=Paper=
{{Paper
|id=Vol-3117/paper7
|storemode=property
|title=MARCUS: An Event-Centric NLP Pipeline that generates Character Arcs from Narratives
|pdfUrl=https://ceur-ws.org/Vol-3117/paper7.pdf
|volume=Vol-3117
|authors=Sriharsh Bhyravajjula,Ujwal Narayan,Manish Shrivastava
|dblpUrl=https://dblp.org/rec/conf/ecir/BhyravajjulaN022
}}
==MARCUS: An Event-Centric NLP Pipeline that generates Character Arcs from Narratives==
https://ceur-ws.org/Vol-3117/paper7.pdf
MARCUS: An Event-Centric NLP Pipeline that
generates Character Arcs from Narratives
Sriharsh Bhyravajjula1 , Ujwal Narayan1 and Manish Shrivastava1
1
International Institute of Information Technology, Hyderabad
Abstract
Character arcs are important theoretical devices employed in literary studies to understand character
journeys, identify tropes across literary genres, and establish similarities between narratives. This work
addresses the novel task of computationally generating event-centric, relation-based character arcs from
narratives. Providing a quantitative representation for arcs brings tangibility to a theoretical concept and
paves the way for subsequent applications. We present MARCUS (Modelling Arcs for Understanding
Stories), an NLP pipeline that extracts events, participant characters, implied emotion, and sentiment to
model inter-character relations. MARCUS tracks and aggregates these relations across the narrative to
generate character arcs as graphical plots. We generate character arcs from two extended fantasy series,
Harry Potter and Lord of the Rings. We evaluate our approach before outlining existing challenges,
suggesting applications of our pipeline, and discussing future work.
Keywords
Computational Literary Studies, Character Arcs, Relations, Events, Narratives
1. Introduction
Characters in narratives attempt to influence their circumstances to resolve conflict, while
circumstance itself shapes the characters with events that develop them [1]. This character
journey is integral to narratives [2]; good stories are driven by the transformative journeys of
compelling characters. This work addresses the challenge of quantifying these journeys using
character arcs modeled around events and relations.
The field of computational literary studies aims to understand, represent and generate narra-
tives. Existing research has focused on various perspectives, notably plot units [3, 4], social
network extraction from narratives [5], and character-centric approaches [6, 7]. This work
builds upon the latter to extract characters from narratives using co-reference resolution and
clustering techniques. We find inspiration in research exploring events in narratives [8] and
argue that the transformation of characters is a consequence of events that involve these char-
acters. We also take advantage of recent advances in semantic role labeling [9] to understand
the function of characters in events and build on advancements in emotion analysis [10, 11] to
represent how events affect both agent and recipient characters quantitatively.
In: R. Campos, A. Jorge, A. Jatowt, S. Bhatia, M. Litvak (eds.): Proceedings of the Text2Story’22 Workshop, Stavanger
(Norway), 10-April-2022
Envelope-Open s.bhyravajjula@research.iiit.ac.in (S. Bhyravajjula); ujwal.narayan@research.iiit.ac.in (U. Narayan);
m.shrivastava@iiit.ac.in (M. Shrivastava)
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org)
67
Figure 1: Relation-Based Character Arc for Frodo Baggins across The Lord of the Rings trilogy. The blue
line represents Frodo the actor, and the red line represents Frodo the experiencer. A: Frodo is stabbed by
the poisoned blade of the Nazgul at Weathertop (valley). B: Frodo reunites with loved ones at Rivendell
(peak). C: Frodo is in grief after the wizard Gandalf falls to a Balrog (valley). D: Frodo is attacked by
Shelob in her lair (valley). E: Frodo succeeds in his quest and returns home to The Shire (peak).
Our approach differs from recent work in the closely related field of extracting relationships
between characters in terms of goals and linguistic features. Some existing approaches either
lack the nuance of emotions [12, 13] or do not leverage semantic role labels for characters [14].
