In this paper, we apply machine learning based predictive models on two large data sets of historical and contemporary 昀椀ction to better understand the role that things play in 昀椀ctional writing. A large body of scholarship known as “thing theory” has attempted to understand the function of 昀椀ctional things within literature mostly by focusing on small case studies. We provide the 昀椀rst-ever estimates of the distribution of di昀erent types of things in English-language 昀椀ction over the past two centuries along with experiments to model their semantic identity. Our 昀椀ndings suggest that the most common 昀椀ctional things are structural in nature, functioning akin to narrative props. We conclude by showing how these 椀昀ndings pose problems for inherited theories of 昀椀ctional things and propose an alternative theoretical framework, embodied cognition, as a way of understanding the predominance of structural things.
Over the past two decades, a large body of research has emerged in the 昀椀eld of literary studies
focusing on the question of “things.” “Thing theory,” as this area has come to be known 8[],
has its origins in di昀erent research traditions, including the rise of material cultural studies
[1], new historicism [
In this paper, we apply machine learning based predictive models on two large data sets of
historical and contemporary 昀椀ction to better understand the role that things play in 昀椀ctional
writing. We de昀椀ne things for our purposes as any non-human object and thus distinguish them
from people, places, or human body parts like faces and eyes. Our aim is to better understand
the function that imaginary objects play in the practice of creative storytelling and how this
has potentially changed over time. Despite a wealth of recent case studies that focus on
particular types of things in individual books1[
Accordingly, we have two primary aims in this paper. First, we provide the 昀椀rst-ever estimates of the distribution of things in 昀椀ctional writing over the past two centuries. For the 昀椀rst time, we can observe the kinds of things that populate 昀椀ctional writing at least with respect to a single major language. Second, we conduct a series of experiments to better describe the semantic identities of things as a way of inferring their narrative function. Doing so can help us engage with longstanding debates about the nature and role of 昀椀ctional objects within the 椀昀eld of thing theory.
Based on the data we provide here, we 昀椀nd that the most predominant type of 昀椀ctional things since the nineteenth century are “infrastructural” in nature. By infrastructural we mean three key qualities: the 昀椀rst, ontological in nature, is that the most common types of things in 昀椀ction are more likely to bestructural rather than, say, instrumental or vehicular or organic. The most prevalent things in 昀椀ction hold other things up. They provide narrative support. Thus for us infrastructure does not more narrowly refer to the contemporary usage of “airports,” “roads,” etc., but rather to human-made structures more generally, such as rooms, doors, tables, walls, and homes that help support other objects.
The second way we mean infrastructural is that these things are also distinctively semantically inert: as we will show through di昀erent experiments, structural things are less descriptively rich than other kinds of entities, less a昀ectively charged, and more associated with physical rather than cognitive or emotional behavior. Such things are more semantically inert than other kinds of entities. They appear to function more as vehicles of embodiment than intellection. Third and 昀椀nally, we mean infrastructural in the sense that such structural things appear regularly throughout narrative time. They are consistently woven throughout the whole narrative rather than functioning as singularly important entities. Their continued prevalence is a key part of their meaning.
Overall, then, the most predominant things in 昀椀ction appear to function as “props” in multiple senses of the word: they physically hold other things up; are descriptively shallow; and predominate throughout narrative space and time. As we discuss at the close of our paper, we think these insights have important implications in modifying the inherited views of thing theory and the role that 昀椀ctional things play in literature.
The primary datasets we use for this paper are, 昀椀rst, the Hathi1M dataset which consists of
1,671,370 randomly drawn pages from over 300,000 volumes of English prose in the Hathi Trust
digital library3[]. These volumes span the years 1800-2000 and consist of an approximate
balance of pages labeled as 昀椀ction and non-昀椀ction. The second, the CONLIT dataset [
In order to detect entity-types in our data, we process all texts using bookNLP4][, which implements BERT-based models and has been shown to outperform other state-of-the-art systems for a variety of NLP tasks when applied to literary texts5][. In addition to performing entity recognition, part-of-speech-tagging, and dependency parsing, bookNLP also provides 41 “super-sense” tags trained on SemCor’s implementation of the Wordnet taxonomy, examples of which are shown in Table 1. In Table 2, we provide a list of noun-types that we use to capture two categories of things as either human-made or organic. According to bookNLP’s documentation, the supersense tagging has an accuracy of 76% when applied to works of 昀椀ction.
