Stance Detection in Twitter Conversations Using Reply Support
Classification
Parul Khandelwal
1 a
, Preety Singh
1 b
, Rajbir Kaur
1 c
and Roshni Chakraborty
2 d
1
The LNM Institute of Information Technology, Jaipur, India
2
ABV IIITM, Gwalior, India
Keywords:
Stance, Twitter, Reply Support Label, LSTM, BERT.
Abstract:
During crisis, social media platforms like Twitter play a crucial role in disseminating information and offering
emotional support. Understanding the conversations among people is essential for evaluating the overall im-
pact of the crisis on the public. In this paper, we focus on classifying the replies to tweets during the “Fall of
Kabul” event into three classes: supporting, unbiased, and opposing. To achieve this goal, we proposed two
frameworks. We used LSTM layers for sentence/word-level feature extraction for classification. We also em-
ployed a BERT-based approach where the text of both the tweet and the reply is concatenated.Our evaluation
on real-world crisis data showed that the BERT-based architecture outperformed the LSTM models. It pro-
duced an F1-score of 0.726 for the opposing class, 0.738 for the unbiased class, and 0.729 for the supportive
class. These results highlight the robustness of contextualized embeddings in accurately identifying the stance
of replies within Twitter conversations through tweet-reply pairs.
1 INTRODUCTION
Social media platforms, especially Twitter, can be a
helpful tool for disseminating information or to share
emotions in case of any crisis. Tweets often pro-
vide a straight line of communication between the
affected population and concerned authorities(Bukar
et al., 2022). Considerable research has been done
on stance classification in events. Stance, has gen-
erally been referred to as public opinion, majorly in
reference to government policies, social movements,
pandemics, etc. (ALDayel and Magdy, 2021). In
most research, attention has been focused on deter-
mining whether a tweet will support, oppose, or main-
tain neutrality toward a certain topic or entity (K
¨
uc¸
¨
uk
and Can, 2020). Stance detection has been done for
political discourses (Lai et al., 2019), health misinfor-
mation (Ng and Carley, 2022), rumors identification
(Zheng et al., 2022; Haouari and Elsayed, 2024), cri-
sis scenarios (Zeng et al., 2016), or other social issues.
However, less attention has been paid to the stance
expressed in individual conversation threads. Classi-
a
https://orcid.org/0000-0003-4963-156X
b
https://orcid.org/0000-0002-6885-1376
c
https://orcid.org/0000-0003-1433-1065
d
https://orcid.org/0000-0003-4476-403X
fying tweet-reply interactions presents a more com-
plex and nuanced understanding of online conversa-
tions. A reply can either be building on, challeng-
ing, or providing a little context to the original tweet
(Zubiaga et al., 2016). This can be indispensable for
judgment regarding public opinion. Tweets provide
the initial viewpoint or information while analyzing
replies can capture the full spectrum of the discourse.
This gap in research can be filled through stance clas-
sification in conversation threads to supplement the
dialogue dynamics especially during crisis manage-
ment. The categorization of the stance of replies can
reveal if important information is being amplified or
contested. This will enable the authorities to evalu-
ate how people are perceiving updates, advisories, or
policies. With this additional knowledge, emergency
responders can modify their communication strate-
gies. This will also help understand the sentimen-
tal conditions of the affected population and gauge
the level of solidarity and aid being shared (Hardalov
et al., 2021).
Reactions to tweets can take forms of support, op-
position, or additional context and can thereby sig-
nificantly affect the dynamics of the virtual discus-
sions. The classification of responses can shed more
light on the actual impact of the seed tweets during
crisis situations. This study focuses on automatically
Khandelwal, P., Singh, P., Kaur, R. and Chakraborty, R.
Stance Detection in Twitter Conversations Using Reply Support Classification.
DOI: 10.5220/0013129800003905
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 235-242
ISBN: 978-989-758-730-6; ISSN: 2184-4313
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
235

classifying tweet-reply pairs to more correctly com-
prehend public engagement and sentiment in conver-
sation threads. Our contributions are summarized be-
low:
• Conversation Stance Classification: We pro-
pose classifying tweet-reply conversations, focus-
ing on both their texts, to assess the support flow
of online social media conversations during a cri-
sis.
• Experimentation on Multiple Frameworks:
We present two distinct classification frameworks,
utilizing LSTM-based models (Hochreiter, 1997)
and a BERT-based architecture (Devlin, 2018).
