A fundamental feature of many genres of fiction is the alternation between a narrative frame (NF) and quoted inset (QI) dialogue. Both formal and representational features distinguish NF and QI segments. This paper is an explorative study on the stylistic diferentiation between frame and inset material in recent commercially successful fiction in Swedish. There are mainly two orthographic options as regards this distinction: Explicitly enclosing inset segments within quotation marks is one. Using an initial dash to indicate utterance display is another, in which case frame and inset material typically alternate in a way not made explicit by the orthography. The corpus behind the present study comprised 450 novels. In order to deal with dash orthography data (135 books), we trained a multilayer perceptron classifier to tell NF and QI segments apart. We relied on the fact that native quotation mark text can be converted to annotated dash orthography data, which can then be used for supervised training and validation. A small-scale manual evaluation on the texts we aim to analyze, yielded an accuracy score around 95%. In order to explore the stylometric relations between NF and QI components in the novels, we looked at a selection of basic grammatical features. A characterization of each feature was made by means of recording the fraction of works in which the relative frequency of the feature is higher in QI than in NF. This summarizes how authors tend to “use” that feature to create a contrast between NF and QI. Another way to examine how the NF and QI styles are related is to apply a correlation test. We then saw, for instance, that QI material in 100% of the books are denser in auxiliary verbs, second person pronouns, and interjections, while NF segments in all or almost all cases are denser in nouns, adjectives, third person pronouns, and prepositions. We could also observe that e.g. noun density in NF and QI correlate in a strong way. The same holds for adverbs and cardinal numerals. This suggests that books and authors exhibit stylistic tendencies which afect both narrator and the characters, as far as the kind of fiction we have studied go.
The present study is concerned with recent commercially successful fiction in Swedish. It will focus on a stylometric exploration of the diferentiation between frame narration and quoted inset material, typically dialogue. A component in this work was the training of a classifier which can tell the two kinds of content apart in novels written in “dash orthography”, which makes this a non-trivial task. CEUR Workshop Proceedings
A fundamental feature of many genres of fiction is the alternation between a narrative frame
(NF) and quoted inset (QI) dialogue, see Figure 1. Dialogue is the mimetic, more or less verbatim
display – using the device of “direct speech” –, of utterances made by the characters in the
story as chains of events play out. QIs are distinguished by both formal and representational
features. One of the former is directness, i.e. “the inset’s syntactic and deictic independence of
the frame” [1, p. 111]. The distinction between quoted dialogue and other modes of narration is
not always a matter of a simple binary opposition, but the analysis here will treat it as such.
This simplification is for the most part adequate for the kind of fiction we analyse here. It is also
possible for authors to use indirect speech and free indirect speech – as illustrated in examples
(3) and (4), respectively, below –, to mention the two most well-known alternatives. Speech
acts can also be described in ways that only summarize their content or focus on other aspects
(cf. e.g. [
In this article we will address two research questions in an exploratory fashion. The first one
is instrumental: How well can we separate inset and frame text automatically? Secondly, we
will investigate whether and how a systematic exploration of stylometric features can be used to
ifnd and illustrate diferences between inset and frame text as well as correlations between the
two embedded styles over a corpus of novels. The method and aims behind this study belong to
the school of “distant reading […] focus[ing] on units that are much smaller or much larger
than the text: devices, themes, tropes” [
In Swedish fiction, there are two main orthographic options for diferentiating between NF and QI material. The most common and most explicit one is enclosing quoted segments within quotation marks (””). The quotation marks serve as an explicit markup unambiguously indicating the start and end of directly quoted material. Example (1) illustrates this, with the oficial English translation exhibiting the explicit style. Using a dash (–) to indicate the start of an utterance display paragraph – as in the original (2) – is a less explicit and somewhat less common (30% of the books in our corpus, see below) orthographic device.
(1) ”För många träfar”, säger han. ”Tiden rinner ut.” [Adapted.]
“Too many results,” he says. “Time’s running out.” [Oficial English translation. 1]
1Lars Kepler, Stalker, Alfred A. Knopf, 2016. Translation by Neil Smith.
(2) – För många träfar, säger han. Tiden rinner ut. [Lars Kepler, Stalker, Bonniers, 2014.]
– Too many results, he says. Time’s running out. [Adapted.]
(3) He said that there were too many results and that time was running out. [Adapted.]
