approaches to metaphor (e.g., Stefanowitsch, Reference Stefanowitsch, Stefanowitsch and Gries2006) in one crucial way: lexical items do not directly map to metaphors, and are not themselves taken as metaphoric for purposes of discovery. Instead, lexical items are associated with frames, which in turn are associated with metaphors; the metaphoricity of a phrase is determined by a cascade exploiting frame inheritance networks, metaphor inheritance networks, and frame-to-metaphor relationships, while mediating via the grammatical constructions in which the candidate expression appears (Dodge et al., Reference Dodge, Hong and Stickles2015). The same lexical item, e.g., cancer, can be metaphoric in one sense (poverty is a cancer in our society) and literal in another (a tough battle with cancer), and the system can determine which is which. This method can be expanded to additional languages, especially given that lexical items are associated with existing frame and metaphor structures. Though no substitute for in-depth qualitative analysis that seeks a broad set of potentially culturally specific linguistic instantiations of these and other conceptual metaphors, this system can paint an initial picture of multilanguage distributions of metaphors known to be dominant in the domains of interest, especially when crawling over very large corpora in sizes (not possible with hand-annotation).
6. Automated metaphor identification results
6.1. poverty in cross-linguistic comparison
A survey of metaphor detection results for metaphor identification in the poverty domain from the Spanish and English Gigaword corpora reveals an abundance of metaphor around target domain expressions such as poverty, destitution, impoverished, impoverishment, indigence, underprivileged, the 47 percent, and the 53 percent, with similar expressions in Spanish. Table 2 shows the target domain-evoking lemmas, and the number of results of linguistic metaphors detected in the English and Spanish Gigaword corpora. These numbers illustrate that metaphoric language gravitates primarily around one or two key target domain terms.
table 2. Raw frequencies of metaphoric poverty-related expressions occurring with specific target lemmas. Metaphoric senses are most common for poverty and pobreza, with some occurrences across other lemmas
notes: AFP: Agence France-Presse, English Service; APW: Associated Press Worldstream, English Service; XIN: Xinhua News Agency, English Service; CNA: Central News Agency of Taiwan, English Service; LTW: Los Angeles Times/Washington Post Newswire Service; NYT: New York Times Newswire Service; WPB: Washington Post/Bloomberg Newswire Service.
Each target-evoking word and phrase is associated with frames that are associated with target domain frame slots of one or more metaphors (e.g., poverty (n.) → Poverty frame → dealing with poverty is physical combat). For this sample search, the target-evoking words and frames are fixed to isolate only those metaphors pertaining to poverty, as many other metaphors are in the system (e.g., communication metaphors, emotion metaphors) and they may be using the combat source domain as well. The metaphors are objects the system uses to create a link between the target and source domain frames, with the latter bringing a set of associated lexical items. For the poverty domain, the stored metaphors are poverty is a crime, poverty is a disease, poverty is a fierce creature, poverty is a motion impediment, poverty is a plant, poverty is physical harm, addressing poverty is treating an illness, addressing poverty is waging war, amount of poverty is size, and amount of poverty is size of geographic feature.
These metaphors are nested in more complex hierarchical networks of metaphor, such that, for instance, if the phrase mushrooming poverty rate is detected, it simultaneously counts as amount of poverty is size (specific) and abstract quantity is an object with measurable size (general). The latter has further entailments, such as reduction in abstract quantity is decreasing in size (e.g., shrink poverty). This entailment is not present for the poverty-specific subcase metaphor, but the latter can still benefit from it in the metaphor identification process, and a phrase such as shrinking poverty now can be detected as well. In short, while only a handful of metaphors are dedicated to poverty in the system, many more metaphors about poverty not limited to that list can be detected by virtue of the hierarchical way in which the metaphors are structured. The rich set of lexical items associated with source domain frames yields this amplified detection effect.
Once the set of metaphoric expressions sought are narrowed down to ones in which the items in Table 2 appear, it is time to detect any source domain language. Lexical items are grouped into frames, such that mushroom, expand, and balloon all evoke the frame Increase In Size. Footnote 5 In Table 3, the source domains of metaphors for poverty, categorized by subgroups and groups of metaphoric source domain frames, are reported with their normalized frequencies (NF) per 1,000 extraction results, for comparability between the two languages and example lexical unit (LU) types encountered in each language. The lexical units are the source-domain frame words and phrases that evoke the conceptual metaphors.
