In this contribution we explore ways of detecting con昀氀ict representation in literary texts. First, we operationalize Glasl's concept of social con昀氀ict for manual annotation and second, we adapt a word embedding-based sentiment analysisS(entiArt) for the attribution of con昀氀ict values based on two scalar con昀氀ict operationalizations. By translating the values of the latter approaches into binary labels, we compare the embedding approaches with the manual annotation. Though correlation between the approaches is low, the paper demonstrates possible approaches to con昀氀ict analysis in literary texts and outlines directions for future research.
By addressing incompatibility as part of a notion of social con昀氀ict, we model incompatibility at sentence level. In contrast, the word embedding approaches (csfe.ction 3) are supposed to tackle con昀氀ict more generally. For these, we 昀椀rst use and then adapt word embedding modelbased sentiment analysis for the attribution of con昀氀ict values to our corpus. For the adaptation to con昀氀ict we use the poles of con昀氀ict and harmony and, in a second operationalization, the poles of con昀氀ict and resolution for assigning con昀氀ict values at the word level. Even though both the manual and the word embedding approaches to con昀氀ict are focused on the—more or less explicit—presence of con昀氀ict in the text, they di昀er with regard to the way they operationalize con昀氀ict. We therefore compare the approaches with regard to similarities and di昀erences, also including the sentiment analysis (cfs.ection 4). Finally we discuss our 昀椀ndings and the next steps for developing a more comprehensive approach to con昀氀ict in literary narratives (cf. section 5).
Our 昀椀rst exploration is based on a operationalization of social con昀氀ict. We use the notion by Friedrich Glasl9[] who synthesizes various previous de昀椀nitions by de昀椀ning social con昀氀ict as an interaction between two agents (‘Aktoren’, i.e. individuals, groups, organisations, etc.), where at least one agent experiences an incompatibility in perceiving, thinking/imagining, feeling, and wanting with the other agent. Furthermore, this incompatibility prevents the realization of what the agent thinks, feels, or wants3.
In terms of de昀椀nition this means that compatibility and impairment are the two necessary conditions in Glasl’s notion. Together they form the su昀케cient condition for the presence of a social con昀氀ict. For the exploration presented in this contribution we focus on the 昀椀rst condition: incompatibility. We assume that the di昀erent kinds of experiencing listed by Glasl (perceiving, thinking/imagining, feeling, and wanting) as well as the experienced incompatibility are explicated in the literary text. Moreover, we assume that these can be detected on sentence level. Since literary texts tend to be less explicit than real live narratives, we de昀椀ne two kinds of incompatibility that can be realized in a text: Firstly, as the narration of an experience of a character as contrary to someone else’s experience and secondly, as the representation of an experience as something negative. For both cases we di昀erentiate between actual incompatibility and hypothetical incompatible experience (e.g., when a character states that they would experience something incompatible if something speci昀椀c happened).
With this, we have a 昀椀rst operationalization of one core aspect of Glasl’s notion of social con昀氀ict. In further studies, it will be connected to the second condition, the resulting impairment. 3For a formalization of Glasl’s notion of social con昀氀ict for literary studies based on category theor5y].cf. [
The manual annotation process of represented incompatibility is based on senten4ceFso.r each sentence, annotators are also given some context that can be used for resolving ambiguities or uncertainties in the sentence in question. The following two conditions have to be true for a sentence to be annotated as incompatibility:
A character is narrated as experiencing something in perceiving, thinking/imagining, feeling, or wanting (by self-statement or narrator’s description)
The experiencing of this something is depicted as somehow contrary to how another character experiences this somethingOR The experiencing of this something is (represented as) negative
• Is the experience tied to something that is a fact →( actual incompatibility) or that is portrayed as possible→( hypothetical incompatibility)?
