Lonely wallets: differences in comprehending anthropomorphic language in Japanese and English

Personification is a type of metaphor where a non-human entity is treated as a person (Lakoff and Johnson, 1980). For example, if someone says “my car is temperamental,” we might infer that their car is unpredictable. This expression works because we can view a car's engine as being akin to a human mind. But in Japanese, there are personification metaphors such as “my wallet is lonely” (meaning the speaker's wallet is empty). This metaphor is unlike the previous example, because a wallet does not have an internal system that is akin to a mind. Thus, personification seems to have distinct features in English and Japanese, and in this work, we examine how these differences might influence sentence comprehension.

Metaphors involve the linking of a target or topic object with a completely different type of object or entity, which is called the source or vehicle (Gibbs et al., 1997; Holyoak and Stamenković, 2018). For example, the target of the phrase “my car is temperamental” is a specific car that is owned by the speaker. In this case, the source is the concept of PERSON (A CAR IS A PERSON; Figure 1). By using the word temperamental, the speaker is highlighting the similarity between the internal components of a car (e.g., the engine, electrical system) and the internal components of a person (e.g., the mind). Personification is argued to depend on an abstract schema, which allows many different features of humans to be predicated on various targets. For example, the sentence “my robot vacuum is lazy” uses the same abstract PERSON schema as “My car is temperamental,” but instead it activates the feature LAZY, which in turn activates the feature MALFUNCTIONING when applied to a vacuum.

In the process depicted in Figure 1, several steps are required to understand the sentence. The noun phrase my car activates the CAR concept, and the word temperamental activates the concept TEMPERAMENTAL, which is a feature of the PERSON concept. If this is a novel metaphor for the listener, then they must think about how to link PERSON and CAR, and they may realize that the internal state in the car is similar to the mind of a person in that it is a complex, hidden controller of the whole entity. Then that allows the interference that UNPREDICTABLE is the equivalent concept for a car that would correspond to the concept of TEMPERAMENTAL within a human mind. This final meaning that is derived from this sequence of steps is equivalent to the literal expression “My car is unpredictable.”

Since there are extra steps for processing a novel metaphor compared to the literal expression, one might predict that they should be more difficult to process than literal sentences (Lai et al., 2009; Lai and Curran, 2013). However, studies show that metaphors can be processed as easily as literal expressions (Gibbs, 1994, 2013), and one explanation is that an abstract relational schema for metaphorical language may be stored as categories (Career of Metaphor, Bowdle and Gentner, 2005), which allow people to quickly access their meaning (e.g., a category that links temperamental to unpredictable for non-human entities). Support for this account comes from studies like those by McGlone et al. (1994), who found that novel variants of idioms (e.g., he didn't spill a single bean) were processed as easily as their literal counterparts. Thus, abstract schemas allow fast processing of novel forms as long as they are similar enough to activate the schema.

One reason that personification metaphors are easy to process is that they reflect universal aspects of human experience (Lakoff and Johnson, 1980). For example, all humans have direct experience of the mental states in their own mind and it is natural to apply this knowledge to understanding others. Epley et al.'s (2007) theory of anthropomorphism called this bias to use our own mental states to understand others elicited agent knowledge. They suggested that another bias that can create personification is effectance, where the behavior of a complex system is summarized by using a single human entity. For example, the phrase “the stock market flirted with the 2000 mark” summarizes the complex behavior of a large number of stocks by treating it like a single person who is approaching a target. Studies have found that children distinguish between animate and inanimate objects early in development (Opfer and Gelman, 2010; Rakison and Poulin-Dubois, 2001), but also have a broad concept for sentient agents and are willing to anthropomorphize a range of inanimate objects (Gjersoe et al., 2015; Goldman and Poulin-Dubois, 2024). If these abilities arise from universal features of the human mind and communication, this would predict that people should easily interpret anthropomorphism in various languages.

In this work, we are interested in whether the processing of personification metaphors is driven by universal factors such as elicited agent knowledge and effectance, or whether it can be changed by culture-specific knowledge. In our view, mind-based personification forms like “my car is temperamental” are common in various cultures. But we have found examples of anthropomorphic language in Japanese that seem to be different from those that are used in English (e.g., “my wallet is lonely”). To understand why we think these forms are different, we will present a dataset of Japanese examples of anthropomorphic language and highlight how these examples differ from the typical forms in English. Then we will report two story processing experiments that test whether English or Japanese readers find these types of anthropomorphic language difficult to process.

Characterizing Japanese personification language

To characterize personification of inanimate objects in Japanese, we collected examples of anthropomorphic language from various sources such as television programs, internet videos, websites, and everyday life. This dataset was an opportunistically-sampled set of utterances that were heard or read by the authors and is similar to early speech error corpora created for studying linguistic processes in production (Fromkin, 1971; Garrett, 1975). Over a few years, we have found 46 spoken examples and 13 written examples. In the text, we will use English translations, but the original Japanese versions are available in the online repository (https://osf.io/vxmkt). It is not possible to argue definitively that these examples are different from those produced in English (since a large corpus of English anthropomorphic language does not exist), but we will try to suggest how each example differs from the type of anthropomorphic language that is typically used in English.

One context where personification often appears is in language between children and adults. For example, a kindergarten teacher tried to get a boy to finish his carrots at lunch by saying “Poor Carrot-san! You ate Hamburger-san, Rice-san, and Orange-san, but you haven't eaten any Carrot-san. Don't you think he feels lonely?” (Hayashi et al., 2009). By using this language, the teacher is not directly commanding the child to finish their food, but instead encouraging the boy to feel empathy for the lonely carrot and to relieve that loneliness (e.g., by eating them). This example also illustrates the use of titles in Japanese. Normally, one uses titles like -san (Mr., Ms., or Mrs.) to show respect to a human listener (e.g., at a company, an employee will always use -san to refer to their colleagues when talking to them, even if they have known them for many years). In child-directed speech, it is common in Japanese to apply -san to foods (e.g., ninjin-san, “Mr Carrot,” Chang et al., 2021). Since anthropomorphization emphasizes that the object is alive, it is strange for English speakers to anthropomorphize a carrot in order to convince a child to eat it.