table 3. Results for Poverty metaphors in English and Spanish from the Gigaword corpora (quartile ranges: 33–130, 10–32, 2.5–9, 0–2.5). NF: normalized frequency; LU: Lexical Unit; Abs. Dif.: absolute difference
Table 3 captures important information about metaphor in these languages. At a high level, metaphors can be grouped into three broad categories that have something in common semantically among the source domain frames – Violence/Harm, Location/Motion, and Properties of Objects and Entities. In Table 3, these groups are sorted from most to least amount of variation between English and Spanish in terms of the absolute difference (Abs. Dif.) between normalized frequencies in the two languages, per 1,000 results. These three broad categories were not instrumental in deriving the results, but are merely groupings provided for expository purposes so as to render the data easier to view. Footnote 6
Second, the data are presented in cells according to quantiles of normalized frequencies (represented as varying cell shadings), and calculated based on log values, which are not reported. This ensures that ranges of normalized frequencies categorizable are presented in a comparable way, as the scales vary greatly from highest-frequency (in the hundreds, darker gray) to the lowest-frequency items (often below 4 occurrences per 1,000 results, unshaded). The main clusters are around the Location/Motion and Violence/Harm in both languages. For Violence/Harm, English is dominated by metaphors that construe poverty as physical struggle or as a disease (e.g., poverty infects the city). Spanish also uses a high concentration of poverty as a disease, but it also employs war-specific language about poverty, more often than English (e.g., guerra contra la pobreza ‘war against poverty’). This is interesting given that the war against X metaphoric construction was apparently introduced and popularized in American news media and political rhetoric (Elwood, Reference Elwood1995; see also Flusberg, Matlock, & Thibodeau, Reference Flusberg, Matlock and Thibodeau2018).
English and Spanish diverge most notably on metaphors that express poverty as physical harm (or dealing with poverty as violent confrontation) or as a state change of an object (growing or shrinking in size). These differences are driven by a few subgroupings. For instance, in Spanish the metaphor poverty is a changing (growing) entity is deeply entrenched (more so than in English), and thus, expressions referring to poverty in terms of physical expansion or extension are common (e.g., aumentar la pobreza extrema ‘augment extreme poverty’, la pobreza se ha extendido ‘poverty has extended’). In English, poverty is a disease is prominent (e.g., poverty infected the city) and thus is linguistically realized in a more robust distribution compared to Spanish. In both languages, Location/Motion metaphors are slightly more common than Violence/Harm metaphors, with more high-density subcategories (categories of over 40 instances per 1,000 results). It is useful to examine a few subframes within this category, especially those pertaining to vertical movement. While the two languages show similar concentrations of static location metaphors (live/dwell in poverty, vivir en la pobreza), as well as metaphors about confinement or containment in a bounded region (mired in poverty, canasta de pobreza ‘container of poverty’), English seems to favor vertical motion metaphors, as in catapult/leap out of poverty and slide/fall into poverty. As Dodge (Reference Dodge2016) observes, there is more emphasis on manner of motion in English, perhaps resulting from the manner-encoding nature of English verbs. Since good is up and bad is down are common primary metaphors for negative social and psychological states, and since poverty is seen as a bad social or personal circumstance, we expect many metaphors that emphasize being (unwillingly) knocked down into a state of poverty, and the desire to rise from that low state.
The high counts of these metaphors for both languages exhibit slight differences that are informative about how emphatically news in each of these languages conveys these metaphors. This invites some questions about the cognitive status of poverty metaphors in the minds of speakers of each language. For instance, given that bad is down is less frequent in Spanish, with expressions like plunge into poverty, does this mean that Spanish speakers feel the negative connotations of poverty to a lesser extent? These computationally driven corpus findings may shed light on possible experimental directions to investigate how the language being used may influence conceptualization and decision-making.
6.2. cancer metaphors in English across two corpora
Using the same system, Footnote 7 and inheriting the same primary metaphor networks as those used in poverty metaphor identification, we see similar patterns emerging for cancer metaphors. The corpora are much smaller and yield comparably fewer results. Also note that the Gigaword corpora used for the Spanish–English poverty comparison were from newswire sources, which are likely to include proportionally more discussion of socioeconomic issues, including poverty. Cancer, on the other hand, is a niche discussion topic. It appears in the news, but more often, texts with a high concentration of metaphoric language about cancer tend to come from blogs and online forums. Using the same source domain semantic categories as for poverty, we supply a within-genre cross-corpus comparison between a general (and large) corpus (filtered for blog data only) and a specialized corpus with only blog and forum data. Table 4 summarizes results surrounding the topic of cancer in English.
table 4. Results for Cancer metaphors in English from a general and a specialized corpus (per 10 results) (quartile ranges: 0.5–4, 0.2–0.5, 0.1–0.2, 0–0.1). NF: normalized frequency; LU: Lexical Unit; Dif.: actual difference
The results in Table 4 are presented in order of actual, rather than relative, difference in normalized frequencies between the GLoWbE corpus and the specialized corpus. The presence of mostly negative values for the specialized corpus in the ‘Dif.’ column indicates that the specialized corpus (predictably) contains a higher quantity of most metaphors for cancer. From this, we infer that more metaphoric language about cancer is employed by writers who are themselves stakeholders in the cancer world, compared to what we would expect in blog texts at large.