Each framework handles tweet-reply embeddings
and sequence processing distinctly.
• Explainable AI: We apply the concept of ex-
plainable AI to interpret the predictions of our
stance classification model.
The paper is organized as follows: Section 2 dis-
cusses existing research on tweet classification and
reply analysis in crisis. Section 3 outlines the method-
ology, including the dataset and description of the
frameworks. Section 4 provides details on the experi-
mental setup. Section 5 discusses the results and anal-
ysis, comparing the performance of each framework.
Section 6 concludes the paper.
2 RELATED WORK
Stance classification has evolved through various
methodologies to analyze public sentiment and dis-
course on social media platforms. (Mohammad et al.,
2016) introduced stance detection by identifying sup-
port, opposition, or neutrality toward specific targets.
(Zubiaga et al., 2016) extended this work to classify
stance in conversational threads using tree structures,
categorizing replies as supporting, denying, querying,
or commenting. (Mutlu et al., 2020) explored public
perceptions using TF-IDF and CNN-based methods,
while (Villa-Cox et al., 2020) focused on replies and
quotes in controversial Twitter conversations. Simi-
larly, (Hamad et al., 2022) developed the StEduCov
dataset to classify online education stances during the
COVID-19 pandemic.
Recent advancements include multi-task frame-
works like (Abulaish et al., 2023), which employed
Graph Attention Networks (GAT) for jointly predict-
ing stance and rumor veracity, and (Bai et al., 2023),
who proposed the Multi-Task Attention Tree Neural
Network (MATNN) for hierarchical classification of
stance and veracity. (Zhang et al., 2024) introduced
DoubleH, leveraging user-tweet bipartite graphs for
stance detection related to the 2020 US presiden-
tial election. While these studies showcase diverse
methodologies, they primarily focus on individual
tweets or structured datasets, overlooking the stance
of replies toward parent tweets in tweet-reply conver-
sations.
3 METHODOLOGY
In this section, we present our methodology to cat-
egorize the support of replies to the parent tweet in
a conversation thread. A high level of support indi-
cates effective communication of the messages. Op-
posing or unbiased replies might suggest misinforma-
tion or confusion among the affected population. Us-
ing such insights can empower crisis managers and
policymakers to make appropriate, real-time adjust-
ments to messaging strategies.
Problem Statement: Consider a dataset D , consist-
ing of n tweet-reply pairs {(T
i
, R
i
)}
n
i=1
, where T
i
rep-
resents the text of the i-th tweet and R
i
represents the
corresponding reply text. Each tweet T
i
has a tweet
label assigned to it.Tweets providing factual data re-
lated to the crisis are labelled as information. Those
reflecting personal reactions and appeals for support
are labelled as emotion. Tweets stating personal opin-
ions are termed as neutral. Classifying tweets offers
insights into the nature of the discourse being car-
ried on during the crisis. A tweet can have multiple
replies, forming many tweet-reply pairs. The goal is
to classify the reply support label into one of the fol-
lowing three categories:
• Supporting: The reply demonstrates alignment
with or endorses the parent tweet.
• Unbiased: Reply neither supports nor opposes
the parent tweet.
• Opposing: The reply expresses opposition to the
parent tweet.
Considering the following example:
Tweet Example (tweet label: information): The
US plans to completely pull all personnel from
its embassy in Kabul over the next 72 hours as
Taliban forces close in on Afghanistan’s capital
https://cnn.it/3CRwPXG
• Supporting Reply: ”Bring them home @POTUS
@JoeBiden !!! Thank you !! You are doing good.”
• Unbiased Reply: #BidensSaigon
• Opposing Reply: Sickens me beyond comprehen-
sion. Biden is not only a liar and self serving
crook, he
ˆ
aC™s a coward. And where is he and
where has he been?? He had all the answers the
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
236

last 4 years and now he disappears and is literally
putting the US in harms way.
There were 123 replies on this tweet at the time of
data collection. Of these, 48 were opposing, 17 were
supporting and 58 were unbiased. Such an analysis
through different conversation threads can provide an
indication on the inclination of the people’s responses
to the developments during the crisis.
