Prompt Distillation for Emotion Analysis
Andrew L. Mackey, Susan Gauch and Israel Cuevas
Deparment of Electrical Engineering and Computer Science, University of Arkansas , Fayetteville, Arkansas, U.S.A.
Keywords:
Emotion Analysis, Natural Language Processing.
Abstract:
Emotion Analysis (EA) is a field of study closely aligned with sentiment analysis whereby a discrete set of
emotions are extracted from a given document. Existing methods of EA have traditionally explored both
lexicon and machine learning techniques for this task. Recent advancements in large language models have
achieved success in a wide range of tasks, including language, images, speech, and videos. In this work, we
construct a model that applies knowledge distillation techniques to extract information from a large language
model which instructs a lightweight student model to improve its performance with the EA task. Specifically,
the teacher model, which is much larger in terms of parameters and training inputs, performs an analysis
of the document and shares this information with the student model to predict the target emotions for a given
document. Experimental results demonstrate the efficacy of our proposed prompt-based knowledge distillation
approach for EA.
1 INTRODUCTION
Sentiment analysis (SA) is a prominent subfield of
natural language processing (NLP) with the goal of
analyzing text documents from which the document’s
polarity is obtained. Emotion analysis (EA) estab-
lishes additional granularity for classes beyond polar-
ity from SA by focusing on the alignment of language
with various emotional categories. For example, the
Paul Ekman model for emotions defines six primary
emotion categories: anger, disgust, fear, joy, sadness,
and surprise (Ekman and Friesen, 1971). Another
approach to illustrate the various emotional dimen-
sions was proposed as the Robert Plutchik model with
eight primary bipolar emotions: anger versus fear, joy
versus sadness, anticipation versus surprise, and trust
versus disgust (Plutchik, 1982). Additional models
have been proposed that projects emotions into a di-
mensional space, such as for valence, arousal, and
dominance (Russell and Mehrabian, 1977).
Various techniques have been proposed for the
task of emotion analysis. The first major area of emo-
tion analysis involves lexicon-based techniques where
the techniques are focused on aligning the emotional
categories of language with the specific words that
were used (Baccianella et al., 2010) (Staiano and
Guerini, 2014). The next major area of emotion anal-
ysis includes various machine learning techniques
that discover latent patterns or representations for the
detection of different emotional categories (Agrawal
and An, 2012) (Calefato et al., 2018) (Hasan et al.,
2019). Some researchers have investigated emotion
representations that seek to achieve emotion represen-
tations that transcend multiple lexicons and datasets
(Buechel et al., 2020). Some work in emotion clas-
sification has concentrated on aligning transformer-
based architectures with emotional categories through
deep contextual representations. Pretrained language
models (PLM) have demonstrated various successes
in outperforming many state-of-the-art techniques in
the field. As the parameters and training data contin-
ued to scale for PLMs, large language models (LLM)
emerged and demonstrated capabilities not seen in
prior work, such as prompt-based learning and rea-
soning.
In this paper, we introduce a prompt-based knowl-
edge distillation model for emotion analysis where the
prompt serves as source of knowledge through which
we distill that information for a student model un-
der the supervision of a much larger teacher model.
The first phase of our model involves a prompt-
based teacher model followed by a knowledge distil-
lation student training model. The teacher model uses
prompt-based techniques to extract information from
the LLM. The student model uses a transformer-based
PLM where probabilities from both teacher and stu-
dent models are aligned so that the student model is
capable of generating similar probability distributions
as the teacher model.
328
Mackey, A., Gauch, S. and Cuevas, I.
Prompt Distillation for Emotion Analysis.
DOI: 10.5220/0012951200003838
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 16th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2024) - Volume 1: KDIR, pages 328-334
ISBN: 978-989-758-716-0; ISSN: 2184-3228
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.