Prior work on the computational detection of objects in 昀椀ction has applied a
dictionarymatching approach using the Wordnet taxonomy to a small set of sixty nineteenth-century
novels [
Fig. 1 provides an overview of the distribution of noun-types across three historical periods. As we can see, humans dominate 昀椀ctional narrative regardless of historical time-frame. Nevertheless, human-made things (“artifacts”) are second, a surprising result given no prior theory has indicated this kind of predominance. Indeed, we 昀椀nd that the rate of artifacts across all three collections is 2.4x higher than all other natural objects combined and 3.3x higher if we condition on the contemporary data. As Fig. 1 indicates, the prevalence of artifacts increases from the nineteenth-century to the present (note the reversed order of persons and artifacts by period).
We can also observe that this increase of artifacts over time is speci昀椀c to 昀椀ction (Fig. 2). Despite the simultaneous growth of artifacts in 昀椀ction and non-昀椀ction over the course of the nineteenth century, by the early twentieth century we see 昀椀ction investing more heavily in the utilization of human-made things. By the end of the twentieth century, artifacts in 昀椀ction are almost 50% more frequent than in non-昀椀ction, averaging roughly 3 more occurrences per page than an equivalent work of non-昀椀ction. Artifacts also represent the semantic category with the single strongest growth over this time period. Human-made things come to dominate 椀昀ction beyond what we expect from their overall prevalence in English-language writing more generally.
To better understand the nature of human-made things and their distribution in 昀椀ction, we break down the artifact category by various sub-types. To do so, we use Wordnet’s hypernym taxonomy. For each word predicted to be an artifact in our dataset, we crawl the hypernym tree for the word’s most common sense and store all types that occur below the level of “artifact.” For example, the word “house” would produce the following sub-types in the following order of ascending generality:dwelling, housing, structure. We condition on 2,500 / 1,800 of the most common artifacts in the contemporary / Hathi collection, which account for just under 80% / 90% of all occurrences of artifacts in each collection. We then manually review and clean the labels, producing the following sub-types and their counts in Table 3.
As we can see, “structures” are the most common kind of object in 昀椀ction by a factor of 1.5 over the next most frequent type, human-made instruments. Indeed, the frequency of structures is higher than all of the organic objects combined. This holds for the Hathi1M collection with the exception that there are considerably more natural objects utilized in the nineteenthand twentieth-century data. Beyond this divergence, the ranking and even the top words of the various categories are remarkably similar (full data provided in the supplementary material).
With a clearer picture of the distribution of the types of 昀椀ctional things, we take the next step of trying to better understand the semantic identity of these things, and structural things in particular. Knowing that human-made things dominate 昀椀ction in terms of object types and knowing that within that group structures are the most common sub-type can give us important information about the potential narrative function of 昀椀ctional things. Here we wish to deepen our understanding by undertaking a series of tests to gain further insights about the semantic frameworks that surround structural things. We note that when we refer to “structures” in this section we combine words labeled as “structures” and “furnishings” because we consider tables, chairs, etc to also function as structural forms for our purposes.
Wall [
As we can see in Table 5, when it comes to modi昀椀cation we see a small increased rate of modi昀椀cation of artifacts when compared to other kinds of entities, which shi昀琀s in the opposite direction when we condition only on structures and furnishings. Structures are unique in their lower levels of description. For a昀ect, we see a small negative relationship between artifacts and their valence scores for the historical data and a large negative relationship for the contemporary collection, which grows stronger as we move from all artifacts to our subset of
Category artifact artifact artifact organic organic organic artifact artifact organic artifact artifact artifact
Type structures. This suggests that structures are functioning less ornamentally and less a昀ectively than other kinds of 昀椀ctional entities.
Another way to understand the meaning of objects in 昀椀ction is to observe their distribution across narrative time 1[0]. Fictional things may play a strategic role by occurring at key turning points or narrative entry or exit points or conversely they may play a more infrastructural role by occurring regularly throughout narrative time. As Boy1d0][has shown, di昀erent linguistic features exhibit meaningful changes over narrative time. In Figure 3, we plot the fraction of structures and other artifacts over narrative time for our contemporary collection. As we can see, we observe a preference for human-made things to appear towards the opening sections of a narrative, suggesting that things function partially as “narrative establishments.” However, in real terms the change is very slight, with a decline of roughly 15 total mentions of structures between the 昀椀rst and 昀椀nal sections of an average narrative. In other words, evidence also suggests that in addition to establishing narrative frames, structures and other artifacts also play an infrastructural role in their continued presence over narrative time.
