A central question for the analysis of literary texts in digital humanities is the identification of characters in the text. A simple approach would be to reduce the analysis to mentions of the character by its proper name as these are easy to retrieve. A more elaborate solution requires coreference annotation, identifying all mentions of a character in the text, irrespective of their form. Using the example of the novel Corpus Delicti by the German author Juli Zeh, we compare these two approaches and show the added value of coreference annotation.
One of the objectives of the analysis of literature in
digital humanities (DH) is the analysis of characters
Both types of analysis require the detection of character
mentions in the text. One simple solution would be to
reduce the analysis to mentions of the proper name of the
character. These can be identified easily and can be
considered an approximation for the actual presence of a
character. The more time-consuming way is a coreference
annotation of the text. Two expressions in a text are
considered coreferent if they refer to the same discourse entity
For German, there is the statistical coreference resolution
tool HotCorefDE
Our work is situated in the context of character network
analysis, made popular by e. g.
For German, several studies on character networks have
been conducted and some tools have been published. The
rCat-Tool1 by
The research project Kallimachos focuses on the analysis
of character networks in narrative texts, namely German
novels
Our object of analysis is Juli Zeh’s novel Corpus Delicti
that was published in 2009. The story is set in a dystopian
future where a repressive system is in power that is centered
around questions of human health. The novel is written in a
realistic style and narrated by a heterodiegetic narrator who
rarely imparts psychological information about the
characters. Many chapters include large dialog sections (about
half of the novel is direct speech) and every chapter
represents a fairly self-contained story episode, which is why we
base our analyses on a segmentation per chapter.
We conducted the manual annotation using the tool
CorefAnnotator3 by Nils Reiter that was developed specifically
for the task of manual coreference annotation. For the
annotation we used the guidelines for coreference annotation
described in
In addition, the text was annotated automatically for
partof-speech with MarMot
From this annotated version of the novel, we extracted the following information for each character mention: 2http://kallimachos.de/kallimachos/index. php/Tutorial_Figurennetzwerke, May 2, 2018. 3https://doi.org/10.5281/zenodo.1228105 the token span, the entity it refers to, the linguistic form (proper name, pronoun...), whether it occurs inside direct speech (detected by quotes) and the chapter in which it occurs.
For copyright reasons, we are unable to publish the annotated text. However, you can find the extracted list of mentions and their annotations at https://doi.org/10. 5281/zenodo.1239701.
4.
4.1.
We will first inspect the forms used to refer to the characters in the novel. Figure 1 displays the distribution of linguistic forms used to refer to the four characters mentioned most frequently in the novel (the white line indicates the absolute numbers, scale on the right). The distribution of forms is very similar for the four characters. Proper names (NE) amount to about one quarter of all mentions, while personal pronouns (PPER) account for about half of all mentions. The rest are mostly possessive pronouns (PPOSAT) and noun phrases (NP). We conclude that while the relation between proper names and other mentions is relatively stable across characters in our data, proper names account for only a minor proportion of all character mentions, making a coreference annotation desirable.
In order to see if the relation between proper names and other mentions is also stable across the text, we display the distribution of expressions referring to the main character of Corpus Delicti, Mia, across chapters in Figure 2. In the horizontal dimension, we can see the 49 chapters of the novel. For each of the chapters the bars display which forms are used for reference to Mia, relative to all references to Mia in the chapter (scale on the left). The white line graph indicates the absolute number of mentions of Mia in each chapter (scale on the right). The distribution of forms for the character Mia shows a considerable amount of variability. Note that the main outliers occur in chapters with very few mentions of the character. We can posit the hypothesis that the relation between different forms of referring expressions is increasingly stable when including longer stretches of text. It can also be noted (having read the novel) that the proportion of proper names is high in chapters with other female characters which make pronouns ambiguous. This applies to, for example, chapters 3 and 28. 4.2.
For studies on the presence of characters, e. g. with the
purpose of creating a character network, we need
information on when characters appear in the novel and which
other characters they appear with. We want to describe how
this information is affected by preprocessing such as
coreference annotation. In addition to this main question, we
also explore the effect of excluding direct speech
In Figure 4 you can see a comparison of three conditions, ranging from no preprocessing (proper names only, Figure 4a) via one preprocessing step (with coreference annotation, Figure 4b) to two preprocessing steps (coreference annotation and exclusion of direct speech, Figure 4c). The bars indicate the relative proportion of character mentions for each of the four main characters, relative to all mentions of all four characters in the chapter.
One expected, but still important change between the three conditions is the number of mentions (white line, scale on the right). When including coreference annotation, the number of mentions increases greatly. This is highly beneficial for analysis of, for example, syntactic contexts the characters occur in. Naturally, the number of mentions decreases again when excluding direct speech.