Figure 1: Overview architecture of the model. We combine a pre-trained language model with a large language model to
extract the emotion embeddings cross-corpus to perform a classification of the emotions. For the final prediction y, we
localize the classification head to a set of possible classes for the respective datatset.
2 RELATED WORK
Recent work in the research community has focused
on tasks involving emotion analysis has concentrated
primarily on PLMs for learning contextual represen-
tations using neural networks (Demszky et al., 2020)
(Turcan et al., 2021) (Alhuzali and Ananiadou, 2021)
(Wullach et al., 2021) (Mackey et al., 2021) (Tora-
man et al., 2022) (Rahman et al., 2024). PLMs un-
dergo various training methods which enables them to
learn latent contextual representations of text. These
models are generally fine-tuned in order to adapt to
task-specific objectives, such as emotion classifica-
tion. Bidirectional Encoder Representations from
Transformers (BERT) is a transformer-based archi-
tecture that bidirectionally encoded embeddings to
learn contextual information in textual data where the
model was pre-trained simultaneously on the tasks
of masked language modeling (MLM) and next sen-
tence prediction (NSP) (Devlin et al., 2019). XL-
Net improves upon BERT by introducing permuta-
tion language modeling where tokens are predicted
in a random order (Yang et al., 2019). RoBERTa
improved upon BERT by modifying the training ap-
proach where the NSP task was removed and dynamic
masking was introduced, and increasing the amount
of training data that was used (Liu et al., 2019).
Other work with LMs has resulted in different
techniques for training methodologies. Knowledge
distillation techniques, where a teacher model trans-
fers knowledge from a complex model to a much sim-
pler model, how shown promising results across dif-
ferent studies (Hinton et al., 2015) (Lukasik et al.,
2022). Brown et al. define various levels of data used
for in-context learning, such as fine-tuning (updating
weights of a pretrained model), few-shot (models are
provided a few demonstrations of a task with no addi-
tional weight updates to the model), one-shot (only
one demonstration is permitted), and zero-shot (no
demonstrations are permitted) (Brown et al., 2020).
Brown et al. also demonstrate that as LMs increase in
scale, their task-agnostic few-shot performance also
increases (Brown et al., 2020). In addition, Halder
et al. acknowledged that tranformer-based LMs fine-
tuned to task-specific objectives curtail their ability to
perform well in zero-shot, one-shot, or few-shot sce-
narios (Halder et al., 2020).
Work involving large language models continues
to demonstrate their task-agnostic capabilities. One
study demonstrated a technique of applying a series
of reasoning steps named chain of thought where
an LLM utilized chain-of-thought prompting that
demonstrated reasoning abilities provided the LLM is
adequately large (Wei et al., 2022). Adversarial dis-
tillation frameworks have also been proposed in re-
search literature for improved knowledge distillation
and transfer learning (Jiang et al., 2023).
3 PROBLEM DEFINITION
Let D represent a dataset comprised of N documents,
where each document in D consists of textual infor-
mation and emotion labels. We observe the following
for each D: (1) the set of text documents in dataset
D is represented as X
D
such that |X
D
| = N; (2) the set
of possible target labels for dataset D is represented
as Y
D
where |Y
D
| = C different emotions; and (3) D
is represented as the following set in the single-label
Prompt Distillation for Emotion Analysis
329

setting:
D = {(x, y) | x ∈ X
D
and y ∈ Y
D
} (1)
and the following serves as the representation for a
multi-label setting:
D = {(x, y) | x ∈ X
D
and y ∈ P (Y
D
)} (2)
Let D represent the input text corpora where D =
{D
1
, D
2
, ..., D
n
}. The task presented in this work is
to train and align a model to recognize the latent emo-
tion representations in a cross-corpus setting using D
for the purpose of single-class and multi-class emo-
tion classification of an emotion label (or set) y from
a given input document x:
ˆ
y = arg max
c

Pr(y = c | x; Θ)