Other approaches fail to utilize fine-grained sentiments [15] or do not share our event-based
perspective on relationships [16, 17]. While there has been progress in the field of emotion arcs,
it has been centered around plot rather than character, capturing a shift in sentiment rather
than a shift in circumstance [18, 19]. To the best of our knowledge, we believe that our task and
approach are novel additions to the field.
2. Key Concepts
2.1. Events
We borrow from [8] and focus solely on events with asserted realis (depicted as actually tak-
ing place, with specific participants at a specific time) instead of those with other epistemic
modalities (hypotheticals, future events). We interpret events as indicators of inter-character
relationship states - for each event, we extract the latent predicate-argument structure to identify
participants. For example, in the sentence “Harry kicked Ron”, “kicked” is the event with Harry
the actor and Ron the experiencer.
2.2. Circumstance
We take inspiration from [2] and use implied sentiment and emotion to quantify circumstance,
the state of fortune associated with a directed pair of characters participating in an event.
Narratives and plots span a plethora of settings - each character ‘reacts to’ as well as ‘influences’
their unique circumstance; it would be beyond the scope of this paper to establish a universal
scale. Instead, we focus on the relative shift of circumstance across the narrative.
68
Table 1 Table 2
Dataset Details RoBERTa Sentiment Regression Model
Data Source Word Count Event Count Metric Score
Harry Potter 1,095,940 93,782 Mean Squared Error 0.01620
Lord of the Rings 478,329 28,670 Mean Absolute Error 0.09693
2.3. Relation Arcs and Character Arcs
As the story progresses, the relationship between a pair of characters evolves. A relation arc
is a plot of the shift in circumstance of a directed pair of characters across their participatory
events in a narrative. While the actor of an event influences the circumstances that affect the
experiencer, the latter’s subsequent actions are captured in other relation arcs, allowing the
effect of events to trickle through multiple connected arcs forming the plot of a narrative. We
posit that a character’s journey at any point in the narrative can be represented by contextually
assessing the amalgamated effect of their interactions with both themselves and other characters
on their circumstance; a character arc is thus defined as a pair of quantitative aggregations of
all the corresponding relation arcs of a character as both actor and experiencer respectively.
Fig 1 shows the character arc for the protagonist of the Lord of the Rings, Frodo Baggins,
capturing the shift of circumstances he experiences as both an actor and experiencer. The events
Frodo participates in change his circumstances - a drop to the valleys (minimas) represents a
deterioration of circumstances, whereas a rise to the peaks (maximas) denotes an improvement.
It is evident from C and the arc slice between D and E that actor and experiencer arcs of a
character are not always aligned, allowing for the intuitive explanation of instances when
characters oppose the circumstances they find themselves in, for better or for worse.
3. MARCUS
We present MARCUS (Modelling Arcs for Understanding Stories), an event-centric NLP pipeline1
that generates character arcs from narratives. In this section, we will elaborate upon the various
components of the pipeline.
3.1. Dataset
We failed to retrieve meaningful character arcs in our preliminary experiments with short stories
containing less than 500 events. Consequently, we choose two larger datasets for character arc
analysis instead, namely The Harry Potter septology and The Lord of the Rings trilogy. For
each source, we ensure that we remove page numbers, unnecessary annotations, unreadable
characters, weblinks, etc. The details of each dataset are detailed in Table 1.
1
https://github.com/darthbhyrava/MARCUS
69
Emotion Identification
Novel BookNLP Event Extraction Participant Extraction Circumstance Relation Arc Character Arc
Sentiment Identification
Scoring
Figure 2: MARCUS (Modeling Arcs for Understanding Stories), an NLP pipeline that plots a character’s
arc as their quantitative interaction with circumstance as both actor and experiencer, represented by
the proxy amalgamation of their event-centric relations across the narrative.
3.2. BookNLP
BookNLP [7] is an NLP pipeline2 that scales to books and long documents (in English) and per-
forms tasks like part-of-speech tagging (Stanford), dependency parsing (MaltParser), named en-
tity recognition (Stanford), character name clustering (e.g., “Tom”, “Tom Sawyer”, “Mr. Sawyer”
: “Tom Sawyer”) and pronominal coreference resolution. MARCUS uses BookNLP to retrieve
character occurrences and linguistic features needed for event and participant extraction.