Formally, our problem becomes a mapping of the
function f : D → y
i
∈ {−1, 0, 1} for the classification
of stances in tweet-reply pairs:
f (T
i
, R
i
) =





1, if R
i
supports T
i
0, if R
i
is neutral towards T
i
−1, if R
i
opposes T
i
The objective is to minimize the classification er-
ror by using the sparse categorical cross-entropy loss
function, computed as:
Loss( f ) = −
1
n
n
∑
i=1
y
i
· log(p
true
( f (T
i
, R
i
)))
where, y
i
is the true reply support label for the
tweet-reply pair and p
true
represents the predicted
probability for the true class y
i
.
3.1 Data Collection, Annotation and
Preprocessing
For the dataset, we focused on a significant geopolit-
ical event of the Fall of Kabul in the year 2021. The
tweets in this dataset are labelled with tweet labels as
Information (I), Emotion (E), Neutral (N), and Irrel-
evant (X). We curated a selection of labelled tweets
from this dataset with at least one reply, ensuring en-
gagement within the conversation. We did not con-
sider the Irrelevant tweets. We extracted only the En-
glish tweets along with their corresponding English
replies. This resulted in the final dataset comprising
of 495 tweets along with their replies. As a tweet can
have multiple replies, our final dataset is comprised of
2864 tweet-reply pairs. The resulting dataset has 200
emotion class tweets with 760 replies, 163 informa-
tion class tweets with 1180 replies, and 132 neutral
class tweets with 924 replies.
For each tweet-reply pair, annotations were done
by providing a reply support label as supporting,
unbiased, or opposing. The annotation process was
meticulously carried out by two independent human
annotators. To maintain inter-annotator reliability, the
annotated datasets were cross-checked by the annota-
tors for consistency and discrepancies were resolved
through mutual discussion. The labelling resulted in
Figure 1: Distribution of supporting, unbiased and oppos-
ing replies across the informative, emotional and neutral
tweets.
943 supporting instances, 1101 unbiased instances,
and 820 opposing instances as shown in Figure 1.
To prepare the data for analysis, we focused exclu-
sively on the textual content of the tweets and replies.
All emojis were converted into their corresponding
textual meanings and hashtags were split into their
component words. The dataset will be made publicly
available after publication of the paper.
3.2 Framework for Classification of
Reply Support Labels
For the classification of the reply support label, we
explored two frameworks, each designed to optimize
the accuracy of reply support classification.
3.2.1 Sentence/Word-Level LSTM Features for
Classification
In this framework, we utilized LSTM modules for
extraction of features. We employed two options:
sentence-level and word-level features. For both ap-
proaches, we generated embeddings for the tweet
and the reply using the GloVe technique which were
passed through two identical LSTM models as shown
in Figure 2a. Each model processed the entire text
sequence as a whole.
For sentence-level approach, the LSTM models
produced two distinct feature vectors, z
t
and z
r
, cor-
responding to the tweet and the reply. The vectors
were concatenated to form a combined vector, z
tr
, and
sent as input to a network of dense layers for the fi-
nal classification. This concatenation step allows the
final classification model to consider both the tweet
and the reply simultaneously.
For word-wise feature extraction approach, the
embeddings were passed through separate LSTM
models, which generated feature matrices rather than
single vectors (refer Figure 2a). The resulting feature
Stance Detection in Twitter Conversations Using Reply Support Classification
237

(a) LSTM-based classification.
(b) BERT-based classification.
Figure 2: Frameworks for classification of reply support labels. (a) Sentence-level and word-level feature extraction from
GLoVE embeddings of tweet and reply using LSTM. (b) Feature extraction from tweet and reply texts using BERT architec-
ture.
matrices, Z
t
and Z
r
for tweet and reply respectively.
Z
t
and Z
r
were concatenated to form a combined ma-
trix Z
tr
. Z
tr
was flattened and sent as input through
dense layers to classify the stance of the reply.
3.2.2 Attention-Based BERT for Classification
In this framework, we leverage the state-of-the-art
BERT (Devlin, 2018) model to perform both em-
bedding generation and classification. This approach
concatenated the tweet and reply text into a single se-
quence for generating embeddings and features in a
unified context. The BERT model outputs a single
embedding of vector size 768 which was then directly
fed into a classification layer to predict the reply sup-
port label.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
238

Table 1: Distribution of Classes in Training, Validation, and
Test Sets across emotional (E), informative (I), and neutral
(N) tweet classes.
