=Paper=
{{Paper
|id=Vol-2769/2
|storemode=property
|title=Analyses of Character Emotions in Dramatic Works by Using EmoLex Unigrams
|pdfUrl=https://ceur-ws.org/Vol-2769/paper_02.pdf
|volume=Vol-2769
|authors=Mehmet Can Yavuz
|dblpUrl=https://dblp.org/rec/conf/clic-it/Yavuz20
}}
==Analyses of Character Emotions in Dramatic Works by Using EmoLex Unigrams==
https://ceur-ws.org/Vol-2769/paper_02.pdf
Analyses of Character Emotions in Dramatic Works by Using EmoLex
Unigrams
Mehmet Can Yavuz
Faculty of Engineering and Natural Science, Sabancı University, Tuzla
İstanbul, Türkiye
mehmetyavuz@sabanciuniv.edu
Abstract medium for establishing antagonisms. Therefore,
the characters in these plays express themselves
In theatrical pieces, written language is through dialogues or monologues clearly. By us-
the primary medium for establishing an- ing the emotional association of each word, from
tagonisms. As one of the most impor- this perspective, it is possible to reveal the emo-
tant figures of renaissance, Shakespeare tional landscape of the plays. Emotionally posi-
wrote characters which express them- tioning the characters relative to each other is im-
selves clearly. Thus, the emotional land- portant for further understanding of the structure
scape of the plays can be revealed from of the plays. It is also important to extract overall
the texts. It is important to analyze such emotional variations throughout the play to have
landscapes for further demonstrating these an insight about the tragedies. In this study, it
structures. is aimed to answer these two questions by using
We use word-emotion association lexicon word-emotion association lexicon with eight ba-
with eight basic emotions (anger, fear, an- sic emotions (anger, fear, anticipation, trust, sur-
ticipation, trust, surprise, sadness, joy, and prise, sadness, joy, and disgust) and two senti-
disgust) and two sentiments (negative and ments (negative and positive).
positive). By using this lexicon, the emo- These tasks are not only important for artifi-
tional state of each character is represented cial literature (Lebrun, 2017; Yavuz, 2020), but
in 10 dimensional space and mapped onto also for the purpose of increasing artistic cre-
a plane. This principle axes planes posi- ativity through computerized analysis. The al-
tion each character relatively. Addition- gorithmic study of literary works has given rise
ally, tempora-emotional evaluation of each to objective criticism in literary theory, (Moretti,
play is graphed. 2013). Accordingly, the discovery of new or
We conclude that the protagonist and the previously theoretically laid out features of lit-
antagonist have different emotional states erary texts through mathematical experiments or
from the rest and these two emotionally statistics brought along conjunctions on literature,
oppose each other. Temporal-Emotional (Moretti, 2000; Yavuz, 2020). Art and criticism
timeline of the plays are meaningful to develop in parallel today as of yesterday. Com-
have a better insight into the tragedies. putational linguistics enriched with fields such as
advanced chatbots, conversational AI, or text style
1 Introduction transfers, the fields give clues that artificial liter-
Shakespeare’s plays are one of the most important ature will also develop rapidly in the upcoming
works of early modernity with their dramaturgy, period. Whether it is a computer-assisted writ-
strong and in-depth characters and poems, that are ing process or fully automated writing machines,
all still contemporary. Antagonistically most pow- these generative models need evaluation metrics.
erful works of Shakespeare, who wrote in three For this purpose, objective metrics should be de-
genres, are tragedies. Tragedies have strong an- veloped for outstanding literary pieces such as
tagonisms and written language is the primary Shakespeare, and for what makes these plays so
great, if possible. . . Aim of this paper is in ac-
Copyright c 2020 for this paper by its authors. Use
permitted under Creative Commons License Attribution 4.0 cordance with the previous works, (Yavuz, 2019;
International (CC BY 4.0). Yavuz, 2020), we would like to discuss objective
and quantifiable reasons behind the literariness of tions along the texts.
a drama.