Here are some examples for incompatibilities. An actual incompatibility based on contrary experience can be found in this sentence: “Because you don’t know what it is to love a man, - to have to love a man, - how that burns!” (Felix Dahn (1886)F:redigundis, our translation). Here, a character states that they know how to feel something (i.e, pain caused by loving a man) that the addressed character does not know how to feel. The following is an actual incompatibility based on an experience represented as something negative: “With the awakening also my malice awoke, which increased still violently, when I heard from over there the mocking words and the laughter of my opponent, who on the other side, a little more mildly than me, might have come to earth.” (Johann Wolfgang Goethe (1811):Aus meinem Leben. Dichtung und Wahrheit, our translation). In this sentence, the 昀椀rst person narrator perceives themself as being malicious, which is a negative feeling.
And 昀椀nally an example for hypothetical incompatibility: “»No, my father is and will remain your and your husband’s faithful friend.« »Well,« said Mrs. Ebermann, »I would feel a cessation of his interest in us painfully, but a reproach would not arise from it in my heart.«” (Ida BoyEd: (1892) Empor, our translation). The second speaker (Mrs. Ebermann) states that she would perceive it painfully (i.e., negatively), if the other character’s father would lose interest in them.
In contrast to the cases exempli昀椀ed by these sentences and in line with Glasl’s notion, descriptions of (static) character traits, combat or other types of physical action caused by (presumed) anger as well as external proscriptions without any subjective perception (e.g., laws, bans, etc.) are not considered incompatibilities. 4This is a brief overview of our annotation guideline. The full guidelines can be founAdpipnendix A.
Based on these principles, we developed a guideline (cAf. ppendix A). Two annotators annotated 1,000 random sentences from our corpo5ra.Both annotators are trained in literary studies, annotator 1 is a student of literary studies and annotator 2 the 昀椀rst author. The annotated categories are depicted inTable 1.
While the inter-annotator agreement of 0.325 Cohenk’sappa is comparably low, the 昀椀ndings in the creation of the gold standard point to possible clari昀椀cations for future annotations. The annotators agree on 50 out of the 165 sentences annotated by annotator 1 (including six cases with disagreement about actual and hypothetical incompatibility). Furthermore, 30 sentences that are only annotated by annotator 1 as well as 27 sentences only annotated by annotator 2 were added to the gold standard. These are cases that seem to have been overlooked by the respective annotator. The remaining 85 annotations of annotator 1 are not included (11 out of these are errors, 44 contain no personal experience and 30 contain some sort of implied incompatibility but are not explicit enough to 昀椀t the current guidelines).
For our second approach to detecting con昀氀ict, we adapt the word embedding-based sentiment
analysis of theSentiArt-toolbox [
According to the distributional hypothesis4][, word embedding models encode semantic similarity based on word use. In sentiment analysis, the advantage of word embeddings over lexicon-based sentiment analysis is that we are not limited to the speci昀椀c words with sentiment values in a sentiment lexicon. Instead, we are able to express semantic similarity in a numerical value which can be calculated for every word in a text. TSheentiArt method allows choosing sentiment label words and thus focusing on speci昀椀c words. The original approach uses emotion label words based on Ekman’s concept of basic emotions3][to calculate valence and arousal values. However, that the algorithmic selection of emotion labels from other sources can be a viable alternative for sentiment analysi1s][.
In our approach, for evaluating a possible relevance of sentiment for con昀氀ict analysis, we
椀昀rst implement the SentiArt-approach for sentiment. For the sentiment labels, we translated
the label words from an emotion model based on James Russel’s ‘Circumplex Model of A昀ect’
[
The second word embedding approach focuses on con昀氀ict by using the the poles of con昀氀ict and harmony. Τhe label words for these are selected from the Dornsei昀-dictionar2y]. [The Dornsei昀-dictionary covers a wide range of topics and organizes subjects into speci昀椀c subgroups, providing a list of related terms for each. Connections between these subgroups are also listed and terms can pertain to multiple subgroups. We use words in the thematic groups of ‘con昀氀ict’ to encode the upper end of the spectrum and words in ‘harmony’-related groups for the lower end of the spectrum.
The third word embedding approach also focuses on con昀氀ict. This time, the scale ranges from con昀氀ict to con昀氀ict resolution and label word selection is based on a prior annotation study of con昀氀ict in German-language short stories. This task aimed at annotating various markers of social con昀氀ict potential, such as indicators of emotion-, action- or condition-based con昀氀ict as well as indicators for con昀氀ict resolution. Based on the average ratings from all annotators, we assessed for each annotated word how well it represented either the tags for con昀氀ict or the speci昀椀c tag for con昀氀ict resolution. The most representative words were selected as label words for con昀氀ict and con昀氀ict resolution.