(4) [He was pessimistic.] There were too many results. Time was running out. [Adapted.]
In dash-style novels we often find an unmarked alternation between frame and inset segments
in the same paragraph, which would have been unambiguous had quotation marks been used.
So, the “inquit” formula he says in (2) belongs to the NF, but could also – quite easily if you
disregard the wider context – be understood as being part of the QI. More extended frame
material surrounded by quotation is known as “suspended quotation” [
Supplementary Materials providing details on the corpus and a more complete presentation of results is available at URL https://github.com/mdahllof/dhnb2022sm.
The analysis of dialogue and narrative goes back to Plato’s opposition between μίμησις and
διήγησις, or between showing and telling, in modern critical parlance. Sternberg [
Still, it has often been assumed in corpus-based literary studies that fiction is one register
and that a novel, for instance, exhibits one style. Egbert and Mahlberg [
Quoted inset recognition as a problem in NLP has been addressed in a number of studies,
which have been motivated by the central importance of the task for computational literary
studies. Recent work, e.g. [
The corpus behind the present study comprises a collection of bestsellers and beststreamers in
Swedish, including translations, in the Swedish book market 2015–2020, along with a batch of
original productions in Swedish from the online streaming platform Storytel. The bestsellers
are the novels which have earned that status according to the Swedish Publishers’ Association
(SvF), either in hardback or paperback. As beststreamers (cf. [
In the classifier training and the stylometric analysis we distinguished the following three categories of paragraph.
• NF only paragraphs, i.e. those not beginning with a dash (–) or containing quotation marks (”). Henceforth: NF paragraphs. • Typical QI paragraphs, i.e. those beginning with QI material as indicated by an intial quotation mark or dash. They are often QI only, but there is often an alternation between QI and NF text. Henceforth: QI/NF paragraphs. (Our classifier targeted the QI/NF separation in the dash-orthography variety of such paragraphs.) • Other paragraphs, i.e. those with non-initial explicitly quoted text or non-matching quotation marks. They were excluded from the quantitative analysis.
The machine learning design, as well as the stylometric analysis, departed from data tagged
with part-of-speech (POS) labels. We used the Stanza [
We did not explore the possibilities for classifier design in depth, but managed to find an approach which can be considered satisfactory, at least for the purpose at hand.
The paragraphs were segmented at punctuation marks, where QI/NF shifts may occur. This leaves us with short snippets to be classified. Features were derived from both the snippet text and the preceding narration context, and involved POS unigrams and bigrams, word forms, as well as pronoun person and verb tense. This allows the classifier to take advantage of both lexical information and point of view diferences between the narration and the dialogue. A multilayer perceptron with three layers, each comprising 15 neurons, was found to be the optimally performing architecture, with 800 features selected. (The classifier pipeline can be found among the Supplementary Materials.)
We applied a cross-validation setup which was three-fold over 315 books (see first row in
Table 1) taking, for training, 80,000 randomly selected snippets from 210 books and and as
testing data all snippets from all dialogue paragraphs from the 105 other books. We then saw
accuracy scores from 94.2% to 94.6% (on word token level: 93.2% to 93.4%).4 A small-scale
manual evaluation of the best-performing of the three classifiers on the kind of data we actually
aim to analyze, i.e. dash-orthography novels, yielded an accuracy score of 95.7%, based on 282
randomly selected snippets from the 135 relevant novels in our corpus. This performance is
dificult to compare with previous approaches, but we can note that [
In order to explore the stylometric relations between NF and QI components in the novels, we
took a selection of very basic grammatical features (listed in Table 2) as our point of departure.
They comprise parts of speech, pronouns according to grammatical person, and negation (only
inte). These were quantified as relative frequency among word tokens. We also included the
most important inflectional categories, quantified as ratio in tokens of the applicable part of
speech. (The full table of novels times feature values is available in the Supplementary Materials.)