When comparing the first two graphs, we can see some changes in the proportions of the four characters. For instance, the relation between the characters Mia and Moritz is almost inverted in chapter 27. A closer look at the chapter reveals the reason: A large part of the chapter is a conversation between Mia and Moritz about something that happened to Moritz. In the conversation, references to Moritz are mainly realized by first and second person pronouns. Another example is chapter 3: As you can see, the proportion of the mentions of Kramer rises relative to the mentions of Mia when coreference annotation is included. The reason is that Kramer is having a conversation in the chapter, so he is often addressed by pronominal expressions in first and second person. Mia, on the other hand, is not communicatively interacting with the other characters but the conversation’s topic. For that reason, Mia nearly ’disappears‘ when direct speech is excluded (see c).
We can derive the hypothesis that references to speaker and addressee in character conversation are rarely realized as proper names and will often be underrepresented if only proper names are considered. In addition to that, the presence of an absent character like Mia would be overestimated.
The distribution changes even more when all mentions in direct speech are excluded (Figure 4c). Generally speaking, the number of characters per chapter is reduced. Most strikingly, the proportions of the character Moritz are considerably smaller. This can be explained by his position in the novel: The character Moritz died before the main narration and is – except for some flashbacks – not present but only talked about.
It is a matter of the individual focus of analysis whether the exclusion of direct speech is appropriate or not, e. g. the fact that Mia talks a lot about her dead brother Moritz does also tell us a great deal about the character constellation. However, we have to bear in mind that this decision can heavily influence the analysis.
In addition to the visual comparison of the graphs, Figure 3 provides Pearson’s correlations for the absolute frequency of mentions of the four characters under the three conditions (in the same order as in Figure 4). A score of 1 would mean that there is a linear relationship between the frequencies: If the frequency of a character measured in proper names doubles from one chapter to another, the corresponding frequency measured in total mentions including pronouns etc. doubles as well. The further away from 1 the score is, the more independent the frequencies are. If the correlation between two conditions is strong, we do not get much additional information when considering both conditions.
We can see that the characters are affected by the changes to different degrees. For Kramer and Rosentreter all the scores are above 0.93, indicating some, but no substantial changes. The scores for Mia and Moritz are much lower (between 0.78 and 0.90). The comparison of the first two conditions reflects the variability we have already seen for Mia in Figure 2: The relation of proper names to other types of mentions is not constant across chapters.
Also the correlations between condition 2 (coreference) and condition 3 (coreference, direct speech excluded) are lower for Mia and Moritz. This can be explained by the fact that they are more frequently mentioned in direct speech than Kramer and Rosentreter.
The correlation between condition 1 (proper names only) and condition 3 (coreference and direct speech excluded) is especially low for Mia and Moritz. This indicates that the two preprocessing steps—coreference resolution and exclusion of direct speech—accumulate for these characters. 4.3.
In addition to increasing the accuracy of analyses based on character mentions, coreference annotation can also give us a first impression of a character’s attributes. For this purpose, we inspect the noun phrases used to refer to a specific character. Note that we use the term ’noun phrase‘ in a narrow sense, referring to appellative noun phrases as opposed to proper names and pronouns. In Table 1 you can see the most frequent noun phrases used for the character Mia, reduced to the head of the phrase. We can see immediately that the story of Mia is centered around court proceedings in which Mia is the defendant and – in the end – the convicted. Additionally, we can see that her family status as a sister is mentioned repeatedly. This is mirrored in the data for her brother Moritz: 43 of 47 noun phrases referring to him have the head Bruder (’brother‘). Here we can see clearly that while the main character Mia is characterized by many different noun phrases, Moritz’s role is limited to his family relation to Mia.
Angeklagte Schwester Beschuldigte Verurteilte Mandantin
defendant sister accused convicted client
Other possibilities for the automatic description of a
character include word clouds based on the context of all character
mentions as in
5.
We have shown that the ratio of proper names to other mentions is about 1:3. While it is surprisingly stable between characters, it varies between chapters of the novel Corpus Delicti. We could show that especially speakers in conversation are highly underrepresented when considering proper names only. For this reason, the analysis of copresence of characters will yield different results when based on proper names only or on all mentions. We therefore argue that, while the analysis of proper names requires much less time, literary character analysis benefits from coreference annotation. In addition, it enhances the possibilities of describing a character by the noun phrases referring to it and its syntactic context.
In the future, we will use the data described here to create character networks and further investigate the influence of preprocessing like coreference annotation for this type of analysis. We hope that this type of analysis will contribute to the identification of genre features, our focus being the dystopia. To allow for genre specific findings, we will annotate and analyze coreference in another dystopia as well as two historic novels.
6.
This work has been funded by the ‘Landesforschungsfo¨rderung Hamburg’ in the context of the hermA project (LFF-FV 35). We thank Lea Ro¨seler and Daniel Fabian Klein for their help with the annotation and Piklu Gupta for checking our English. All remaining errors are our own. (a) based on proper names only (b) with coreference annotation (c) with coreference annotation, direct speech excluded
Figure 4: Mentions of main characters by chapters under different conditions 5