For our analysis we 昀椀rst trained twoWord2Vec-models [
With the sentiment and con昀氀ict values, analyses of the most/least con昀氀ictual novel or trends in a corpus become possible. While the outcomes need further inspection, we can see that, for example, there is an increase in the con昀氀ict values of both approaches during the Romantic 7For a previous iteration of this method cf.10[]. There, we calculated values for verb phrases. For evaluation reasons we now use whole sentences, as verb phrases where o昀琀en too short for allowing annotation decisions. period (cf.Figure 2).8 This aligns with the assumption in literary studies that con昀氀icts increase in the Romantic period (cf., for example,1[4, p. 53]).
(a) (b)
Furthermore, examining speci昀椀c novels can, among other, provide insights on the development of con昀氀ictuality, coincidences of con昀氀ict value with key con昀氀icts in the plot or correlations between sentiment and con昀氀ict. For example, in Schlegel’sLucinde (cf. Figure 2), we see that lower valence, i.e. negativity, correlates with con昀氀ict. The Pearson correlation between con昀氀ict value and valence amounts to p=-0.85, orp=0.85, if the valence values are reversed as in Figure 2. The correlation between con昀氀ict values and arousal, on the other hand, is only p=-0.42. 8Cf. Appendix C for additional outcomes.
Part of our research interest is exploring whether the manual annotation of incompatibility and the word embedding approaches relate to similar phenomena. While the manual annotation approach operationalizes con昀氀ict as incompatibility in social interaction, the embedding approaches implement con昀氀ict as scale from con昀氀ict to harmony respectively con昀氀ict resolution. Nevertheless, since all approaches address con昀氀ict, it is interesting to see if they are connected in some way. Therefore, we explore some possible comparisons in this section.
For now, we focus on the con昀氀ict vs. harmony-approach as the label words represent two distinct poles and examine the sample of 1,000 sentences, where incompatibility was annotated manually. Since we use binary labels (con昀氀ict/no con昀氀ict) for manual annotation and scalar values for the word embedding approach we transform the latter into binary labels by de昀椀ning threshold values.
Since a sentence is similar to the concept of ‘con昀氀ict’ or ‘harmony’ if it has a positive or negative value, respectively, we set the threshold to zero. This results in 577 sentences with con昀氀ict (i.e., a positive value) and 423 sentences without con昀氀ict (i.e, a value below zero, cf. Table 3). If we compare these to the 107 sentences that were manually annotated as containing actual or hypothetical con昀氀ictual incompatibility and 893 sentences that contained none we get a precision value of 0.11, a recall of 0.61, and an F1-score of 0.19.
For a second comparison, we evaluate whether comparably high con昀氀ict values correlate with the manual incompatibility annotation. Therefore, we set the threshold at the upper quartile of the con昀氀ict values of the entire corpus, assuming that sentences with high con昀氀ict values also contain represented incompatibility. This results in 255 sentences above the upper quartile (value of 0.0089) and 745 sentences below the threshold, resulting in a precision of 0.11, a recall of 0.25 and an F1-score of 0.15 (cfT.able 4).9 9The comparison with the other word embedding approach with the poles con昀氀ict and con昀氀ict resolution demonstrated a similar overlap (F1-score of 0.19 if the threshold is set at zero as well and 0.11 with the upper quartile as
manual incompatibility annotation
Analogously to the comparison of con昀氀ict vs. harmony-values, we now compare valence and arousal values with the incompatibility annotations. In order to select the high and negative values respectively, we use all sentences with valence values lower than the lower quartile and arousal values higher than the upper quartile. This yields low overlap with a recall of 0.26 for low valence and 0.27 for high arousal (cTf.able 5and Table 6).
