Character-distinguishing features in fictional dialogue: the case of War and
Peace

Daniil Skorinkin

skorinkin.danil@gmail.com

National Research University Higher School of Economics, Russia

Introduction

The study of character speech is a topic of fairly consistent interest among
digital literary scholars. It is usually acknowledged that voices of characters
are essentially different from narrator's own voice and should be treated
separately. Some researchers have fictional dialogue removed from the texts
they studied before any tools of computational investigation are applied
(Hoover, 2004). Quite a lot of effort has been made recently to address the
problem of identifying character speech in prose and attributing it to the
correct speaker (ccbMku!). One of the outcomes of such

research is the possibility to study voices of different

characters on relatively large scale and apply computational tools that measure
their recurring stylistic parameters.

Method

The study of character speech has traditionally had strong ties to the fields
of stylometry and authorship attribution, as their methods proved quite useful
for studying idiolect of a fictional speaker. Suffice it to say that one of the
seminal works in stylometry, Computation into criticism# by Burrows (Burrows,
1987), was focused on the study of character speech in Jane Austen's novels.
The method developed by Burrows grew into what is currently known as Delta, a
widely-adopted standard for authorship attribution. Delta has been consistently
and successfully applied to identifying the author of an unattributed text of
different languages and genres, but at the same time it saw considerable usage
as a purely stylometric tool for the study of text where authorship is
undisputed. Among other things, this included research into the specific
idiolects

of fictional characters (see, for example, Rybicki,

2006).

In our research Delta was used as one of the two possible approaches to
studying character voices in Leo Tolstoy's War and peace. Much like in case
of Senkewic (Rybicki, 2006), there's certain critical opinion (Eikhenbaum,
2009) that Tolstoy's characters are quite distinct from each other in their
speech. Our own experience of carefully reading speech instances extracted from
War and peace (for details on extraction procedure see (Skorinkin,
Bonch-Osmolovskaya, 2015) supports the opinion. So it seemed natural to try and
test computational methods that already showed their applicability to precisely
such task. We used R package stylo by (Eder et al, 2013)

Testing the method on Russian material

Surprisingly enough, we were unable to find any work that applied Delta to any
Russian material. Therefore we felt obliged to conduct a couple of experiments
that would test its general applicability to Russian before we proceed with
character speech. At the first stage we tried Delta's ability to distinguish
between Tolstoy and Dostoevsky. The training set contained one of the six parts
of Dostoyevsky's Crime and Punishment and three of the fifteen books of
Tolstoy's War and Peace. The remaining 18 pieces of text (5 by Dostoevsky and
13 by Tolstoy) constituted the test set. The results with different settings
can be seen in Table 1 and Figures 1,2:

┌───────────────┬──────────┬─────────┬─────────┬──────────┐
│N most frequent│Words     │grams    │grams    │grams     │
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│25             │80% (4/5) │60% (3/5)│60% (3/5)│100% (5/5)│
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│30             │80% (4/5) │80% (4/5)│60% (3/5)│80% (4/5) │
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│35             │80% (4/5) │60% (3/5)│60% (3/5)│80% (4/5) │
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│40             │80% (4/5) │60% (3/5)│60% (3/5)│80% (4/5) │
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│45             │80% (4/5) │60% (3/5)│80% (4/5)│100% (5/5)│
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│50             │80% (4/5) │60% (3/5)│80% (4/5)│100% (5/5)│
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│55             │80% (4/5) │60% (3/5)│80% (4/5)│80% (4/5) │
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│BO             │100% (5/5)│60% (3/5)│80% (4/5)│80% (4/5) │
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│65             │100% (5/5)│60% (3/5)│80% (4/5)│80% (4/5) │
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│70             │80% (4/5) │60% (3/5)│80% (4/5)│100% (5/5)│
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│75             │80% (4/5) │80% (4/5)│80% (4/5)│80% (4/5) │
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│80             │80% (4/5) │60% (3/5)│80% (4/5)│100% (5/5)│
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│85             │80% (4/5) │80% (4/5)│80% (4/5)│100% (5/5)│
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│90             │100% (5/5)│80% (4/5)│80% (4/5)│100% (5/5)│
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│95             │80% (4/5) │80% (4/5)│80% (4/5)│100% (5/5)│
├───────────────┼──────────┼─────────┼─────────┼──────────┤
│100            │80% (4/5) │80% (4/5)│80% (4/5)│100% (5/5)│
└───────────────┴──────────┴─────────┴─────────┴──────────┘