A survey of a population that includes all elements defined by a set of criteria, like the present
one, does not face a problem of sampling bias. This means that simple descriptive statistics are
relevant, while significance testing is not applicable as in sample-based studies [
A direct way of visualizing the QI and NF values for the various books and features is to scatter-plot them against each other as in Figure 2, which shows four examples of features. The NF values for preterite clearly distribute the works bimodally, forming a larger high-value cluster and a smaller low-value cluster. The QI values are clearly less dispersed, suggesting that authors to a high extent “agree” on the proper average amount of preterite in fictional speech. This does not correlate to any visible degree with the preterite density in the NF. (The plot for present tense, see Supplementary Materials, presents a similar shape upside down.) Looking at another plot, we see that the noun density is quite dispersed both in NF and QI. However, the NF value is always higher than the QI value, even if the ranges of possible values overlap. Furthermore, the QI and NF values correlate. In other words, there is a tendency four noun-dense narrators to be coupled with noun-dense characters. We see the same kind of correlation for adverbs, but in most books (86% of them) the QI material is denser in adverbs than the NF. Finally, in Figure 2, there is a plot for negation only showing books by the ten most prolific Swedish authors in the corpus. The distribution is roughly as for adverbs in general, but we also see that most authors cluster in particular regions, i.e. they use negation consistently in their books in both NF and QI. Lars Kepler, for instance, is low-negation in NF, but high in QI. Sofie Sarenbrant is high-negation in both NF and QI, even extremely so in NF. Fractions of works in which the QI value exceeds the NF value for the features explored and Spearman’s Feature p1 (first person pronoun) p2 (second person pronoun) p3 (third person pronoun)
QI > NF (%) neg(ation) VBPRT (preterite) VBPRS (present) VBSUP (supine) VBIMP (imperative) VBINF (infinitive) VBSFO (s-form) NNIND (indefinite) NNDEF (definite) NNSIN (singular) NNPLU (plural) 87 100 1 99 19 78 9 100 65 100 4 0 53 34
0.19 0.22 0.29 0.53 0.10 0.08 0.44 0.32 0.47 0.56 0.58 0.58 0.58 0.60
A straightforward characterization of the “behaviour” of each feature was made by means of recording the fraction of works in which the relative frequency of the feature is higher in QI than in NF (i.e. above the diagonals in Figure 2). This score summarizes how authors tend to “use” that feature to create a contrast between NF and QI material. Another way to examine how the NF and QI styles are related is to apply a correlation test. Not wishing to assume that these variables are normally distributed, we used Spearman’s correlation coeficient ( ), which is rank-based and consequently non-parametric (i.e. not based on a normality assumption). Noun and adverb densities, as shown in Figure 2, will yield quite high values.
Table 2, whose values are plotted in Figure 3, shows the results. We find, as it were, a left column of NF-oriented features and a right one of QI-oriented ones. In eight cases the tendencies hold for all novels (0% or 100%) in the corpus. The NF-oriented features comprise elements of general and definite nominal reference, prepositions, passive and past tense verb forms, and numerals. The QI material, by contrast, is generally denser in interjections, first and second person pronouns, and present tense and imperative verb forms. Nouns are indefinite to a higher degree than in the NF. A plausible explanation is that there is less of anaphoric reference to already introduced entities. This is as can be expected from dialogue. The higher incidences of auxiliary verbs, adverbs, including negation, and subordinating conjunctions does not as immediately chime with what can be expected.
We see that there is a positive correlation between QI and NF values for all features, but the tendency is of a stronger kind ( > 0.55 ) as regards nouns, cardinal numerals, adverbs, and the inflectional categories of nouns. The lowest degree of correlation is seen in features related to deictic reference, i.e. tense and grammatical person, whose use in narration or dialogue is based on pragmatic principles.
We have introduced a method for visualizing and quantifying how authors use various
grammatical features to diferentiate between NF and QI material in fiction. Our observations clearly
agree with the findings of Egbert and Mahlberg [
We have found that NF and QI segments can be told apart with a high degree of accuracy by means of a multilayer perceptron. It should be stressed that there is a potential methodological danger in using a classifier that to a large extent rely on the same kind of features that are explored in the study. As always, patterns revealed by corpus-based “distant reading” should be seen as an invitation to look closer at the data.
Another open and interesting issue is to what extent the choice of QI orthography influences other stylistic options. The absence of quotation marks in the dash orthography is likely to prompt authors to use other means of making the QI status of segments obvious to the readers.
The present study has, in a sense, provided a map to the comfort zones of Swedish popular ifction and to the vast and fascinating regions of stylistic space this genre tends to shun.
This work was carried out in the project “Patterns of Popularity: Towards a Holistic Understanding of Contemporary Bestselling Fiction” funded by Vetenskapsrådet (2019-02829). PI Karl Berglund contributed to the design and data curation for the present study.