Tweet Reply Support Label
Label −1 0 1
Training
E 116 205 307
I 206 464 259
N 334 212 188
Validation
E 13 19 33
I 25 71 26
N 44 20 35
Test
E 13 22 32
I 20 68 41
N 49 20 22
4 EXPERIMENTAL SETUP
We used 2864 tweet-reply pairs for our experiments
as explained in Section 3.1. These pairs are labelled
as supporting (+1), unbiased (0), or opposing (-1)
for the reply support label. The dataset was split
into training, validation, and test sets in the ratio of
80 : 10 : 10 using a stratified approach. Table 1 shows
the distribution of the tweet-reply pairs according to
their reply support labels and across the tweet label
classes. We used two frameworks for the classifica-
tion task as presented in Section 3.2. To evaluate how
well our models classify the reply support label, we
used Precision, Recall, F1-score, and Accuracy.
4.1 Setup for LSTM-Based
Classification
Two LSTM models, each with 64 units and a dropout
rate of 0.3 were used to process the tweet and the re-
ply in both sentence-level and word-level feature ex-
traction. For sentence-level features, the concatenated
feature vector output from the LSTMs served as an
input to a dense layer of 64 units activated by ReLU,
with a batch size of 8. A dropout of 0.2, batch nor-
malization and L2 regularization were applied.
For word-level features, the concatenated output
was sent to three dense layers composed of 2048,
512, and 64 units respectively with ReLU activation.
Batch normalization and a dropout rate of 0.2 was em-
ployed. In both cases, networks were trained using the
Adam optimizer with a learning rate of 1e − 4.
4.2 Setup for BERT-Based
Classification
The concatenated text served as input to the BERT
model for feature extraction. The output of the BERT
model was sent through a dense layer with 256 units
with ReLU activation. Batch normalization and L2
regularization and dropout were applied. A dense
layer with softmax activation was used for the fi-
nal classification. The model was trained for 150
epochs using the Adam optimizer with a learning rate
of 1e − 5 using sparse categorical cross-entropy loss
function.
5 RESULTS AND DISCUSSION
The results for the classification of reply support
label for the tweet-reply pairs are presented in Table
2.
5.1 Results for LSTM Models
The sentence-level features from LSTMs obtained an
overall accuracy of 64.1%. It achieved the highest
precision of 0.697 for the opposing class though the
recall for this class was low. For the unbiased class,
the model achieved a strong recall value of 0.736 and
an F1-score of 0.661. The model achieved the lowest
F1-score of 0.622 for the opposing class. The con-
fusion matrix for the framework is shown in Figure
3a. Few opposing and supporting replies have been
identified as unbiased replies.
For word-level feature extraction, the overall ac-
curacy was 65.2%. Individual class metrics revealed
that the model’s performance on all metrics was fairly
similar for all classes. The confusion matrix for the
framework is shown in Figure 3b. Few opposing
replies were falsely labelled as supporting or unbi-
ased. Some supporting replies were identified as un-
biased replies.
5.2 Results for BERT Model
BERT classification significantly outperformed
LSTM models. It achieved an overall accuracy of
73.2%. The opposing class showed a high recall of
0.793. There was improvement seen in the unbi-
ased class also The supportive class achieved high
precision at 0.767 and an F1-score of 0.729, demon-
strating relatively well-balanced classification for all
classes. The confusion matrix for this framework is
given in Figure 3c. It shows the robustness of the
Stance Detection in Twitter Conversations Using Reply Support Classification
239

Table 2: Performance Metrics for Sentence-level LSTM Features, Word-level LSTM Features and BERT Features for Classi-
fication of Reply Support Label.
Framework Support Label
Performance Metrics
Precision Recall F1-score Accuracy
Sentence-Level LSTM −1 0.697 0.561 0.622
0.641Features 0 0.600 0.736 0.661
1 0.663 0.600 0.630
Word-Level LSTM −1 0.675 0.634 0.654
0.652Features 0 0.640 0.646 0.643
1 0.647 0.674 0.660
BERT Embeddings −1 0.670 0.793 0.726
0.7320 0.760 0.718 0.738
1 0.767 0.695 0.729
(a) Sentence-level LSTM Features. (b) Word-level LSTM Features. (c) BERT Embeddings.
Figure 3: Confusion matrices for classification using: (a) Sentence-level features extracted from LSTM (b) Word-level features
extracted from LSTM (c) Embeddings from BERT.