3.3. Event Extraction
We feed the features processed from BookNLP into a tagger consisting of a BiLSTM model with
BERT embeddings to extract events and corresponding entities from the narrative. The tagger
is trained3 using the LitEvents dataset [8], which provides an annotated list of events from over
100 public domain novels from Project Gutenberg. MARCUS uses tagged events and entities for
extracting participant characters, sentiment, and implied emotion.
3.4. Participant Extraction
We process each event with a BERT-based Semantic Role Labeller provided by AllenNLP4
to extract the latent predicate-argument structure of the event and identify roles of partici-
pant characters. If there are multiple events in the same sentence with the same actors and
experiencers, we consider only the first event to avoid redundancy.
3.5. Circumstance
Every actor/experiencer pair corresponding to an event in the narrative is assigned a quantitative
measure of circumstance. We argue that proxy indicators such as sentiment and implied emotion
implicitly capture an event’s circumstances and are specifically well suited to our focus on
shift of circumstance. Absolute measures of circumstance can therefore be interpreted as a
characteristic of the genre or trope of narrative, while their shifts are a characteristic of the
character’s journey. Since circumstances evolve over time (or event sequences), MARCUS
considers the effect of previous circumstances between characters when defining relation arcs.
2
https://github.com/dbamman/book-nlp
3
https://github.com/dbamman/litbank
4
https://docs.allennlp.org/models/main/models/structured_prediction/models/srl_bert/
70
3.5.1. Sentiment Identification
We pose sentiment extraction as a regression task to capture the subtleties of relationships. We
fine-tune a RoBERTa model on the Stanford Sentiment Treebank (SST) [20] dataset to obtain
a fine-grained sentiment score in the range of 0 to 1. The SST dataset provides 12𝑘 sentences
and phrases with their associated sentiment scores lying between 0 to 26, which we normalize
before training for ten epochs in a 60:20:20 split. The metrics for this model are reported in Table
2. MARCUS uses the model to assign sentiment scores to events extracted earlier in Section 3.3.
3.5.2. Emotion Identification
Sentiment alone may not give us enough information about circumstance - we argue that in
such cases, multi-faceted emotional states help capture shift in circumstance by leveraging
the nuances of relationships. To identify emotions, we use a BERT model5 trained on the
GoEmotions [10] dataset in a multi-label setting, as interactions can have more than one
emotional undertone. The GoEmotions dataset consists of 58k Reddit comments manually
annotated for 27 emotion categories: admiration, amusement, sorrow, fear, etc. MARCUS uses
the confidence of the model’s predicted labels as well as a manually assigned value for each
label to contribute to the measure of circumstance. These labels, in the range of −2 to 2, are
assigned based on intensity of emotion; higher intensity corresponds to higher absolute value.
3.6. Relation and Character Arcs
MARCUS generates relation arcs by plotting the measure of circumstance, 𝑡, for every event, 𝑒,
belonging to an actor/experiencer pair of characters across the narrative.
𝐿
𝑡e = 𝛼 ∗ 𝑠e + ∑ 𝛽𝑖 ∗ 𝑐ie (1)
𝑖=1
where 𝑡e is the measure of circumstance for event 𝑒, 𝛼 ∈ (0, 1) is the sentiment co-efficient that
controls how much influence the fine grained sentiment score should have over relation arcs, 𝑠e
∈ (0, 1) is the sentiment of that event, 𝐿 is the total number of emotion labels for that event,
𝛽i ∈ [−2, 2] is the fixed score for emotion label, and 𝑐ie ∈ (0, 1) is the corresponding confidence
score for each emotion label in the event. We run multiple experiments to choose optimal values
of 𝛼 and 𝛽: their final values are listed in the code.