2.1 EmoLex Unigrams
In order to convey such analyses, each charac-
ter in any play represented in 10 dimensional emo- Psychologists proposed many theories for classi-
tional space and mapped onto a plane. The emo- fying human emotions, (Dalgleish, 2000). Some
tional timeline of the plays are also revealed. We emotions are considered basic, while others are
would like to overview the word-emotion associ- considered as complex. The distinction can be
ation lexicon and dimension reduction algorithm. between emotions that we can sense and per-
Then, the last section is left to discussions on the ceive (instinctual), and emotions that can arrive af-
emotional states of the characters and temporal ter some thinking and reasoning (cognitive), (Za-
analyses of the emotions. jonc,1984). There are also oppositions, (Lazarus,
1984). According to Plutchik, (Plutchick, 1985),
1.1 Related Works the discussion may not be resolvable because there
Distant reading is a established ap- is no empirical basis. There is a high correlation
proach/methodology in digital humanities between basic and instinctive emotions, similarly
(DH) and mostly deals with the quantitative between complex and cognitive emotions. Many
analyses of literary and cultural studies (Clement, of the basic emotions are also instinctual.
2008;Crane, 2006). Inside DH, ”Drametrics” is There are theoretical studies on what the ba-
a sub-research field, specialized on quantitative sic emotions are. Ekman argues about about
analysis of the literary genre of drama (Romanska, the existence of 6 basic emotions: joy, sadness,
2015). Digital Shakespeare research and projects anger, fear, disgust, and surprise, (Ekman, 1992).
have gotten attention since the 2000s (Hirsch, Plutchik includes two new emotions: trust and an-
2014; Mueller, 2008). The dramatic structures ticipation, (Plutchick, 1994). Plutchik displayed
in the form of antagonisms are revealed by topic the emotions on a wheel. The distance from the
modeling algorithms, (Yavuz, 2019). Exceptional center in the circle indicates intensity. Plutchik
characters, such as Deus-ex-Machina, is also also state the basic emotions as opposing pairs:
detected in semantic space, (Yavuz, 2020). The joy–sadness, anger–fear, trust–disgust, and antic-
graphs are also used to extract secondary antago- ipation–surprise. The opposing places and neigh-
nisms, (Yavuz,2020). Machine learning based text borhoods on the circle are formed accordingly.
analyses are also carried out for genre classifica- Emolex is the dictionary of English words
tions (Yavuz, 2019; Ardanuy, 2014; Hope, 2010; and their associations with eight basic emotions
Schöch, 2016; Underwood, 2013; Yu, 2008). In (anger, fear, anticipation, trust, surprise, sadness,
literature, structural elements such as dramatis joy, and disgust) and two sentiments (negative and
persone are also analyzed and applications are
developed for further analyses (Dennerlein, 2015;
Krautter, 2018; Schmidt, 2019, Trilcke, 2015;
Wilhelm, 2013). In addition to dramatic structure
works, there is literature that apply sentiment
analyses for dramatic works, (Nalisnick, 2013;
Schmidt, 2018; Schmidt, 2018a; Schmidt, 2018b).
2 Methodology
Our approach is lexicon based. Lines uttered by
each character treated as a document and repre-
sented with a tf-idf. Emotional weights are multi-
plied with the vectors and the summed up to have
the final 10 dimensional emotion space and then
reduced to a plane by linear dimension reduction.