Figure 4: Results of employing LIME on the BERT model for classification of a specific reply to a tweet. The weights
assigned to the words contribute to the decision making process of the model in classifying the reply support label.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
240

Table 3: Stance detection on varied events in current research. Our proposed model works on a more granular level by
determining stance of replies in conversation threads.
Author Objective Stance Dataset Models Results
(Mutlu et al., 2020)
Stance on neutral, COVID-CQ Logistic Accuracy:
treatment against, dataset of regression 0.76
for Covid-19 favor 14374 tweets
(Hamad et al., 2022)
Stance on online agree, StEduCov BERT Accuracy:
education during disagree, dataset of 68%
Covid-19 neutral 16, 572 tweets
(Zhang et al., 2024)
Stance on US pro-Biden, Self curated DoubleH Accuracy:
presidential pro-Trump dataset of 0.8579 ±0.04
elections 1, 123, 749 tweets
Proposed
Stance of supporting
Twitter dataset Sentence-Level F1-score:
Reply support label
reply in tweet unbiased
of event LSTM 0.637
conversations opposing
“Fall of Kabul”, Word-Level F1-score:
2864 tweet-reply LSTM 0.652
pairs BERT F1-score:
0.731
model in correctly classifying opposing and unbiased
responses with moderate misclassification between
unbiased and supportive replies. To summarize,
the results indicate that BERT provides the most
effective classification of the reply support labels,
which can play a crucial role in understanding social
interactions in crisis contexts.
5.3 Results for Explainable BERT
Using LIME
To provide insights into the model’s decision-making
process, we focus on enhancing the interpretability of
our BERT-based classification model by employing
LIME (Local Interpretable Model-agnostic Explana-
tions). Figure 4 demonstrates the results generated by
LIME for the following tweet-reply pair.
Tweet: This is unbelievable, potentially lethal
negligence on the part of the Morrison Government
towards Afghans who have stood with Australia
over a decade. I
ˆ
aC™ve been pleading with the
government to bring them to Australia for months.
Where the hell is Dutton.https://t.co/2OrI4HcfTU
Reply: https://theguardian.com/australia-
news/2020/nov/19/australian-special-forces-
involved-in-of-39-afghan-civilians-war-crimes-
report-alleges Australian elite soldiers recently killed
39 unarmed Afghanistan civilians. Probably more
went unreported. And you expect them to trust
Australia?!?
The visualization demonstrates that the model has
classified the reply as opposing with a high probabil-
ity of 0.89. The probabilities for unbiased and sup-
porting classes are very low, being 0.06 and 0.05 re-
spectively. This suggests that the model is quite con-
fident about its decision. LIME also highlighted that
words such as {trust, expect, negligence, soldiers}
highly influenced the classification of the reply stance
as opposing, being assigned weights of 0.21, 0.20,
0.17 and 0.14 respectively. Words such as {killed, sol-
diers} contributed towards the supporting class also
and {reported} was responsible for aiding the unbi-
ased class, but the overall influence is minimal.
5.4 Discussion of State-of-Art
Usually stance detection has been done in totality
for specific events/topics. Few of these works are
shown in Table 3. Our research differs from them
as the replies were not considered in those studies.
Our framework extends the work by analyzing replies
of tweets, deepening the granularity of online dis-
courses. (Mutlu et al., 2020) and (Hamad et al., 2022)
reported results for stance on singular tweet classi-
fications like COVID-19 treatment opinions or on-
line education. (Zhang et al., 2024) predicted the bi-
nary stance context on political tweets. Our contribu-
tion stands out by the fact that it focuses on tweet-
reply pairs, a rather unexplored field in stance de-
tection. This conversation analysis can deliver even
more subtle insights during crisis where engagement
in replies significantly shapes the discourse. Our best
framework makes use of a BERT-based framework
optimized for the reply support label classification
achieves an F1-score of 0.73, showcasing reasonable
performance in the multi-class setting.
Stance Detection in Twitter Conversations Using Reply Support Classification
241

6 CONCLUSIONS
This paper proposed a novel idea of determining the
reply support labels of tweets during crises to un-
derstand the conversation dynamics. By evaluating
LSTM and BERT-based models, we demonstrated the
reply stance classification in social media conversa-
tions, achieving an F1-score of 0.731, recall of 0.735,
and accuracy of 0.732. The interaction analysis of
tweets and replies is informative in terms of public en-
gagement compared to a sole focus only on the stance
of tweets. This knowledge can contribute towards im-
proved information management and offer actionable
insight for authorities to tailor their strategies in real-
time.