While English anthropomorphic language with children often involves inanimate object characters with faces (e.g., Thomas the Tank Engine), Japanese adults use it to discuss the relationships between real-world objects and other entities. For example, an adult woman sold her air conditioner unit, and her mother wrote “The air conditioner was happy being taken care of by you. I hope it will do well with its next master” (T. Tatsumi, personal communication, 2023). By talking about the air conditioner's feelings, the mother was trying to make her adult daughter feel better about her effort in arranging its sale. This anthropomorphic language highlights the transfer of ownership of a real-world machine, which seems like a less common pattern in English.

Another function of anthropomorphization is to amuse the audience. For example, the satirical newspaper The Onion once had a headline “New Liver Complains of Difficulty Working with Lou Reed,” which humorously treats his cancerous liver as being akin to his human collaborators (New Liver Complains of Difficulty Working with Lou Reed, 2013). An example of this in Japanese is a video where a female adult YouTuber was eating a sandwich where the tomato and lettuce were tightly stuck together and she commented “Tomato-chan and Lettuce-kun, why are you guys such close friends?” (Shion, 2023). By treating these ingredients as people who are stuck together, she turned her lunch break into a funny episode in the video. In addition like the -san title mentioned earlier, she used Japanese diminutive titles that are typically used with girls (-chan) and boys (-kun). These titles are rarely used for sandwich ingredients, even in child-directed speech (Chang et al., 2021), so by using them, she is highlighting the anthropomorphic nature of her speech. Since a liver is a complex living entity with unknown internal states, it is a more likely target of English anthropomorphic language compared to simple food like lettuce.

In the above examples, the situations are casual and informal. But Japanese anthropomorphic language also occurs in more formal settings. One example occurred during a medical exam where an older male veterinarian was talking to an adult couple about their sick cat, who eventually died (S. Yagata, personal communication, Sept. 21, 2023). While looking at the X-rays and talking about the cat's constipation, the veterinarian said “if Feces-chan gives us the benefit of coming out, then things will be more comfortable.” It is not normal to apply the diminutive -chan to feces and to talk about their intentional leaving of the fecal track (“give us the benefit”), but in this case, this anthropomorphic language may soften the authority of the doctor and help to show some empathy with the distress of the owner of the cat. While English-speaking doctors may use anthropomorphic language with children, it would be more unusual to use it with adults in a serious situation such as the imminent death of a beloved pet.

English has some anthropomorphic idioms (e.g., “opportunity knocks,” “the wind howled”). But some idioms like “the walls have ears” are actually not about an inanimate target (e.g., the walls), but instead imply that there are people nearby who might hear what is being said. Japanese also has anthropomorphic idioms like “my mouth is lonely,” which can refer to the mouth's desire for more taste sensations (e.g., foods), and this example was used in an infomercial for electric cigarette flavors (Horiemon, 2023). Another example from a commercial context is a person writing about a new belt for a kimono where they said, “Previously I had this belt dyed in a different color, but it quickly became a wife, so I made it again in a new color” (Yuzuru, 2023). Traditionally in Japan, a woman moved out of her family home and into her husband's house, so “becoming a wife” is a kind of idiom that can mean moving to the customer's house. While English idioms are typically used in informal situations, Japanese anthropomorphic idioms are conventional enough to be used in formal business situations with customers.

As we have seen, anthropomorphic language in Japanese often employs titles like -san, -chan, and -kun that are normally used with humans (e.g., o-tera-san Mr Temple). Another common title is -sama, which is used to show a higher level of respect than -san (similar to the meaning of “Sir/Madam”). Anthropomorphic examples of this are o-tsuki-sama (Sir Moon), o-tento-sama (Sir Sun), and o-hoshi-sama (Sir Star). There are various other examples using job titles such as fuyu-shogun (General Winter) and guuguru-sensei (Teacher Google). Finally, there are some examples that attach names (-tarou is a common boy's name, e.g., Momo-tarou) to inanimate objects like natto-tarou (natto is a kind of fermented bean food; Japan Natto cooperative society, n.d.). In English, titles are used less frequently than in Japanese, so this method of anthropomorphization of inanimate objects is less available.

Some types of personification in our dataset appear to be idiomatic, while others are not. To understand the degree that this language is idiomatic in Japanese, we searched for these forms in the National Diet Library Ngram corpus which includes 2.3 million books and magazines with approximately 1.7 billion words and phrases (National Diet Library, 2023). We created regular expression search terms for each argument-predicate form in our dataset. Since our goal is to ensure that we can find examples with inanimate objects in the corpus, the search terms were designed to match different word forms that have the same basic predicate. For each inanimate object in the search term, we also searched for the same predicate paired with a human argument (e.g., hito, 人, “person”). For example, the phrase “the beans do not dance” (豆が踊らない) had the search term “(豆|人)[がは]踊ら,” which could match the predicate “dance” paired with either “beans” or “person,” and it also allowed various particles and verb morphology. For each term, we collected the count of the number of exact phrase matches for humans and for the original inanimate object (a table with the results is present in the FrequencyDataset.pdf file in the OSF archive). Since these predicates are normally used with people, the number of matches for the human argument should be higher than the number of matches for the inanimate object argument. In our dataset, 17 examples appeared to be idioms because the predicate was present with the inanimate argument in the NDL Ngram corpus. Of these, 11 were used with titles and the rest were verb predicates (“discuss with my wallet,” “my money ran away,” “my mouth is lonely,” “bread is born,” “weeds show their face,” and “the beans don't dance”). The rest of the 37 forms appear in the corpus with a person as an argument, but not with the inanimate object that occurred in our dataset. Thus, the majority of the anthropomorphic argument-predicate combinations in the dataset are novel forms.