Table 1. Delta authorship attribution, Tolstoy vs Dostoevsky

Tolstoy vs Dostoevsky Principal Components Analysis

┌────────────────────────────────────────────────────┬────────────────────────┐
│                                                    │Tolstoy_VoinalmirTom4Ch2│
│                                                    │                        │
│                                                    │Tolstoy_VoinalmlrTom4Ch3│
├────────────────────────────────────┬───────────────┼────────────────────────┤
│Dostoevsky PrestuplenielNakazanieCh2│TolstoyJ       │/„¡n,lni?<^t^:“^.^      │
│                                    │               │[irTom2ch2]             │
│DosToevsky_PrestuplenielNakazanieCh1├───────────────┼────────────────────────┤
│                                    │               │Tolstoy VoinalmirTom1Ch3│
├────────────────────────────────────┼───────────────┼────────────────────────┤
│                                    │               │Tolstoid[(]S|t5)P!      │
│                                    │               │VMiemWfn4Ch4            │
├────────────────────────────────────┼───────────────┼────────────────────────┤
│Dostoevsky_PrestuplenielNakazanieCh5│               │VoinalmirT^^^'^^        │
│                                    │               │OJvolnalmlrTomlChl      │
├────────────────────────────────────┼───────────────┼────────────────────────┤
│                                    │               │Tolstoy_VolnalmlrTom2Ch5│
├────────────────────────────────────┼───────────────┼────────────────────────┤
│                                    │PC1 (40.9%)    │                        │
│-0.4    -0.2                        │               │0 °-^2                  │
│                                    │0 MFC 4-grams  │                        │
│                                    │Culled @ 0%    │                        │
└────────────────────────────────────┴───────────────┴────────────────────────┘

Fig. 1. Delta PCA on 150 most frequent character 4-grams, Tolstoy vs Dostoevsky

[625-1]

Fig. 2. Delta PCA on 100 most frequent words, Tolstoy vs Dostoevsky

The second experiment involved four Russian authors Tolstoy, Dostoevsky,
Goncharov and Turgenev. All four represent (roughly) the same epoch of
Russian literature and all four are recognized as masters of realistic prose.
We used three novels by each author for our experiment. At the first stage two
out of each three were placed in the training corpus, and Delta was supposed to
attribute the remaining one. All four novels from the test corpus were
attributed correctly. At the second stage we reverted the experiment and left
only one novel by each author in the training set. In this case Delta
consistently showed 7 out of 8 correct attributions (the only mistake being
Tolstoy's Family Happiness incorrectly attributed to Dostoevsky.A possible
explanation could be that Family Hap-piness is written in first person from the
point of view of a young woman, something uncommon for Tolstoy; and the only
Dostoevsky's work the training corpus contained was The Insulted and Humiliated
, also a firstperson narrative). Fig. 3 shows Delta scores for all the

novels visualized with help of principal component

analysis.

XIX century novels Principal Components Analysis

┌────────────────────────────────────────────┬────────────────────────────┐
│roHMapo^fii^pa^gph.B                        │TypreHeB_flBopnHCKoe rHe3flo│
├────────────────────────────────────────────┼────────────────────────────┤
│Tyf                                         │)reHeB_OTpbi „ fle™         │
├────────────────────────────────────────────┼────────────────────────────┤
│rOHHapOB_06blKHOBeHHafl MCTOpriB            │TypreHeB_flbiM              │
├────────────────────────────────────────────┼────────────────────────────┤
│                                            │ToncToti_BoMHa h wnp        │
├────────────────────────────────────────────┼────────────────────────────┤
│flocToeBCKMM_npecTynneHne u HaKa3aHne       │ToncTOu_AHHa KaperinHa      │
├────────────────────────────────────────────┼────────────────────────────┤
│flocToeBCKkiii_yHki>i<eHHbie m ocuopSreHHbi'│3                           │
├────────────────────────────────────────────┼────────────────────────────┤
│fl0CT0eBCKnn_5eflHbie moflii                │ToncTori_CeMeiinoe cqacTbe  │
└────────────────────────────────────────────┴────────────────────────────┘

-15    -10    -5    0    5    10    15

PC1 (24.1%)

250 MFW Culled @ 0%

Correlation matrix

Fig. 3. Delta PCA for 12 Russian novels of 1850-1870-ies, 250 most frequent
words

Applying Delta to Tolstoy's characters

Having thus shown that Delta is applicable to Russian, we proceeded with our
experiment. In the first place we applied the method to top 5 characters by
the total number of speech instances. We split the total sets of speeches by
each character and then tried authorship attribution The results are shown in
Table 2.