Overall emotional state of any character, thus, is Figure 1: Eight basic emotions (anger, fear, antic-
represented in relative point to each other. Lexicon ipation, trust, surprise, sadness, joy, and disgust)
is also used to extract temporal emotional varia- on Plutchik’s Wheel of Emotions
positive), (Mohammad, 2013). Thus, represent- The opposing emotional pairs (given by Plutchik)
ing each word in 10 dimensional emotion space were represented as a time series as positive / neg-
is effective as well as theoretically relevant. The ative. The states are represented as a cumulative
dataset is unigram. Each single word represented temporal sum and the emotional landscape of the
in 10 dimensional feature space: 8 basic emotions tragedies are revealed.
and 2 sentiment labeling. Each emotion can be
either 0 or 1. No intensity information is given. 3.1 Analyses of Character Emotions
The words are crowd-sourced and manually la-
beled with Mechanical Turk. There are 14182 un-
igrams in the dictionary.
2.2 SVD (Golub, 1970)
Linear dimension reduction methods are for get-
ting a n-dimensional plane over the hyperspace.
For instance, if you have data cloud in 10 dimen-
sion, by mapping onto a plane, one can visualize
such points. SVD is one of the relevant methods,
it breaks any A matrix into three,
A = U SV 0 which (1)
0 0
U U = I and V V = I (2)
S is a diagonal matrix that consists of r singular
values. r is the rank of A. Truncated SVD is a re-
duced rank approximation. Only the most relevant
dimensions are selected, these are the largest sin-
gular values. The dimensions of truncated SVD Figure 2: Emotional position of each character is
are [uxk] ∗ [kxk] ∗ [kxv] Therefore A matrix is represented, Hamlet, Othello, Romeo and Juliet,
approximated by k dimensions, this is the dimen- Macbeth, respectively.
sion reduction. A descriptive subset of the data is
called T, which is a dense summary of the matrix The weights of Tf-Idf features help to position
A, character emotions. Each word contributes to the
resulting emotion with their weights. Weights of
T = U Sk (3) more frequent terms affect the resulting emotional
state more, while less frequent terms affect less.
Sk denotes k largest singular values, which is the The 10-dimensional emotional space mapped onto
number of reduced features. Each feature is repre- an abstraction plane by linear dimension reduc-
sented with a percentage of variance. Higher vari- tion. The dimensions of these planes correspond
ance means more information gain. to abstract emotions or a mixture of the other 10
dimensions with certain proportions. The impor-
3 Discussion
tant thing in these graphics is the position of the
There are two analyzes we would like to discuss. characters relative to each other. The emotional
The first is about whether we can get ideas about positioning of Hamlet in the upper left, Othello in
the plays based on the emotional state of the char- the upper right, Romeo and Juliet in the lower left,
acters. For this purpose, each character is treated and Macbeth in the lower right.
as a text and represented with Tf-Idf features. In these graphics, the basic characteristics of the
The weights of each word for a specific character plays can be observed. The main characters or
multiplied with its 10-dimensional emotion vec- pairs of characters are emotionally different from
tor. Right afterwards, SVD lets you extract prin- the rest. The protagonist and the antagonist al-
ciple axes and mapped onto the plane. This is ways have emotional contrast. For example, the
the first analysis to be discussed. Secondly, the Hamlet play is basically determined by the ten-
temporal dimensions of the plays are considered. sion between the two people, Hamlet and King
Claudius. Although Hamlet is emotionally very or not. The cumulative sum of a emotional con-
different from other characters, King Claudius is trast pair,
emotionally close to the main character cluster. In
T
all tragedies, there is a cluster of emotionally in- X
C({e1 , e2 }, T ) = E[e2 , t] − E[e1 , t] (5)
different characters, we can call the main cluster. t=0
Characters like Lord Polonius and Laertes are also
located around King Claudius with the main clus- which {e1 , e2 } are the random variables for emo-
ter. The Ghost character, like Hamlet, is different tional contrast pairs. The cumulative total for
from all other characters and is in an opposite po- each pair is specified for 5 curves. Cumula-
sition to Hamlet. These observations follow the tive sums of (Negative-Positive), (Sadness-Joy),
readings of the play. In Othello, Iago sets traps to (Disgust-Trust) pairs for all four plays constantly
harm Desdemona. Desdemona is also compatible increases. (Fear-Anger), (Anticipation-Surprise)
with the main cluster. But Iago and Othello are are more neutral. All in all emotionally, the tem-
positioned far apart and apart from the main clus- poral word distributions for tragedies are similar.