A unique feature of Japanese anthropomorphic language is its application to simple inanimate objects like food which do not have features that are associated with minds like the ability to think. To characterize this explicitly, we used the animacy ratings of VanArsdall and Blunt (2022), who had English speakers rate various features related to animacy for 1,200 common nouns. One feature that raters had to judge was the ability of the noun's referent to think. Another feature was the degree of goal-directedness, where a referent might be able to have a goal, even if it cannot think (e.g., the goal of a train is to move along its track). To connect our Japanese dataset with these ratings, the inanimate object in each example was mapped to a similar English word in the VanArsdall and Blunt lexicon (the word concept here refers to the term that is in the lexicon). There were seven words that did not match (e.g., Google, price, taste, feces, mouth, credit card, and chicken leg). In addition to the original concepts in the dataset, five extra concepts (woman, boy, dog, ant, computer) were added to allow us to understand how animate and inanimate concepts are related.

Figure 2 shows that humans are in the top right corner, because they can both think and have goals. Pets are in the middle in their ability to have goals, and they have a higher ability to think compared to insects and computers. The words from the dataset cluster down at the bottom left (there is a magnified view of this cluster), which are low in both thinking ability and goal-directedness. These ratings show that English speakers view the 49 concepts used in Japanese anthropomorphic language in our dataset as having very little capacity for thought or goals, even compared to an ant or a computer.

The above analysis demonstrates that most of the examples in this dataset are applied to targets (e.g., carrots, bread, belts) that are rated low on their capacity for thought or goals. This is puzzling from the point of view of the personification metaphor mechanism outlined earlier, because that mechanism assumes that there is a mind-like element in the object (Figure 1). When applying personification to a car, the engine is a complex, hidden object that is akin to a human mind. But in the sentence “Tomato-chan and Lettuce-kun, why are you guys such close friends?”, these sandwich ingredients are simple foods without a mind-like internal element, so it is difficult to use mind-based personification (left side of Figure 3).

An alternative explanation for this personification language is that this phrase activates the concept FRIENDS, which is a feature of the concept of HUMAN GROUP. In turn, this concept can be aligned with the concept of sandwich INGREDIENTS, where the corresponding concept for FRIENDS is STUCK TOGETHER (right side of Figure 3). Using a human group schema for personification does not require a mind-like element, and hence it can explain many examples from our Japanese dataset. For example, the carrot is lonely, because it is separated from the other foods that are members of its group. The air conditioner is happy, because it is joining a new group with a new master. In addition, this explanation can also justify the use of titles. Japanese titles reflect hierarchical relationships between members in a group (-sama/-san is typically for higher senior members, -chan/-kun is for equal or junior members, Figure 3). Japanese speakers often use titles to show their relationship with someone (e.g., Tanaka-san), even when talking about that person to a third person. Thus, a group-based schema can explain a range of examples from our dataset as well as the use of relational titles.

In this section, we described a dataset of Japanese examples of anthropomorphic language applied to objects that English speakers code as not being able to think or have goals. Many of the argument-predicate combinations in this dataset were not found in a 1.7 billion-word corpus, which suggests that these are not conventionalized idioms. Unlike personification involving complex objects like cars and computers, many of these examples use simple non-moving objects like bread or carrots, which should make it more difficult to apply mind-based schema (X-IS-A-PERSON). We argued instead that Japanese readers may have some culturally-unique schemas (e.g., X-IS-A-GROUP-OF-PEOPLE) that allow them to process anthropomorphic language more easily than English readers. On the other hand, it is also possible that English readers have a powerful personification schema that allows them to easily understand any type of anthropomorphic language. In the next section, we will compare and contrast these two possibilities in two story comprehension experiments.

The comprehension of anthropomorphic language

The analysis of the dataset suggested that Japanese speakers are willing to anthropomorphize simple inanimate objects. Since these utterances involve a mismatch between the argument and the predicate, they should trigger some comprehension difficulty within theories that use linguistic constraints to check the compatibility of arguments and predicates (Bornkessel and Schlesewsky, 2006). These mismatches should also create difficulty for theories of comprehension, which are shaped by statistical regularities in the input (the words carrot and lonely rarely occur together, MacDonald and Hsiao, 2018). In this section, we examine how argument-predicate mismatches in personification language might be related to the literature on sentence comprehension.

Within the language processing literature, a range of evidence from event-related potential studies suggests that comprehenders have difficulty with mismatching argument-predicate combinations (e.g., inanimate subjects as the subject of active transitive verbs: the hearty meal was devouring the kids, Kim and Osterhout, 2005). This type of sentence triggers a P600, suggesting that comprehenders had trouble parsing online the idea that the meal was doing the act of devouring (they also rated these sentences as unacceptable offline). Yano and Sakamoto (2016) found a related mismatch effect in Japanese for sentences with active transitive verbs with inanimate subjects. Related evidence is found in studies with discourse story contexts. (Nieuwland and Van Berkum 2005) had Dutch participants read a story that introduced a context (e.g., a woman is helping a male tourist, who is checking in a suitcase), and then it was followed by a coherent discourse (e.g., the tourist asks the woman for a date) or an incoherent anthropomorphic discourse (e.g., the suitcase asks the woman for a date). They found a late positivity ERP effect, which they interpreted as evidence that readers found it difficult to understand stories with anthropomorphized inanimate objects in the incoherent discourse.