In comparing the automatic con昀氀ict annotation labels with the sentiment labels we make the same observation as insection 3 of high con昀氀ict correlating rather with low valence than high arousal. The comparison of the more con昀氀ictual sentences (threshold=Q3) and the sentences with lower valence (threshold=Q1) yields an F1 of 0.43 and with higher arousal (threshold=Q3) of 0.25 (cf. Table 7and Table 8).
We have presented two possibilities of approaching con昀氀ict in literary texts. The 昀椀rst possibility focuses on incompatibility represented by the text. This is a necessary condition for social con昀氀ict and can thus be considered an indicator for con昀氀ict. We implemented the approach as a manual annotation procedure. The second possibility was implemented by using a word embedding-approach with con昀氀ict-speci昀椀c label words. We adapted the approach ofSentiArt for a sentiment analysis as well as for two con昀氀ict analyses. For the latter, con昀氀ict label words were chosen from a comprehensive lexicon and from annotated data from another study.
Regarding manual annotation, our agreement evaluation and error analyses show that the operationalization of incompatibility needs re昀椀nement for tackling the currently unclear cases. These cases range from descriptions of circumstances that could be interpreted as personal experience to cases where a di昀erence in level of knowledge could be interpreted as a marker of con昀氀ictuality. Further elaborations of the annotation task could de昀椀ne more clearly which explicit and, if any, implicit expressions of incompatibility shall be annotated. A subsequent important step will be to take more elements from Glasl’s de昀椀nition into account, especially by de昀椀ning impairment. With better agreement scores, the annotated data set can be used for machine learning.
For the word embedding approaches further evaluations would be interesting. Generally, our comparisons can also be seen as an attempt for better understanding word embedding approaches and improving their validity. For now, the presented comparison to the manual annotation yielded rather low F1-scores. This may indicate that the phenomena measured by the di昀erent approaches do not really correlate. Furthermore, for the con昀氀ict vs. harmonyapproach di昀erent threshold-settings may make a di昀erence: While the precision between zero or the upper quartile was nearly the same, setting the threshold at zero resulted in higher recall. These outcomes may be due to the di昀erent con昀氀ict notions connected to the approaches but other reasons are also possible. We therefore consider a further exploration of possible connections between the approaches fruitful. This includes a more in-depth examination of the heuristic potential (e.g. by looking at what happens in larger passages with high con昀氀ict values, cf. the con昀氀ict arc in Figure 2) as well as developing additional strategies for transferring the scalar values to binary values and checking for resulting di昀erences in correlations.
We would like to thank Michael Weiland for annotating our data set and the anonymous reviewers for valuable feedback.
The following guidelines implement an aspect of Glasl’s de昀椀nition of social con昀氀ict, namely incompatibility. Goal is to identify situations of incompatibility depicted in a narration that are/could likely be a prerequisite for social con昀氀ict. The goal of the annotation is not to annotate Glasl’s notion of social con昀氀ict in its entirety.
The annotation span are sentences. Sentence splitting was done automatically. The following guidelines have to be applicable to at least one phenomenon in the sentence for the sentence to be tagged accordingly. One sentence before and a昀琀er are given only as auxiliary context, annotated are no phenomena expressed over multiple sentences.
Note: Annotation span is the automatically split sentence. Errors by the sentence splitter do not have to be corrected, e.g. if a sentence contains direct speech by multiple characters (recognizable by guillemets (»«)) the annotation guidelines can be applied to the entire sentence (cf. examples from Dahn’sFredigundis below).
A sentence contains con昀氀ictual incompatibility if both of the following twnoecessary conditions are true: 1. Is someone represented as experiencing something in perceiving, thinking/imagining, feeling, or wanting (by self-statement or narrator’s description)? 2. Is the experiencing of this something represented as somehow contrary to how someone else experiences this something? OR
Is the experiencing of this (represented as) something represented as negative? Furthermore, the following question constitutes the type of incompatibility: 3. Is the experience tied to something that is a fact or that is portrayed as possible? → actual or hypothetical incompatibility10
Please note: There are only rare cases where a text addresses incompatibility in one of the versions mentioned above in an explicit manner. Instead, in most cases the decision about the presence of a compatibility involves some interpretation of the text. Nevertheless, this interpretation should be strongly related to the text.