ter. In the Macbeth play, the two enemies, Mac-
duff and Malcolm, are opposite and separately po- 4 Conclusion
sitioned. Lady Macbeth is emotionally compatible The traces of the tensions between characters are
with the main cluster. In an interesting observa- observable from the emotional aspect. As we
tion on Romeo and Juliet, the positioning of the show, the emotional positions of the protagonist,
clusters is placed in symmetry in accordance with the antagonist and the main cluster gives much
Renaissance thought. It is known that the play insight about the greatness such pieces. Any
is written symmetrically. There are three family great tragedy needs emotional contrast between
positions in symmetry: Ruling house of Verona, the main characters and there is always the main
House of Capulet, House of Montague. cluster. The temporal-emotional characteristics of
The graphs show that the emotional positioning the plays are also important and very much simi-
of tragedies is compatible with the readings of the lar to each other. There are constantly increasing
play. What we mean is the protagonists and antag- emotions as well as neutrals. Each play grows to-
onists are clearly observable. Distances or orien- wards positive, joyful and trusty emotional state.
tations, or rather relative positions, are significant. This might be the reason behind a followable play.
The main characters that experience basic tensions Positive feelings should accumulate.
could be demonstrated. In the play of Romeo and Either it is computer-assisted or fully automated
Juliet, the affinities are observed and there is sym- writing machine, artificial literature needs emo-
metrical positioning of the families. tional aspect. The emotional aspect of literary
works should be conditional of such generative
3.2 Temporal-Emotional Evaluation of the
models. The common acceptance on this early
Tragedies
Temporal-emotional evaluation of the tragedies
are drawn. We can assume that each play is a
temporal series, and we summed up the emotional
state in each timestep, the cumulative emotional
curves are calculated. Therefore, the emotional di-
rections are determined. Emotions are positioned
in contrast as (Negative-Positive), (Fear-Anger),
(Anticipation-Surprise), (Sadness-Joy), (Disgust-
Trust). For each timestep, or the word, the contri-
bution is the expected emotion,
E[e] = e ∗ p(e) (4)
which p(e) is the occurrence probability of the Figure 3: Emotional landscape of each character
emotion in the lexicon and e is the Bernoulli ran- is represented, Hamlet, Othello, Romeo and Juliet,
dom variable, either 0 or 1, either has the emotion Macbeth, respectively.
modern author is his greatness as a tragedy writer. Hope, J., & Witmore, M. (2010). The Hundredth
The theatrical pieces by Shakespeare are in dialog Psalm to the Tune of ”Green Sleeves”: Digital
Approaches to Shakespeare’s Language of Genre.
form, each character express themselves clearly.
Shakespeare Quarterly, 61(3), 357-390. Retrieved
Therefore, it is shown that any dramatic antago- from http://www.jstor.org/stable/40985589
nism is also emotional. Any artificial dramatic
work should have a similar emotional resonance Krautter, B. (2018). Quantitative microanalysis? Dif-
ferent methods of digital drama analysis in com-
with such tragedies. With this analyses, we try to parison. Book of Abstracts, DH 2018. Mexico-City,
further develop evaluation metrics for artificial lit- Mexico, pp. 225-228.
erature. A baseline metric to emotionally evaluate
such theatrical forms. Lazarus, R. S. (1984). On the primacy of cog-
nition. American Psychologist, 39(2), 124–129.
On the way to Artificial Literature (ALit), there https://doi.org/10.1037/0003-066X.39.2.124
needs more criteria and more complex tools to
analyze literariness of such pieces (literariness, Le, Q., Mikolov, T. (2014, January). Distributed rep-
resentations of sentences and documents. In Inter-
2020). As we shown so far, emotion aspect of the national conference on machine learning (pp. 1188-
plays are very crucial at establishing antagonisms. 1196).