We apply a window function 𝑅(.) over te , the set of all measures of circumstance corresponding
to the event set e, to calculate the relation arc, r, given by
r = 𝑅(te , 𝑛, 𝑝) (2)
where 𝑅(.) is the window function that helps smoothen the arcs while retaining previous
state information, 𝑛 is the window size, and 𝑝 is an optional parameter for specifying order for
polynomial fitting.
As shown in Fig 3a, the relation arcs are too noisy without smoothing or retention of previous
circumstance information. We experiment with three standard window functions: Rolling
5
https://github.com/monologg/GoEmotions-pytorch
71
(a) (b)
Figure 3: (a) Character Arc for Hermione, No Rolling Function Applied; (b) Character Arc for Hermione,
with a savgol filter of window size 1/10th of her event sequence length, fitted with a third degree
polynomial.
Mean, Triangular Rolling Mean and Savitzky-Golay Filter [21]. We find that the Savitzky-Golay
Filter represents the narrative most accurately as seen in Fig 3b, and use the same for all arcs
represented in the paper. We generate the character arc, c, by adding up the corresponding
relation arcs r of that character over all events e in the role of actor and experiencer respectively.
4. Evaluation
4.1. Survey
We ask a set of 16 human volunteers (avid fiction readers aged 20-31) to peruse both the Harry
Potter and Lord of the Rings series, following which they evaluate our system by answering
surveys on two tasks: idchar, where the volunteers are given a list of relation arcs and character
pairs and asked to match the arcs to their corresponding pairs, and idplot, where the volunteers
are given character arcs, pertinent plot events and asked to identify the points in the arc that
they think represent the corresponding events.
For both these tasks mentioned above, we calculate accuracy. Task idchar achieved an
accuracy of 71.5% and task idplot had an accuracy of 72.9%. We also evaluate with the Fleiss
Kappa metric where category 1 indicates a complete match, and 0 indicates otherwise. For both
the tasks, we have a score of 0.675 and 0.528 indicating strong and moderate inter-annotator
agreement, respectively. Thus, most of the volunteers consistently identified both the character
pairs and the relevant points in the graph given the plot sequence.
4.2. Gold Labels
We have a volunteer extremely familiar with the story annotate the first 300 events of Lord of
the Rings trilogy involving Frodo Baggins as a participant character. The annotator marks each
event with three labels denoting a positive, neutral or negative shift in circumstance. Tables 3
and 4 illustrate the positive and negative shifts as tagged by our system and the corresponding
gold labels for the same events provided by our annotator. Our system tends to assign positive
labels to neutral events and has higher accuracy for negative shifts of circumstance.
72
Table 3 Table 4
Positive Shifts Negative Shifts
Gold Label Percentage Gold Label Percentage
Positive 0.36 Positive 0.12
Neutral 0.30 Neutral 0.15
Negative 0.33 Negative 0.73
5. Challenges and Future Work
We identify five notable challenges in our approach that can be addressed in future work.
Firstly, since MARCUS is a sequential pipeline, it is challenging to determine the effect of
errors cascading through the system quantitatively. Secondly, our rolling window makes the
arcs dependent on the availability of data; event paucity in short stories or characters with
low interactions hinders accurate arc generation. Thirdly, we observe in our arcs that our
fine-grained events do not represent an abstract view of the discourse - a more contextual
representation of events is needed. Fourth, in our understanding of a character’s circumstance,
localized interactions with other event-specific characters heavily influence shifts; we need an
effective means of capturing the latent relative importance of character-specific interactions.
And lastly, our approach does not aim to handle non-linear narratives where events are not
sequentially presented.
6. Applications
Providing tangibility to the theoretical concept of character arcs, MARCUS can be employed
in a variety of applications. Character arcs can be used in a more nuanced approach for
detecting similarity between narratives by focusing on character journeys, leading to a possible
improvement in book recommendations and movie recommendations based on stories and
scripts. Character arcs can also help with digital enrichment in e-readers, adding to the rich
metadata provided by devices like Kindle. Character arcs can also function as guidance for
natural language generation tasks in the field of fiction. And lastly, they can help narrative
studies by identifying character tropes and for identification of personality traits.