Finally, we explore collocate word types associated with structures to better understand how
they are semantically contextualized. As the theory of distributional semantics suggests, the
company that words keep has a strong in昀氀uence on their meaning [
As we can see, structures are most strongly associated with actions associated with contact and motion and entities associated with substances and locations, while they are negatively correlated with actions of communication, cognition, and emotion. In other words, when readers encounter structural things in 昀椀ction, they are considerably less likely to encounter moments of a昀ect or cognitive re昀氀ection. Structural things are more likely to cause characters to move and touch, not think and emote.
Summarizing our 昀椀ndings to this point, we can say that according to our data the most common kinds of things in 昀椀ction are structural in nature, i.e. human-made and supportive. These may include rooms, houses, doors, windows, tables, roads, kitchens, walls, stairs, apartments, among many other kinds of things. Such things are less likely to be modi昀椀ed and less likely to be associated with a昀ective feelings. They are consistently deployed throughout narrative time and they are more likely to facilitate corporeal rather than cognitive or communicative behavior in characters.
These 昀椀ndings raise challenges for some of the more prominent scholarly arguments that have fallen under the heading of thing theory. Brown’s claim that 昀椀ctional things are important because of their “labor of infusing manufactured objects with a metaphysical dimension”11[] or Calvino’s claim that “in a narrative any object is always magic” (331)3[] do not 昀椀t well with the semantic nature of the structural things that we have shown here. An individual door or room may be magical or metaphysical (e.g. the door in Tieck’Bsluebeard’s Castle or Gregor’s room in Ka昀欀a’s Metamorphosis), but broadly speaking such structural things are less likely to be associated with ornamentation, a昀ect, or cognition. Their semantic inertia appears to matter more to their identity rather than their individual a昀ective or descriptive depth.
Similarly, Freedgood’s1[
Barthes [
The framework we would suggest falls under the heading “embodied cognition,” an
increasingly studied (and still debated) framework within cognitive science26[] that has also found
increasing resonance in literary studies2[
One way to understand the predominance of structural things in 昀椀ction, then, would be as a means of activating this idea of “embodied cognition” through 昀椀ctional narration. Structural things shi昀琀 the focus, as Randall Beer, one of the early proponents of embodied cognition, argued, “from accurately representing an environment to continuously engaging that environment with a body so as to stabilize appropriate co‐ordinated patterns of behavior” (97)7][. The strong corporeal and low decorative aspects of such objects may potentially produce, in Andy Clark’s words, “a constantly available channel that productively couples agent and environment” (15) [15]. The persistent recurrence of such objects are thus neither narratively useless nor to be understood principally as vehicles of introspection. Rather, they may be means of activating an experience of “the extended mind,” enabling the experience of embodied cognition in the minds of readers through these objects’ recurrent physical and semantically inert presence. Seen in this light, one of 昀椀ction’s modern social functions could be that it helps readers activate the particular mode of thought known as embodied cognition.
By empirically accounting for the types of things in 昀椀ction and their semantic identity across large number of documents, our 昀椀ndings pose challenges to inherited critical theories about the role of things in 昀椀ctional narratives. While those theories are undoubtedly valid for the individual objects and works they address, they fail to account for the most predominant kinds of 昀椀ctional things and their semantic behavior across large amounts of literature. Rather than argue that such things are unimportant, we contend that it is highly important to account for the most prevalent kinds of things in 昀椀ction if we are to understand the function of 昀椀ctional things.
While we suggest embodied cognition as one possible avenue for understanding the function of such structural things, future work can test this theory further either through more text samples drawn from di昀erent cultural domains and languages or through empirical reader studies. Work in embodied cognition has a long history of measuring human attention and problem solving and many of these approaches could be productively applied towards understanding reader behavior. The coupling of large-scale observational data with empirical reader studies o昀ers an ideal synthesis to continue to better understand the social and psychological functions
This research was generously supported by the Social Sciences and Humanities Research Council of Canada (895-2013-1011). [3] S. Bagga and A. Piper. “HATHI 1M: Introducing a Million Page Historical Prose Dataset in English from the Hathi Trust”. InJ:ournal of Open Humanities Data 8 (2022).