Lebrun, T. (2017, May). Who Is the Artificial Author?.
References In Canadian Conference on Artificial Intelligence
(pp. 411-415). Springer, Cham.
Ardanuy, M. C., & Sporleder, C. (2014, April).
Structure-based clustering of novels. In Proceedings literariness. Oxford Reference. Re-
of the 3rd Workshop on Computational Linguistics trieved 8 Jun. 2020, from
for Literature (CLFL) (pp. 31-39). https://www.oxfordreference.com/view/10.1093/oi/
authority.20110803100108912.
Clement, T., Steger, S., Unsworth, J. and
Uszkalo, K. (2008). How not to read Mohammad, S. M., Turney, P. D. (2013). Crowdsourc-
a million books. Available online at ing a word–emotion association lexicon. Computa-
http://people.brandeis.edu/ unsworth/hownot2read.html tional Intelligence, 29(3), 436-465.
Crane, G. (2006). What do you do with a mil- Moretti, F. (2013). Distant reading. Verso Books.
lion books? D-Lib Magazine. Available online at
http://www.dlib.org/dlib/march06/crane/03crane.html Moretti, F. (2000). Conjectures on world literature.
New left review, 54-68.
Dalgleish, T., Power, M. (Eds.). (2000). Handbook of
Mueller, Martin. (2008). Digital Shakespeare, or to-
cognition and emotion. John Wiley Sons.
wards a literary informatics. Shakespeare. 4. 284-
Dennerlein, K. (2015). Measuring the average popu- 301. 10.1080/17450910802295179.
lation densities of plays. A case study of Andreas Nalisnick, E. T., Baird, H. S. (2013, August).
Gryphius, Christian Weise and Gotthold Ephraim Character-to-character sentiment analysis in Shake-
Lessing. Semicerchio. Rivista di poesia comparata speare’s plays. In extitProceedings of the 51st An-
LIII: 80–88. nual Meeting of the Association for Computational
Linguistics (Volume 2: Short Papers) (pp. 479-483).
Ekman, P. (1992). An argument for basic emotions.
Cognition emotion, 6(3-4), 169-200. Plutchik, R. (1985) On emotion: The chicken-and-
egg problem revisited. Motiv Emot 9, 197–200 .
Golub, G. H.; Reinsch, C. (1970). ”Singular https://doi.org/10.1007/BF00991576
value decomposition and least squares solu-
tions”. Numerische Mathematik. 14 (5): 403–420. Plutchik, R. (1994) The psychology and biology of
doi:10.1007/BF02163027. MR 1553974. emotion. HarperCollins College Publishers.
Gao, J., Galley, M., Li, L. (2018, June). Neural ap- Ramsay, S. (2011). Reading Machines: Toward an Al-
proaches to conversational AI. In The 41st Interna- gorithmic Criticism. University of Illinois Press.
tional ACM SIGIR Conference on Research Devel-
opment in Information Retrieval (pp. 1371-1374). Romanska, M. (2015). Drametrics: what dramaturgs
should learn from mathematicians. In Romanska,
Hirsch, B., & Craig, H. (2014). ”Mingled Yarn”: M. (ed.), The Routledge Companion to Dramaturgy.
The State of Computing in Shakespeare 2.0. In T. Routledge, pp. 472-481.
Bishop, & A. Huang (Eds.), The Shakespearean In-
ternational Yearbook (Vol. 14: Special Section, Dig- Schmidt, T., Burghardt, M., Dennerlein, K. & Wolff, C.
ital Shakespeares, pp. 3-35). United Kingdom: Ash- (2019). Katharsis – A Tool for Computational Dra-
gate Publishing Limited. metrics. In Book of Abstracts, DH 2019.