7. Conclusion
We propose MARCUS (Modeling Arcs for Understanding Stories), an NLP pipeline that addresses
the novel task of generating character arcs from narratives. We explain key concepts like events
and circumstance and delve into the details of our event-centric approach which leverages
proxy markers like sentiment and emotion. We then evaluate our pipeline, discuss challenges,
elucidate future work and outline potential applications.
73
References
[1] K. M. Weiland, Creating Character Arcs, The Masterful Author’s Guide to Uniting Story
Structure, Plot, and Character Development, PenForASword Publishing, 2016.
[2] K. Vonnegut, Kurt Vonnegut on the Shapes of Stories, YouTube, 1995. URL: https://youtu.be/
oP3c1h8v2ZQ.
[3] M. Elsner, Character-based kernels for novelistic plot structure, EACL, 2012.
[4] W. Lehnert, Plot units and narrative summarization, Cogn. Sci. 5 (1981) 293–331.
[5] A. Agarwal, A. Kotalwar, J. Zheng, O. Rambow, SINNET: Social interaction network
extractor from text, IJCNLP (System Demonstrations), 2013.
[6] L. Flekova, I. Gurevych, Personality profiling of fictional characters using sense-level links
between lexical resources, EMNLP, 2015.
[7] D. Bamman, T. Underwood, N. A. Smith, A Bayesian mixed effects model of literary
character, ACL, 2014.
[8] M. Sims, J. H. Park, D. Bamman, Literary event detection, ACL, 2019.
[9] P. Shi, J. J. Lin, Simple bert models for relation extraction and semantic role labeling (2019).
[10] D. Demszky, D. Movshovitz-Attias, J. Ko, A. S. Cowen, G. Nemade, S. Ravi, Goemotions: A
dataset of fine-grained emotions, ACL (2020).
[11] S. Zad, M. Finlayson, Systematic evaluation of a framework for unsupervised emotion
recognition for narrative text, in: Proceedings of the First Joint Workshop on Narrative
Understanding, Storylines, and Events, ACL, 2020, pp. 26–37.
[12] S. Srivastava, S. Chaturvedi, T. M. Mitchell, Inferring interpersonal relations in narrative
summaries, CoRR abs/1512.00112 (2015).
[13] S. Chaturvedi, S. Srivastava, H. D. III, C. Dyer, Modeling dynamic relationships between
characters in literary novels, CoRR abs/1511.09376 (2015).
[14] M. Iyyer, A. Guha, S. Chaturvedi, J. Boyd-Graber, H. Daumé III, Feuding families and
former Friends: Unsupervised learning for dynamic fictional relationships, NAACL-HLT,
ACL, 2016.
[15] F. Rashid, E. Blanco, Characterizing interactions and relationships between people, EMNLP,
2018.
[16] E. Kim, R. Klinger, Frowning frodo, wincing leia, and a seriously great friendship: Learning
to classify emotional relationships of fictional characters, CoRR abs/1903.12453 (2019).
[17] S. Chaturvedi, M. Iyyer, H. D. III, Unsupervised learning of evolving relationships between
literary characters, AAAI, 2017.
[18] S. Somasundaran, X. Chen, M. Flor, Emotion arcs of student narratives, in: Proceedings of
the First Joint Workshop on Narrative Understanding, Storylines, and Events, ACL, 2020.
[19] A. J. Reagan, L. Mitchell, D. Kiley, C. M. Danforth, P. S. Dodds, The emotional arcs of
stories are dominated by six basic shapes, CoRR abs/1606.07772 (2016).
[20] R. Socher, A. Perelygin, J. Wu, J. Chuang, C. D. Manning, A. Ng, C. Potts, Recursive deep
models for semantic compositionality over a sentiment treebank, EMNLP, 2013.
[21] A. Savitzky, M. J. Golay, Smoothing and differentiation of data by simplified least squares
procedures., Analytical chemistry 36 (1964) 1627–1639.
74