In a later study (Nieuwland and Van Berkum, 2006), they presented Dutch readers with stories about a human entity engaged in some activity that only humans do (e.g., a sailor talks with a therapist about his problems). This animate condition was compared with the same story with the human entity replaced with an inanimate object (e.g., a yacht talks with a therapist about his problems). When participants read this sentence about an inanimate object, they exhibited an N400 ERP effect, which they interpreted as evidence that this anthropomorphic sentence was more difficult to understand than the non-anthropomorphic sentence in the animate condition. But as participants read more sentences about this object, the N400 effect disappeared, which suggested that they created a discourse representation where this anthropomorphic entity was expected. This research suggests that readers start off comprehending mainly using world knowledge and then gradually build a temporary fictive discourse context when they experience stories that violate those constraints.

To explain how anthropomorphic language is processed, we propose three simplified accounts that can generate falsifiable predictions for our study. Comprehension theories often explain parsing as the building of representations that encode dependencies between linguistic elements such as arguments and predicates (Bornkessel and Schlesewsky, 2006; Gibson, 1998). When there is a mismatch between these elements, then Dependency theories predict comprehension difficulty, and that has been confirmed in various studies (Kim and Osterhout, 2005; Nieuwland and Van Berkum, 2005, 2006; Yano and Sakamoto, 2016). Since these Dependency accounts are argued to be universal theories of parsing, they predict comprehension difficulty when reading the argument-predicate mismatches found in Japanese anthropomorphic language. Alternatively, Anthropomorphism theories might predict that this type of language should be universally easy to process, because every human has a rich understanding of personhood (elicited agent knowledge, Epley et al., 2007; Lakoff and Johnson, 1980), and this might lead to a powerful schema that can be used easily with all inanimate objects. The final possibility is the Culture account, which argues that there are differences in the processing of anthropomorphic language depending on cultural knowledge. For example, if Japanese readers can readily access a culturally-specific schema (e.g., a group-based personification schema), then they might find the comprehension of personification of simple inanimate objects to be easier. To test the predictions of these accounts, we did a story comprehension study with anthropomorphic language in English and Japanese. The Dependency account predicts difficulty in both languages, while the Anthropomorphism account predicts similar ease in both languages. The Culture account predicts differences in processing the two languages.

Experiment 1: Comparing the processing of anthropomorphic language in Japanese and English readers

To examine how Japanese and English readers would respond to anthropomorphic sentences with inanimate subjects, we designed a multi-sentence story comprehension study based on the method in (Nieuwland and Van Berkum 2005). We created experimental stories that were five sentences long and the fourth sentence had an inanimate object where the speaker referred to its emotions or goals. One example is shown in Table 1, which is a story where a gambler is telling his friend about losing all of his money. In the expected condition, sentence 4 is “my wallet is empty,” which is a literal sentence that fits in this context. A more unexpected way to say the same thing is “my wallet is crying,” which anthropomorphizes the wallet. We manipulated the expected or unexpected status (Expectation) of the fourth sentence, and it was followed by a fifth sentence that was the same for both stories. By comparing the processing of the fourth sentence with the processing of the third and fifth sentences, we can determine if the unexpected anthropomorphized sentence is harder to process.

Table 1. An example of an experimental story for Exp. 1 with both Japanese and English versions.

The unexpected sentence is in bold, the expected sentence is underlined.

If people find the unexpected anthropomorphic condition hard to process, they should slow down in their reading of the fourth sentence relative to the expected literal condition. However, if processing anthropomorphic language is easy, then we might expect no difference due to the Expectation condition. Since the lack of a difference is a null effect, it is necessary to show that participants are generating expectations in our task. Therefore a separate set of control stories was created. These stories were also five sentences long and they created an expectation that the fourth sentence would be about an inanimate object. If the fourth sentence described an inanimate object, then this was the expected condition, but if the sentence described an animate object, it would be unexpected given the context (unexpected animacy condition). In the example story in Table 2, a man hikes up to a lake, but realizes that he left something in his car. It is typical to forget an inanimate object like a bag, so readers should find the sentence with an animate entity like a dog more surprising. We predict that both Japanese and English speakers will show an expectation difference for the control sentences, which shows that they are generating expectations based on animacy in this task as they read. If there is no difference in expectation between the experimental sentences in one of the two languages, then the control items will show that the result is not due to a lack of sensitivity to expectation in general in those participants.

Table 2. An example of a control story for Exp. 1 with both Japanese and English versions.

The unexpected sentence is in bold, the expected sentence is underlined.

Method

Participants

Japanese participants were 64 adults recruited from an online Japanese crowd-worker website called Crowdworks. Since the crowd-worker website did not restrict the study to native speakers of Japanese, participants were given a series of culture and Japanese language multiple-choice questions, and 3 participants were excluded for scores less than 70% on these questions. To ensure that participants were engaged with the task, the average accuracy at answering the comprehension questions was recorded, and 1 participant was excluded based on a comprehension accuracy of less than 70%.

English participants were 61 adults recruited from the Prolific online crowd-worker website. We restricted participants to those whose first language was English and who were monolingual English speakers. To ensure that participants were engaged with the task, the average accuracy at answering the comprehension questions was recorded, and two participants were excluded based on a comprehension accuracy of less than 70%. All of the studies here were conducted with the formal approval of the Ethics Review Committee for Experimental Research with Human Subjects at Kobe City University of Foreign Studies.

Stimuli

There were eight experimental stories, where the fourth sentence in the unexpected condition was based on one example of anthropomorphic language from our dataset (right side of Table 3).