┌───────────────┬──────────┬─────────┬──────────────┬──────────┐
│               │          │cn&mata- │              │uiameca-  │
│N most frequent│WocUSs    │         │Cnamei digrams│          │
│               │          │grams    │              │grams     │
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│25             │80% (4/5) │60% (3/5)│60% (3/5)     │100% (5/5)│
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│30             │80% (4/5, │80% (4/5)│60% (3/5)     │80% (4/5, │
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│35             │80% (4/5) │60% (3/5)│60% (3/5)     │80% (4/5) │
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│40             │80% (4/5) │60% (3/5)│60% (3/5)     │80% (4/5) │
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│45             │80% (4/5, │60% (3/5,│80% (4/5,     │100% (5/5,│
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│50             │80% (4/5) │60% (3/5)│80% (4/5)     │100% (5/5)│
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│55             │80% (4/5) │60% (3/5)│80% (4/5)     │80% (4/5) │
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│60             │100% (5/5)│60% (3/5,│80% (4/5)     │80% (4/5) │
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│65             │100% (5/5)│60% (3/5)│80% (4/5)     │80% (4/5) │
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│70             │80% (4/5) │60% (3/5)│80% (4/5)     │100% (5/5)│
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│75             │80% (4/5) │80% (4/5)│80% (4/5,     │8U% (4/5) │
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│80             │80% (4/5) │60% (3/5)│80% (4/5)     │100% (5/5)│
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│85             │80% (4/5) │80% (4/5)│80% (4/5)     │100% (5/5)│
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│90             │1UU% (5/5)│80% (4/5,│80% (4/5)     │100% (5/5)│
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│95             │80% (4/5) │80% (4/5)│80% (4/5)     │100% (5/5)│
├───────────────┼──────────┼─────────┼──────────────┼──────────┤
│100            │80% (4/5) │80% (4/5)│80% (4/5)     │100% (5/5)│
└───────────────┴──────────┴─────────┴──────────────┴──────────┘

Table 2.

The most common mistakes are between princess Marya Bolkonskaya and Natasha
Rostova and between prince Andrew Bolkonsky and Pierre Bezukhov.
Their closeness can be seen in Figure 4:

T op5CharactersDelta Principal Components Analysis

[625-2]

-5    0    5    10

PC1 (34.8%)

100 MFW Culled @0%

Correlation matrix

Fig. 4. Delta PCA for top 5 most talkative characters in War and Peace, 100
most frequent words


Topi SCharactersDelta Principal Components Analysis

[625-3]

-5    0    5    10

PC1 (19.3%)

100 MFW Culled @ 0%

Correlation matrix

Fig. 6. Delta-based PCA for top 15 most talkative characters in War and Peace,
100 most frequent words


TopSCharactersDelta Cluster Analysis


HaTaniaPocTOBa


MapbflEojiKOHCKaa


ribepEe3yxOB


AHflpeÜEOJlKOHCKMÜ


One can easily see the clustering of Rostov family, to a lesser extent this
applies to Bolkonsky family as well. Dolokhov, Denisov and Kutuzov could
constitute the 'war' cluster.

We then made another expansion and moved from 15 to 30 characters. Figure 7
demonstrates PCA of Delta scores for this selection.


HnnonafiPocTOB

1.5    1.0    0.5    0.0

100 MFW Culled @0%

Classic Della distance

Fig. 5. Delta-based hierarchical clustering for top 5 most talkative characters
in War and Peace, 100 most frequent words

The quality of speech authorship attribution inevitably got worse once we
expanded our selection from 5 to 15 characters. The results were still quite
tolerable reaching 10 out of 15 with certain settings. The analysis of mistakes
showed that a) they're less likely to occur between characters of different
gender and b) the mistaken characters have quite a lot of mutual conversations.