Schmidt, T., Burghardt, M. (2018, August). An evalu- the Computational Processing of Portuguese (PRO-
ation of lexicon-based sentiment analysis techniques POR 2020), Évora, Portugal., CEUR-WS.org, on-
for the plays of gotthold ephraim lessing. In Pro- line http://ceur-ws.org/Vol-2481/paper75.pdf.
ceedings of the Second Joint SIGHUM Workshop
on Computational Linguistics for Cultural Heritage, Yavuz, M. C. (2020, November) ”Analyses of Charac-
Social Sciences, Humanities and Literature (pp. ter Networks in Dramatic Works by Using Graphs”
139-149). The 7th International Conference on Behavioural
and Social Computing), Bournemouth UK.
Schmidt, T. and Burghardt, M. (2018b). Toward a Tool
for Sentiment Analysis for German Historic Plays. Yu, B. (2008). An evaluation of text classification
In Piotrowski, M. (ed.), COMHUM 2018: Book of methods for literary study. Literary and Linguistic
Abstracts for the Workshop on Computational Meth- Computing 23(3): 327-343.
ods in the Humanities 2018. Lausanne, Switzerland:
Laboratoire laussannois d’informatique et statistique Xanthos, A., Pante, I., Rochat, Y and Grandjean,
textuelle. pp. 46-48. M. (2016). Visualising the dynamics of character
networks. Book of Abstracts, DH 2016. Kraków,
Schmidt, T., Burghardt, M. and Dennerlein, K. (2018a). Poland, pp. 417-419.
Sentiment Annotation of Historic German Plays: An Zajonc, R. B. (1984). On the primacy of affect. Ameri-
Empirical Study on Annotation Behavior. Sandra can Psychologist, 39(2), (1984). 117–123.
Kübler, Heike Zinsmeister (eds.), extitProceedings
of the Workshop on Annotation in Digital Humani-
ties (annDH 2018) Sofia, Bulgaria. pp. 47-52.
Schmidt, T., Burghardt, M. and Dennerlein, K.
(2018b). Kann man denn auch nicht lachend
sehr ernsthaft sein?“ – Zum Einsatz von Senti-
ment Analyse-Verfahren für die quantitative Unter-
suchung von Lessings Dramen. In Book of Abstracts,
DHd 2018 , Cologne, Germany.
Schöch, Christof. (2016). Topic Modeling Genre:
An Exploration of French Classical and En-
lightenment Drama. Digital Humanities Quarterly.
http://doi.org/10.5281/zenodo.166356
Trilcke, P., Fischer, F. and Kampkaspar, D. (2015).
Digital Network Analysis of Dramatic Texts. Book
of Abstracts, DH 2015. Sidney, Australia
Underwood, T., Black, M.L., Auvil, L., & Capitanu,
B. (2013). Mapping mutable genres in structurally
complex volumes. 2013 IEEE International Confer-
ence on Big Data, 95-103.
Wilhelm, T., Burghardt, M., and Wolff, C. (2013). “To
See or Not to See” - An Interactive Tool for the
Visualization and Analysis of Shakespeare Plays.
In R. Franken-Wendelstorf, E. Lindinger, and J.
Sieck (Eds.), Kultur und Informatik: Visual Worlds
& Interactive Spaces. Glückstadt: Verlag Werner
Hülsbusch, pp. 175–185.
Yavuz, M. C.: (2019, November) Analyses of Liter-
ary Texts by Using Statistical Inference Methods.
In: Proceedings of the Sixth Italian Conference
on Computational Linguistics, CLiC-it’19. Bari,
Italy., CEUR-WS.org, online http://ceur-ws.org/Vol-
2481/paper75.pdf.
Yavuz, M. C.: (2020, March) Analyses of Characters in
Dramatic Works by Using Document Embeddings.
In: Proceedings of the Workshop on Digital Human-
ities and Natural Language Processing (DHandNLP
2020) co-located with International Conference on