Table 3. Expected and unexpected version of the fourth sentence in the English experimental condition in Exp. 1.

This fourth sentence combined a predicate that is normally used with humans (e.g., crying, dying, hurting) with an inanimate object (e.g., bread, wallet, toy camera). The story was made up of three sentences that introduced the anthropomorphic 4th sentence and there was one additional sentence after the fourth sentence. The expected condition had the same story as the unexpected condition, but the fourth sentence was changed to be a non-anthropomorphic literal form that expresses the same meaning (left side of Table 3). There were also eight control stories which created an expectation for a sentence about an inanimate element. In the expected condition, the element was inanimate, while in the unexpected condition, the inanimate concept was replaced with an animate element. To control the length of the critical fourth sentence in both experimental and control stories, we required that both the unexpected and expected versions were within two characters of each other (two letters in English, two kana in Japanese). For each story, a true/false comprehension question was created to make sure that participants read the stories carefully. Each trial started with the five sentences from the stories, a sequence of asterisks that signaled the end of the story, and then the comprehension question.

There were also 10 Japanese culture and language questions that were used to check that participants were native speakers of Japanese. The culture questions asked about typical experiences that a person would have growing up in Japan. The language questions were taken from the Japanese Language Proficiency Exam Level 1. Each question had four answers.

Since the stories differ in various ways, we presented each story two times counterbalancing the expectation condition (the fourth sentence was different). This allows for a within-participant comparison of the two versions. Each version was followed by a different comprehension question, which helped to justify why similar stories were presented twice. To keep the two versions isolated from each other, we blocked the items into halves, so one version was seen once in the first half and once in the second half. Since the experimental and control stories are not related to each other, we alternated between them such that each type acts as fillers for the other type. Half of the lists had one order of the stories and the other half had the reverse order. The lists were designed so that the first half was a full counterbalance of the experimental materials. Two practice stories preceded the main test items. Stimuli and statistics files for all studies are available in the online repository (https://osf.io/vxmkt).

Procedure

The study was done using a script that was run in the PCIBEX system (Zehr and Schwarz, 2018). The task involved reading a story sentence by sentence by pressing the spacebar. At the end of the story, a comprehension question was presented. Participants had to press 1 if this sentence was true based on the story and 2 if the sentence was false. After they read all of the stories, the Japanese participants were given the 10 Japanese language and culture questions. For these trials, a question was shown on the screen and four answer buttons were shown below it. Participants had to press the button corresponding to the best answer for the question. On-screen instructions were provided before the study. At the end of the study, a completion code was provided which could be used to receive payment for the study.

Results

To analyze the reading times, the raw reading times were log-transformed to remove the skew in their distribution. It is also important to remove various low-level factors that influence reading times in order to focus more on high-level sentence meaning processing. To reduce the effect of these factors, we fitted a regression model that predicted the log reading times based on various predictors and then extracted the residuals, which are the part of the reading times that cannot be explained by these predictors. One factor that can influence reading times is the position of the sentence in the story, where reading speed increases as more sentences are read. We included sentence position as a categorical predictor in the model to remove the average reaction time for each sentence position. Since longer sentences take more time to be read, another predictor was the length of each sentence in characters. A final factor in this study is that each participant saw both versions of each story and therefore they were faster to read the second version. We included a predictor for the first and second halves of the experiment to remove the difference between them. Finally, the effect of sentence position can depend on first or second reading, so we included the interaction of these two factors in the regression. When this model was fitted, nearly half of the variability in the data was explained (R² = 0.48). The residuals of this model provide reading time information that factors out variation in position in story, length of the sentence in characters, and study half. The analyses focused on sentence positions 3, 4, and 5. Position 4 is where expectation was manipulated and position 3 provides a natural comparison where participants read the same sentence in both versions. Position 5 allows us to see if there are spillover effects, which might be easier to detect since the same sentence was used in both versions. To remove outliers, we removed residual log reading times that were two standard deviations above and below the mean.

A linear mixed-effect model was fitted to residual log reading times with language (English/Japanese), story (experimental, control), expectation (unexpected, expected), and sentence position crossed (Table 4). To allow for correct interpretation of interactions (Brehm and Alday, 2022), the language, story, and expectation variables were sum-coded and sentence position was helmert-coded with one contrast for position 4 vs. the other two positions (p4vs35) and a second contrast for position 3 vs. 5 (p5vs3). Random effects for participants and items were included and the maximal model based on the data had a random slope for language in participants and a random slope for position in items. Likelihood ratio tests were used and chi-square and p-values from those comparisons are reported in the tables (significant terms are marked with a *). The residual log reading times for language, position, story, and expectation are shown in Figure 4. Error bars in this figure and all other figures in this paper show the standard error after random effects have been removed.

Table 4. Mixed model results for Exp. 1.

With respect to the goals of this study, there was an interaction of expectation and contrast for position 4 vs. the other two positions (expectation:p4vs35), which means that reading times for the critical 4th sentence were slower for unexpected than expected sentences. But this was modulated by a four-way significant interaction (lang:story:expectation:p4vs35), which is due to an expectation difference in both stories in English readers and control stories in Japanese readers, but no effect for anthropomorphic stories in Japanese readers (post hoc tests show no difference in this condition, Table 5). These results demonstrate that in the critical position 4, Japanese readers did not find anthropomorphic sentences (e.g., my wallet is crying) to be any more difficult than literal sentences (e.g., my wallet is empty).

Table 5. Post hocs test for a difference between expected and unexpected sentence by position, story, and language for Exp. 1.