REFERENCES
Abulaish, M., Saraswat, A., and Fazil, M. (2023). A multi-
task learning framework using graph attention net-
work for user stance and rumor veracity prediction.
In Proceedings of the International Conference on Ad-
vances in Social Networks Analysis and Mining, pages
149–153.
ALDayel, A. and Magdy, W. (2021). Stance detection on
social media: State of the art and trends. Inf. Process.
Manage., 58(4).
Bai, N., Meng, F., Rui, X., and Wang, Z. (2023). A multi-
task attention tree neural net for stance classification
and rumor veracity detection. Applied Intelligence,
53(9):10715–10725.
Bukar, U. A., Jabar, M. A., Sidi, F., Nor, R. B., Abdullah,
S., and Ishak, I. (2022). How social media crisis re-
sponse and social interaction is helping people recover
from covid-19: an empirical investigation. Journal of
computational social science, pages 1–29.
Devlin, J. (2018). Bert: Pre-training of deep bidirec-
tional transformers for language understanding. arXiv
preprint arXiv:1810.04805.
Hamad, O., Hamdi, A., Hamdi, S., and Shaban, K. (2022).
Steducov: an explored and benchmarked dataset on
stance detection in tweets towards online education
during covid-19 pandemic. Big Data and Cognitive
Computing, 6(3):88.
Haouari, F. and Elsayed, T. (2024). Are authorities denying
or supporting? detecting stance of authorities towards
rumors in twitter. Social Network Analysis and Min-
ing, 14(1):34.
Hardalov, M., Arora, A., Nakov, P., and Augenstein,
I. (2021). A survey on stance detection for mis-
and disinformation identification. arXiv preprint
arXiv:2103.00242.
Hochreiter, S. (1997). Long short-term memory. Neural
Computation MIT-Press.
K
¨
uc¸
¨
uk, D. and Can, F. (2020). Stance detection: A survey.
ACM Computing Surveys (CSUR), 53(1):1–37.
Lai, M., Tambuscio, M., Patti, V., Ruffo, G., and Rosso,
P. (2019). Stance polarity in political debates: A di-
achronic perspective of network homophily and con-
versations on twitter. Data & Knowledge Engineering,
124:101738.
Mohammad, S., Kiritchenko, S., Sobhani, P., Zhu, X., and
Cherry, C. (2016). Semeval-2016 task 6: Detecting
stance in tweets. In Proceedings of the 10th inter-
national workshop on semantic evaluation (SemEval-
2016), pages 31–41.
Mutlu, E. C., Oghaz, T., Jasser, J., Tutunculer, E., Rajabi,
A., Tayebi, A., Ozmen, O., and Garibay, I. (2020). A
stance data set on polarized conversations on Twitter
about the efficacy of hydroxychloroquine as a treat-
ment for COVID-19. Data in brief, 33:106401.
Ng, L. H. X. and Carley, K. M. (2022). Pro or anti? a so-
cial influence model of online stance flipping. IEEE
Transactions on Network Science and Engineering,
10(1):3–19.
Villa-Cox, R., Kumar, S., Babcock, M., and Carley, K. M.
(2020). Stance in replies and quotes (srq): A new
dataset for learning stance in twitter conversations.
arXiv preprint arXiv:2006.00691.
Zeng, L., Starbird, K., and Spiro, E. (2016). # unconfirmed:
Classifying rumor stance in crisis-related social me-
dia messages. In Proceedings of the international
aaai conference on web and social media, volume 10,
pages 747–750.
Zhang, C., Zhou, Z., Peng, X., and Xu, K. (2024). Doubleh:
Twitter user stance detection via bipartite graph neural
networks. In Proceedings of the International AAAI
Conference on Web and Social Media, volume 18,
pages 1766–1778.
Zheng, J., Baheti, A., Naous, T., Xu, W., and Ritter, A.
(2022). Stanceosaurus: Classifying stance towards
multicultural misinformation. In Proceedings of the
2022 conference on empirical methods in natural lan-
guage processing, pages 2132–2151.
Zubiaga, A., Kochkina, E., Liakata, M., Procter, R., and
Lukasik, M. (2016). Stance classification in ru-
mours as a sequential task exploiting the tree struc-
ture of social media conversations. arXiv preprint
arXiv:1609.09028.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
242