The quartile color concentrations in Table 4 illustrates how genres dedicated to cancer talk, such as blogs and fora, show greater frequency of the two metaphor families, Violence/Harm and Location/Motion, which dominate discussions of cancer more generally. Motion metaphors concentrated around discussions of moving forward (cancer journey, the path forward), dealing with obstacles on a path (get through this, overcome cancer), and being collocated with cancer (live with cancer). Among Violence and Harm metaphors, cancer is either an enemy or antagonistic entity that attacks, beats, or hits the patient, or it is a more abstract force-dynamic antagonist that grips, holds, or squeezes the patient. In the first case, the patient is an equal opponent who can fight back (and potentially win the war), while in the second case, the patient is demoted to a helpless entity that is manipulated and held down. Aside from the two dominant metaphor groups, there are also some spikes in the source domain of object states, especially changes in object size. This is due to the frequent description of cancers as growing, spreading (or conversely, shrinking or diminishing) entities inside the body or in society at large.
This latter finding points to an important observation we make about conceptual categorization in cancer-related metaphor. Namely, there is frequent compartmentalization into three different levels in the target domain: the societal level, in which cancer treatment policies and funding are of primary importance; the individual level, in which cancer is reified as a person that the patient ‘fights’ or ‘travels with’; and the physiological level, at which cancer cells are seen as moving, invading, or attacking within the body (see Semino et al., 2004, for discussion of the last type). Sentences (9), (10), and (11) illustrate these three perspectives on the metaphoric description of cancer.
(9) The ICR’s mission is to make the discoveries that defeat cancer. (societal) (can_spec_1052)
(10) My own brand of faith simply provided me the will to endure in the face of cancer’s randomness. (personal) (can_spec_4)
(11) Kris Carr has halted tumors growing in her liver for over 7 years by focusing on nutrition. (physiological) (glo_3914)
This difference is important because it highlights the presence of slightly different metaphor systems from what we observed for the domain of poverty, or other social problems. Both poverty and cancer are personal problems that are metaphorically construed as opponents (enemies) or states to travel through and out of, the two most frequently used metaphors in both domains. But cancer also has a physical, physiological component that is relevant to the level of the individual (Semino et al., 2004, p. 1279). We add that, although cancer is a physical property of the body and ‘grows’ within the body, it usually cannot be readily perceived via the senses, and therefore the physical bodily effects of cancer are more-or-less metaphorically construed.
7. Discussion
The metaphor identification system made a large variety of data instantly available in two languages. By analyzing specific sentences in the output, we can see that although the movement-related words (slide, fall, rise) and violence words (battle, attack, fight) may both be used in metaphoric expressions about both poverty and cancer, the metaphors motivating their use differ in the two domains. Cancer is talked about as a journey, with paths and impediments (rarely the representation of poverty); the motivation in the cancer journey is to arrive at a healthy state (location), while movement in poverty is motivated by trying to flee a bad (often low) location. For the cancer patient, this is an extension of the already entrenched metaphor life is a journey, where the experience of cancer is a new phase in the already-ongoing journey of life. This integration of the cancer experience with the life journey is already attested in English narratives by sufferers of breast cancer (Gibbs & Franks, Reference Gibbs and Franks2002). The journey frame is evoked with not only the lexical item journey, but also constructions involving paths (road to cancer recovery), obstacles (a bump in the road), and enablements (cancer treatment is going smoothly) (see also Magaña & Matlock, Reference Magaña and Matlock2018). Construals of poverty (and socioeconomic states in general) as specifically a journey are fairly uncommon. This is perhaps because it is less common to associate one’s personal socioeconomic status with life is a journey than with one’s own health state, as is the case for cancer is a journey. Nevertheless, some instances of poverty as a journey do appear, as in (12).
(12) Her journey out of poverty has been marked by single-mindedness and luck. (engw_114719)
Although (12) is a journey metaphor, it conveys the vertical motion typical of poverty metaphors, via the construction NOUN out of X. This reflects how discussions of poverty focus less on forward motion and more on vertical motion, where poverty is often a low location that one inadvertently gets into (or is forced into) and wants to escape. This is true in both languages, but more so in English. The low location is also sometimes confining, drawing together not only action is motion, but also bad is down and states are locations. This fusion of multiple primary metaphors is exemplified in (13) and (14).