One possibility is that the differences in reaction time on the fourth sentence could be due to the frequency of these forms in Japanese and English. To examine this in Japanese, we collected exact phrase frequency information for the argument-predicate combination in the NDL online corpus. For the control items, the expected and unexpected items had the same predicate, but differed in the animacy of the object. The argument-predicate pairs for the expected condition had a mean frequency of 286 in the NDL corpus, while the unexpected condition had a mean frequency of 6,051. The reason that the unexpected condition was more frequent is that animate entities tend to occur more frequently with all control item predicates compared with inanimate entities. This frequency pattern would predict faster reading in the unexpected condition, but in our story task, readers are slower in this condition for control items. For the experimental sentences, the expected and unexpected fourth sentences had the same inanimate object, but differed in whether the predicate was normally used with inanimate entities or not. For the expected literal predicate (e.g., my wallet is empty), the mean frequency was 1,058, which shows that most of these predicates are compatible with inanimate objects. However, the unexpected predicate (e.g., my wallet is crying) had a mean frequency of 3, which shows that the anthropomorphic argument-predicate pairs in our stimuli were not common in the NDL corpus. Even with the large difference in frequency, readers found these unexpected sentences easy to read and not different from the expected sentences. This analysis demonstrates that ease in processing Japanese anthropomorphic language was not due to the frequency of these forms in the input.

A similar exact phrase frequency analysis was done in English using the English NGRAMs website (Trenkmann, 2023). For the experimental items, the expected items were more frequent (688; e.g., my wallet is empty) compared to the unexpected items (0; e.g., my wallet is crying). Thus, input frequency might be one factor that influences the reading times in our study. But for the control items, there were similar mean frequencies for the expected items (4,322) and unexpected items (5,821). The fact that the expectation difference for the control stories was similar in magnitude to the experimental stories suggests that the frequency in the input does not fully explain the reading time differences in the English participants.

Experiment 2: Replication of novel cultural difference finding with additional items

The first experiment found a difference in the processing of anthropomorphic language. English speakers read anthropomorphic sentences more slowly than literal sentences, while Japanese speakers read both equally fast. These findings are at odds with the Dependency account, which predicts difficulty with anthropomorphic sentences in both languages (Bornkessel and Schlesewsky, 2006; Gibson, 1998). They are also not consistent with the Anthropomorphism theory, which predicts ease of processing personification in both languages (Epley et al., 2007; Lakoff and Johnson, 1980). Since this is the first study that has found that Japanese anthropomorphic language can be processed as easily as literal language, we thought it was important to replicate our results with a new set of participants. To increase the reliability of the results, we added 4 experimental and four control stories such that each condition had 12 stories (increasing the number of stories by 1.5 times).

Method

Participants

Japanese participants were a new set of 67 people recruited from the Crowdworks website. To ensure that participants were engaged with the task, a threshold of 70% on the comprehension questions was used, and six participants were excluded. An additional four participants were excluded due to scoring less than 70% on language and cultural questions. English participants were a new set of 69 English monolingual speakers (first language is English) recruited from the Prolific website. We excluded seven participants due to a comprehension accuracy of less than 70%.

Procedure and stimuli

The study was done using a script that was run in the PCIBEX system in a method similar to Exp. 1. The original 8 experimental and control stories from Exp. 1 were used. In addition, four more experimental stories based on the examples in the dataset were added (Table 6).

Table 6. Additional expected and unexpected fourth sentences in the English experimental condition in Exp. 2.

Expected and unexpected versions of the fourth sentence in the experimental condition.

These stories involved different objects from the original set and included examples from our dataset produced by older men in more formal settings (e.g., a veterinarian, a dresser for geisha). An additional four control stories were also created. Comprehension questions were created for the new stories. The same Japanese culture and language questions for Exp. 1 were used. A new counterbalance was created with each of the 12 stories appearing two times following the same principles as the counterbalancing for Exp. 1.

Results

To remove the effects of sentence length, position in sentence, and experiment half from the reading times, reading times were residualized and cleaned in the same way as Exp. 1 (R² = 0.44). A linear mixed-effect model was fitted to residual log reading times with language (English/Japanese), story (experimental, control), expectation (unexpected, expected), and sentence position crossed (Table 7). Sentence position was helmert-coded with one contrast for position 4 vs. the other two positions and a second contrast for position 3 vs. 5. All other variables were sum-coded. Random effects for participants and items were included and the maximal model based on the data had no random slopes. The residual log reading times for language, position, story, and expectation are shown in Figure 5.

Table 7. Mixed model results for Exp. 2.

resLogRT ~ story ^* expectation ^* p4vs35 + story ^* expectation ^* p5vs3.

With respect to the goals of this study, our expectation manipulation at sentence 4 worked, because we found that unexpected sentences were read more slowly than expected stories at that sentence relative to other positions (interaction of expectation:p4vs35). As in Exp. 1, there was a four-way interaction (lang:story:expectation:p4vs35), which is due to a lack of an expectation difference at sentence 4 for anthropomorphic stories in Japanese readers in contrast to the significant difference in all of the other sentence 4 combinations (post hoc tests for each position, Table 8).

Table 8. Post hocs test for a difference between expected and unexpected sentence by position, story, and language for Exp. 2.

In this study, the effect of expectation spilled over into sentence 5 (interaction of expectation:p3vs5) even though the same sentence was used for each story, but there were no interactions of language and expectation with position 5, so language differences in expectation did not seem to be strong at that position.