Further on, we decided to pay more attention to overall Delta scores of
character voices and see if they

give us any meaningful clustering of characters. Figure 6 shows PCA of
characters based on Delta.

[625-4]

Fig. 7. Delta-based PCA for top 30 most talkative characters in War and Peace,
100 most frequent words

The most striking thing here is the obvious separation of Vera Rostova from the
rest of Rostov family. The difference between cold, tempered and rational Vera
and her emotional and very sentimental relatives is outspoken and obvious to
the reader, but it seems valuable to have this potential quantiative support
in the form of different Delta scores. What is even more striking is that Vera
is quite close to Berg, a rationalizing careerist who becomes her husband. Note
also the

closeness of Boris and Julie Karagine, another pragmatic couple happily united
in a marriage of convenience.

Alternative Features Top 5 Cluster Analysis


HuKOJiaAPocTOB


Applying alternative features

Having tried Delta, we proceeded with our own set of alternative features for
character voice analysis (a typical step, as dectibed in Eder, 2015). These
features are not related to the lexical makeup of character speech and attempt
to reduce the influence of gender-related morphological features of Russian
language and the factor of mutual interactions between characters. At this
stage we limited ourselves to four features

only: the average number of words, the ratio of exclamatory sentences, the
ratio oa question sentences and the ratio of punctuation marks (latter being
a crude approximation of the ‘disruptedness' of speech, which seems rather
typical of certain more emotional and lively characters).

When the character set is limited to 5 characters these features even manage to
distinguish character speech with some tolerable accuracy (though worse than
Delta). However, the analysis of mistakes shows that they capture fundamentally
different types of similarities than Delta does. For instance, joyful
Natasha in this case is never mistaken for sentimental and melancholic Marya,
but rather for her boisterous brother Nikolai. Pierre, on the other hand, is
mistaken for Marya rather than for Andrey, who is distinct from them all.
Figures 8 and 9 show the results of PCA and hierarchical clustering for these
characters based on our own alternative features.

[625-5]

Fig. 8. PCA for top 5 most talkative characters in War and Peace, 4 alternative
features


HaTawaPocTOBa

MapbflBojiKOHCKaa

ribep6e3yxoB

AHflpeiiBonKOHCKMM

0.08    0.06    0.04    0.02    0 00

4 MFW Culled @ 0%

Cosine distance

Fig. 9. Hierarchical clustering for top 5 most talkative characters in War and
Peace, 4 alternative features

If we compare figures 8 and 9 to their counterparts from the Delta experiment
(figures 4 and 5) we can see that the alternative features ignore gender or
mutual interactions/ The hypothesis is that they enable us with a more indepth
view of a character personality, his/her emotional type and so on.

Figures 10-12 show data on wider selections of characters using alternative
features.

Alternative Features Top 15 Principal Components Analysis

┌─────────────────────────────────────┬──────────────┬────────────────┐
├─────────────────────────────┬───────┤tfguOHCKan    │rpact>HHHPOCTOBa│
│AHfl pew Bon KOHCKMM         │flonox │              │                │
├────┬────────────┬───────────┼───────┼──────────────┼────────────────┤
│    │            │           │       │              │HiiKOnanPoCTOB  │
├────┼────────────┼───────────┼───────┼──────────────┼────────────────┤
│    │            │           │KyTy3OB│              │                │
├────┼────────────┼───────────┼───────┼──────────────┼────────────────┤
│Bac,│>™nKypar[M]H│roHKapaTaeB│       │fleHMCOB      │                │
├────┼────────────┼───────────┼───────┼──────────────┼────────────────┤
│    │            │           │       │HwKoriaiiBonKO│MnuaPoCTOB      │
└────┴────────────┴───────────┴───────┴──────────────┴────────────────┘

Fig. 10. PCA for top 15 most talkative characters in War and Peace, alternative
features


[625-6]

[625-7]

Fig. 11. Hierarchical clustering for top 15 most talkative characters in War
and Peace, alternative features

Note that here we do not see any similarity between Andrey and Pierre.
Moreover, Andrey is close to Napoleon, which is rather striking given Napoleon
is

his hero and role model for a considerable part of the novel.

[625-8]

The separation of Vera, on the other hand, is still rather visible - she is far
from Moscow-centered Rostov world and close to Saint-Petersburg world of
Ku-ragine family and berg.