(13) The Congo we’re talking about has mountains of debt and canyons of poverty, so it’s a candidate for the debt cancellation that was trumpeted last year. (engw_113705)
(14) But it does not necessarily lead us to a more upwardly mobile middle class or rescue those drowning in poverty. (engw_109128)
Mountains and canyons are topographic features of the terrain, with extreme height and depth. The mountains of debt, although high, do not employ good is up, rather difficulties are impediments to motion (mountains are hard to traverse on a journey) as well as more is up. Canyons evokes bad is down and unchangeable states are confining locations.
Although cancer is also a negative state, and presumably, should use the bad is down entailments of movement metaphors, we never see expressions like *fall into cancer or *live in cancer. The cancer experience is most commonly mapped to the path of movement (my cancer path), a co-mover or companion (live with cancer), or an obstacle in one’s normal life path (overcome cancer), and less as a location where one ends up in or tries to move away from. Perhaps this is due to seeing cancer as a more transient state, as health states tend to be, and one that happens to be particularly unwelcome as one proceeds with the normal journey of life.
An enormous caveat here is that many trigger words evoke violence and movement scenarios simultaneously. As the numbers are reported in the current work, it is not immediately clear that some datapoints can be cross-categorized into two or more frame families (indeed, in the calculations we only report each datapoint along with the first metaphor suggested by the metaphor identifier). But lexical items often evoke multiple concrete frames in a complex way, as is the case with the use of shackles in (15), and invasive in (16), and Spanish desierto ‘desert’ in (17).
(15) Unfair trade rules do not only prevent poor people from throwing off the shackles of poverty, but shackle poor people and poor communities still further. (engw_81227)
(16) The authors of the report provide a picture of the number of cancer survivors who had previously been diagnosed with an invasive cancer. (glo _15306)
(17) Y hay muchas formas de desierto: el desierto de la pobreza, el desierto del hambre y de la sed. (esgw_501545722)
‘There are many types of deserts: the desert of poverty, the desert of hunger and of thirst.’
In the system, shackle is classified as belonging to a family of frames that express confinement. However, a deeper analysis of the semantics of shackles in (15) would reveal how it should also be categorized as a form of Burden (shackles weigh you down), a type of Harm (shackles injure the body), and a type of Impediment to motion (shackles restrict movement). Similarly, invasive in (16) is categorized as Movement, but also used to describe movement of unwanted entities that can create harm, and invade is often associated with military encroachment. Finally, a desert (17) is a vast open area through which one can move, and a terrain that can be dangerous or fatal (the first being an action is motion metaphor, the second being a states are locations metaphor). Although the salient semantics of this word is spatial, the metaphorically relevant semantics concern the dangers (and helplessness) of this type of terrain, which might be more relevant in poverty metaphors. The problem of primary and secondary classification is one of polysemy in the metaphoric usage of individual lexical items, discussed in the literature on the challenges of grouping metaphors (Cameron, Reference Cameron, Cameron and Maslen2010; Deignan, Reference Deignan, Cameron and Maslen2010). This should be considered as this (or other computational metaphor identification) systems are further refined.
In this study, we illustrated the effectiveness of an automated method for finding a large number and broad variety of metaphoric expressions across two domains and two languages. These discoveries were made possible from the inclusion of high-level primary metaphors in the metaphor database that propel the automated identification system. Cancer and poverty metaphors were found by means of primary metaphors in the higher levels of the metaphor networks, of which more specific metaphors such as poverty is a disease and cancer is physical combat, among others, are subcases.
MetaNet allows researchers to identify metaphor in very large texts, and across languages. It may not provide as broad a coverage as some automated metaphor identification systems, but it can be used to query particular target domains, or domains of knowledge in which we would like to observe distributions of metaphoric language, and to do so cross-linguistically. Though not exhaustive, the results provide insights that could lead to further quantitative and qualitative analyses. Its large data output yields a good starting point for understanding conceptual similarities and differences in the domain(s) of interest. Other domains can be added to the core primary metaphor network with minimal tweaks, mostly in the form of adding subcase metaphors and additional lexical items and frames. This iterative process relies on a computationally operationalized version of primary metaphor networks and semantic frames as the source and target domains of conceptual metaphors. We have thus presented one example of MetaNet’s pipeline and a selection of results from the domains of cancer and poverty as proof of concept, hoping that such a system will make a valuable addition to metaphor analysts’ computational linguistic toolkit.