As with the previous study, we collected Japanese frequency information in the National Diet Library online corpus for the argument-predicate combinations in the 4th sentence from our data. Due to the preference for animate objects, the unexpected control items had a higher mean frequency (5,041) than expected control items (843), which would predict the opposite pattern that was found in the control story condition. For the experimental items, the expected condition was higher (731) than the unexpected condition (2) and this would predict faster reading times for expected items, which was not found. Thus, frequency does not explain why Japanese speakers found anthropomorphic language easy to process. For the English data, we collected frequency for search terms related to the fourth sentence in this study from the NGRAMs website (Trenkmann, 2023). For the 12 experimental items, the expected items were more frequent (1,385) compared to the unexpected items (69). These corpus frequencies are consistent with the large slowing for unexpected sentences compared to expected sentences at position 4 in this study. For the 12 control items, the expected items (2,946) were less frequent than the unexpected items (3,896), due to the fact that animate entities were more common in general, and this would predict the opposite pattern that was found in the control sentences results.

Combined Japanese/English analysis

Since these studies are the first to claim that there are culturally-unique Japanese anthropomorphic forms, it is important to provide a strong test of this claim. To do this, it is useful to combine all of the Japanese and English data into one analysis. In the experiments, the stories were presented twice to each participant (with a different fourth sentence) and the second presentation may be influenced by the first. Although previously we factored out this effect by using residualization, a stronger approach is to analyze only trials from the first exposure using the first half of each participant's data. Furthermore, in the previous analyses, we included position 5 in order to test for spillover effects in case there was no difference at position 4, but in this combined analysis, we now remove position 5, which allows us to focus on the difference between position 3 and position 4. This combined analysis provides a new insight into our data by focusing on the reading of the first story while at the same time increasing power by using a combined pool of 261 participants.

Log reaction time was residualized using the first half of each participant's data and cleaned as before. Since the two replications in each language are just different in the number of stories, we did not distinguish the replications in order to focus on the language differences. A linear mixed-effect model was fitted to residual log reading times with story (experimental, control), expectation (unexpected, expected), sentence position (3, 4), and language (English, Japanese) crossed (Table 9). All variables were sum-coded. Random effects for experiment-specific items and participants were included and the maximal model based on the data had a random slope for position in items.

Table 9. Mixed model results for combined analysis of data from both languages.

With respect to the goals of our study, participants were slower for unexpected than expected sentences at position 4 vs. position 3 (interaction of expectation:position). But this was modulated by a four-way interaction (language:story:expectation:position), which is due to the lack of an expectation effect at position 4 in experimental stories in Japanese readers (Figure 6). The lack of a difference is not a problem of power, because the direction of the effect is in the wrong direction in both experiments and this combined analysis (expected is slightly slower than unexpected at position 4). If this effect of expectation in the Japanese experimental stories was made significant with additional participants or items, the wrong direction of the effect would still provide evidence that Japanese readers are processing these sentences differently from English readers. Furthermore, significant interactions require a larger sample size compared to main effects (Gelman, 2018), so the fact that we have a four-way interaction in all of the analyses suggests that there was sufficient power to detect a difference in the anthropomorphic stories in Japanese readers, if that effect existed.

Over the two studies in English and Japanese, readers generated expectations about the animacy of elements in the stories that they were reading, and they were slower when those expectations mismatched the contents in the fourth sentence. But the one exception was that Japanese readers showed no surprise when they read an anthropomorphic sentence, which suggests that they view this type of language as a normal way to talk about inanimate objects. This is especially interesting in light of the fact that they appear to show a larger difference for expectation in the control stories, which shows that they may have stronger expectations about animacy. The difference in processing in each language is not consistent with theories that predict similar processing in both languages (the Dependency account predicts similar difficulty, the Anthropomorphism account predicts similar ease). Instead, the results appear to support the Cultural account, where the Japanese readers have a way to easily understand the anthropomorphic language.

Discussion of cultural factors in the comprehension of anthropomorphic language

The Cultural account argues that there are differences in the processing of anthropomorphic language in English and Japanese. In this section, we review some features of Japanese culture, which might make anthropomorphic language using simple objects more common. We are not claiming that this cultural knowledge directly explains the results in our studies, but rather it provides a background context that might indirectly influence readers. One factor that may help to explain why anthropomorphism is relatively common in Japan is religion (Hatano and Inagaki, 1994). In contrast with many western religions, the Japanese religion of Shinto is a belief system that centers around god-like supernatural entities called kami (Teeuwen and Scheid, 2002). These kami are contained within non-moving inanimate objects like trees or rocks (called shintai, Martinez, 2005) and these shintai can then be subject to anthropomorphic thinking. For example in Koshikiiwa Shrine near Kobe, there is a rock shintai that contains a kami related to fertility and the protection of women (left side of Figure 7), but since people may not distinguish between the kami and the shintai, some may think of the rock itself as the god that supports childbirth. Another example are “married” rocks or trees, which are found all over Japan. In Sumiyoshi Shrine in south Osaka, there are two camphor trees that are called the Married-Couple Trees of Nankunsha (middle of Figure 7). These two trees are tied together with a large shimenawa rope that is usually placed around shintai. The idea of treating non-human objects as married couples has a long history evidenced by wood-block prints from the 1800s of these married-couple rocks, meoto iwa (Kunisada, 1844; right side of Figure 7).

Unlike western religions where believers tend to select one religion as their main belief system, Japanese people often engage in religious ceremonies for different events. Marriages are often performed with Christian or Shinto rituals, while funerals are often performed using Buddhist ceremonies called kuyō. Funerals in Japan typically involve burning the body and chanting prayers for the deceased person. In addition to these funeral rites for people, Japanese Buddhist temples also perform funeral rites for inanimate objects that have acquired some human associations through their use. There are yearly funeral ceremonies for dolls, personal stamps, glasses, mirrors, needles, writing brushes, and tea ceremony whisks (Figure 8). In some special cases, there have been funeral rites for cell phones, robot dogs, and statues (Burch, 2018; Uechi, 2024). These ceremonies appear to have a long history with a needle funeral hall created at Horinji Temple in Kyoto by 881 (Harikuyo horinj [“Needle Memorial Service” at Horinji Temple], 2025). As in human funerals, the inanimate objects are burned and prayers are chanted. By giving these objects a human-like funeral, they are anthropomorphized and treated as equal to humans in this way. Thus, Japanese people participate in various religions that appear to anthropomorphize inanimate objects and this culture has existed in Japan for a long period of time.