Examples for actual incompatibility: The experience regards something that is actually happening or has actually happened. • “Mit dem Erwachen erwachte auch meine Bosheit, die sich noch he昀琀ig vermehrte, als ich von drüben die Spottworte und das Gelächter meiner Gegnerin vernahm, die an der andern Seite, etwas gelinder als ich, mochte zur Erde gekommen sein.” (Johann Wolfgang Goethe (1811): Aus meinem Leben. Dichtung und Wahrheit) → The 昀椀rst person narrator perceives to be mean, which is typically a negative feeling, in the following sentence the character also attacks somebody. • “Gleichgültig brach sie mit mir – und das war der Bruch meines Lebensp!r”io(r sentence), “Ich mochte nichts mehr von ihr hören;”ta(rget sentence), “ich sandte ihr das Letzte, was ich von ihr besaß, die Börse, durch den Grafen Malariva zurückf.”oll(owing sentence) (Heinrich Zschokke (1844):Die Rose von Disentis) → The 昀椀rst person narrator want’s not to be associated with someone anymore, breaking o昀 contact as a negatively connoted desire. • “»Weil du nicht weißt, was es ist, einen Mann lieben, – einen Mann lieben müssen, – wie das brennt!« »Nein!« lachte die andre, das Haar lustig schüttelnd.«” (Felix Dahn (1886): Fredigundis) → The speaker states that they know how to feel something (pain caused by loving a man) that the addressed person does not know how to feel (and even con昀椀rms it). 10If a sentence contains multiple phenomena (actual and hypothetical incompatibility), the dominant phenomenon should be assigned as tag. Examples for hypothetical incompatibility The experience regards something that is only possible.
• “Nein, mein Vater ist und bleibt dein und deines Mannes treuer Freund.« »Nun,« sagte Frau Ebermann, »ich würde ein Au昀栀ören seines Interesses an uns schmerzlich emp昀椀nden, aber ein Vorwurf würde ihm in meinem Herzen nicht daraus erwachsen.” (Ida Boy-Ed (1892): Empor) → The second speaker (Frau Ebermann) states, that she would perceive it painfully (i.e., negatively), if the other character’s father would lose interest in them.
• What does not count as (relevant/propere)xperience? – Merestatements without any hint towards anexperience and an experiencing subject.
∗ “Sich absondern von der Menge, verräth wenig Muth.” (Karl Bleibbtreu (1888):
Größenwahn) → There is no information about the origin of this statement. – Descriptions of (static)character traits → e.g. looks, state of mind, even if stereotypically negative ∗ “Er ist mein Lieblingsdichter, und ich kann ihn auswendig, so wenig ich mir sonst, trotz gelegentlich eigener Versündigungen, aus der Dichterei mache.” (Theodor Fontane (1895):E昀케 Briest ) → The character does not like poetry, not liking poetry can be seen here as static character trait. – Combat orphysical actions caused by (presumed) anger ∗ “Der Fürst wandte sich ab.” (Joseph von Eichendor昀 (1833): Dichter und ihre Gesellen) → Could only be result of a con昀氀ictual situtation, no explicit expression of the character experiencing something. – Mere facial expressions.
∗ “»Was kommt Euch denn zu von den Marken?« sprach der Graf von Anhalt, der gar 昀椀nster vor sich blickte.” (Willibald Alexis (1842D):er falsche Woldemar) → The character’s (Graf von Anhalt) facial expression does not explicitly match a mental experience and is rather a con昀氀ict indicator, than an expression of proper incompatibility. • What does not count as (relevant)cause for incompatibility? – External proscriptions (laws, bans, etc.), where the incompatibility does not stem from a subjective perception ∗ “Mir ist es bisher versagt geblieben, irgend welchen Ein昀氀uß auf das junge Mädchen auszuüben.” (Wilhelm von Polenz (1893)D:er Pfarrer von Breitendorf ) → The character 昀椀rst does not express that they want to in昀氀uence this person and the incompatibility (of presumed want to in昀氀uence a person) is second not of internal nature (like e.g. the character stating that they are angry about not being allowed to in昀氀uence this other person).