Although Japanese religion may have influenced anthropomorphic language in the past, it is not clear that religion creates unique cultural beliefs about inanimate objects in modern Japanese people. For example, the Japanese General Social Surveys (2000–2003, 2004) found that less than 10% of the sample were associated with a religious group and none reported Shinto as their personal religion (less than 5% were Buddhist, Roemer, 2010), so it is not clear the extent that people know the beliefs of these religions. If Japanese people are culturally-biased to apply personification to inanimate objects (perhaps due to the view that all objects have gods within them, yaoyorozu no kami), then we might expect children to show this most strongly (Gjersoe et al., 2015). Experimental studies have found that 5–6 year old Japanese children do not believe that a rock can think, feel happy, or feel pain (Inagaki and Hatano, 1987), but later studies found that Japanese children were more likely to attribute animal properties like the ability to see or move to inanimate objects like a rock or chair (Hatano et al., 1993). An alternative possibility is that Japanese people may be using a human-group-based personification schema (Figure 3). One reason that this schema might be accessible is that Japanese society is thought to be more collectivist and group-oriented than western society. However, research on attitudes about conformity and collectivism often yields mixed evidence with some meta-analyses showing that Japanese people are more individualistic than western people (Oyserman et al., 2002; Takano and Sogon, 2008). Future studies are needed to determine how religion and collectivism influence the processing of anthropomorphic language in contemporary Japanese people.

Conclusion

Anthropomorphism is typically used to refer to the attribution of human-like mental states to animals (e.g., my cat loves me) or complex machines (e.g., my computer is sleeping). But in this work, we examined the use of anthropomorphic language in Japanese with a range of simple non-moving inanimate objects that should be hard to anthropomorphize (e.g., lettuce, belts). We collected a dataset with a range of Japanese examples of this type of language and found that many of them were inanimate objects that are low in the ability to think or have goals. The analysis of our dataset demonstrated that many of these forms are not found in a 1.7 billion-word Japanese corpus, which suggests that many of these argument-predicate combinations are not idiomatic.

Dependency theories of comprehension predict that argument-predicate mismatches should induce comprehension difficulties (Bornkessel and Schlesewsky, 2006), and experimental studies have confirmed this prediction in various languages including Japanese (Kim and Osterhout, 2005; Nieuwland and Van Berkum, 2005, 2006; Onishi and Murphy, 1993; Yano and Sakamoto, 2016). Consistent with these theories, our story-reading experiments found that argument-predicate mismatch in the control stories yielded slower reading times in both Japanese and English readers. Further support can be found in the difficulty that English readers had with reading the anthropomorphic language in the experimental items. But when Japanese speakers read these sentences, they showed no difference, suggesting that the argument-predicate mismatch mechanisms that created the other slowing effects were not active in this condition in this language. A minimal modification of Dependency theories to explain this result would be assume that argument-predicate linking processes can be influenced by culture-specific schemas, and Japanese readers have some unique schemas which support the processing of anthropomorphic language.

One possible factor in reading time differences is the frequency of the anthropomorphic and literal forms in each language (MacDonald and Hsiao, 2018). But corpus analyses in both languages found that the literal form (e.g., my wallet is empty) was always far more frequent than the anthropomorphic form (e.g., my wallet is crying). In fact in the last two studies, the average frequency of the literal form in Japanese was more than 365 times as frequent as the average frequency of the anthropomorphic form, while in English it was only 21 times as frequent. Even with this larger frequency multiple, Japanese speakers showed no difference in their reading speed for the literal and anthropomorphic forms. Therefore, input frequency does not seem to be the main factor controlling reading times in the two languages.

The independence from exact phrase frequency in the dataset and experiments suggests that the processing of anthropomorphic languages uses abstract schemas that can deal with novel forms. One way to explain the ease with which Japanese readers dealt with anthropomorphic language is to assume that they have culturally-specific schemas (such as the group-based personification schema, Figure 3), which allows them to easily apply human predicates to simple non-moving inanimate objects. If this schema is abstract, it might collect frequency across a range of forms, which should increase its accessibility. If English speakers do not have access to these schemas, they may be slower to process language with these types of targets.

Although we found a difference between English and Japanese processing of anthropomorphic sentences, much is still unknown about this finding. These results do not imply that English speakers cannot process this type of anthropomorphic language, but rather that there may be a difference in the strength or broadness of the ability in the two cultures. Although differences in the anthropomorphic schemas in each language may explain why the Japanese sentences are easier to process, it is also possible that non-anthropomorphic cultural/linguistic knowledge is influencing these results. In addition, anthropomorphic language applied to objects with mind-like elements might yield different results. There may be culturally-unique examples of English anthropomorphic language that are difficult for Japanese readers. Furthermore, there may be examples of anthropomorphic language in the dataset which do not fit with either mind-based or group-based schemas. This research is lonely and there are still many issues that need to be examined.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. Grammarly.com was used to check the grammar and spelling in the article. AI was not use for generation of the text.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Al Ryanne Gatcho, Hunan Institute of Science and Technology, China

Reviewed by: Xinmiao Liu, Beijing Foreign Studies University, China

Jiashen Qu, Kanda University of International Studies, Japan