(3)
4 PROPOSED APPROACH
We present our proposed solution in this section for
the single-class and multi-class cross-corpora emo-
tion classification task. In Figure 2, we provide
an overview of our framework for learning the la-
tent emotion distribution of text documents. There
are three major components to our approach: (1) a
prompt-based knowledge distillation paradigm for ex-
tracting information from an LLM to facilitate the
alignment of a task-specific model; (2) a task-specific,
emotion classification model that leverages a pre-
trained, transformer-based language model, which is
fine-tuned for the emotion classification task; and (3)
a cross-corpora framework for learning latent emotion
representations.
4.1 Prompt-Based Methodology
For a given dataset D = (X
D
, Y
D
), each input
and target is represented as (x
doc
, y
emo
) such that
(x
doc
, y
emo
) ∈ D. The target y
emo
of the model is the
emotion class for each document where y
emo
∈ Y
D
(i.e. anger, grief, disgust, etc.) in the respective
dataset D. To facilitate knowledge distillation from
an LLM, we define (x
prompt
, y
llm
) to represent the
prompt-based input and label generated from an LLM
for each (x
doc
, y
emo
) ∈ D.
Prompt Template. The following template is used
to the generate each x
prompt
:
You will be given a human written sentence. Classify
the sentence into one of the following categories:
⟨y
emo
0
, y
emo
1
, ...⟩. Return the following format only for
each category as a probability distribution (the sum
should be 1): ⟨y
emo
i
, probability⟩.
The following is the document: x
i
.
The target y
llm
represents the emotion distribution
produced by the LLM for the given input prompt
x
prompt
, which is modeled as follows:
ˆ
y
llm
i
= Pr(y
emo
| x
prompt
i
) (4)
= LLM(x
prompt
) (5)
Hallucinations are a known problem in research
literature where an LLM produces a response that is
either factually incorrect or unaligned with the input
prompt it was provided (Farquhar et al., 2024). To ad-
dress the problem of hallucinations, we conduct a val-
idation step for
ˆ
y
llm
to ensure the format of the output
is aligned with the targets in the training data. Docu-
ments failing the validation step will undergo a fixed
interval of reprompting where the input and interac-
tions are returned to the LLM for further processing
in the form:
ˆ
y
llm
′
= LLM( ⟨x
prompt
′
, ⟨x
prompt
, y
llm
⟩⟩ ) (6)
4.2 Emotion Classification Model
The task-specific emotion classification model begins
by employing the use of a transformer-based language
model to provide contextual representations h
emo
=
⟨h
emo
1
, h
emo
2
, ..., h
emo
k
⟩ for input tokens x
doc
where k
represents the number of time steps. The transformer-
based encoder LM is parameterized with φ for all
datasets D ∈ D to generate the contextualized word
representations h
emo
i
for each time step i:
h
emo
i
= LM
φ
(x
doc
i
) (7)
The last layer of h
emo
i
is used to compute the dis-
tribution for the emotion classes, where it is param-
eterized by φ
d
for each D
d
∈ D to obtain the target
prediction distribution
ˆ
y
emo
i
and the softmax layer is
applied to normalize the logits:
ˆ
y
emo
i
= Pr(y
emo
i
| h
emo
i
) (8)
= Softmax(W
φ
d
h
emo
i
+ b
φ
d
) (9)
The model shares a common set of parameters φ
between all members of D to facilitate latent emo-
tion representation learning in a cross-domain envi-
ronment, while the task-specific classification head
maintains a specific set of a parameters φ
d
.
4.3 Knowledge Distillation
The goal of a prompt-based teacher model is to extract
knowledge from an LLM and transfer it to the task-
specific student model, which is responsible for fine-
grained emotion classification. The prompt-based
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
330

model instructs the emotion classification model to
enable the smaller model to generalize in a manner
that resembles the teacher model. The student model
minimizes a loss function which focuses on both cor-
rectly predicting the target label y
emo
while simulta-
neously aligning the model with the teacher model’s
responses y
llm
.
The model utilizes a cross-entropy loss function
for the single-class emotion classification task
L
CE
= −
1
N
C
∑
i
y
emo
i
log( ˆy
emo
i
) (10)
and a binary cross-entropy loss function for multi-
class emotion classification
L
BCE
= −
1
N
C
∑
i

y
emo
i
log( ˆy
emo
i
) + (1 − y
emo
i
)log(1 − ˆy
emo
i
)

(11)
for when there exists multiple emotion labels for a
given document.
We use τ to represent the temperature rate hyper-
parameter to produce a softer probability distribution
over all possible classes for class imbalances through
knowledge distillation techniques. For these models,
the losses from the emotion detection task and the
prompt-based alignment model are summed together
after each batch by using the adjustable hyperparam-
eter λ, which balances the terms below:
L
φ
= λL
emo
+ (1 − λ)τ
2
L
llm
(12)
5 EXPERIMENTS
In this section, we provide an empirical analysis of
our proposed model and investigate the following re-
search questions:
• RQ1: What is the effectiveness of the proposed
model for the emotion classification task in terms
of model performance metrics?
• RQ2: Does the choice of LM contribute to the
performance of the proposed model?
• RQ3: How does the knowledge distillation from
an LLM to the proposed model contribute to the
overall performance?
5.1 Data
Our experiments are conducted on two benchmark
datasets: WASSA-21 dataset and Real World Worry
dataset (Buechel et al., 2018) (Kleinberg et al.,
2020). The WASSA-21 dataset was provided in
the 11th Workshop on Computational Approaches
Figure 2: Distribution of the emotion labels by dataset. The
RWW dataset emphasized fear and sadness labels. The
GoEmotions dataset had a stronger presence of documents
labeled as neutral and joy. The WASSA dataset contained
more labels with the sadness and surprise labels in compar-
ison to other datasets.
to Subjectivity, Sentiment, and Social Media Analy-
sis (WASSA) Shared Task: Empathy Detection and
Emotion Classification (Tafreshi et al., 2021). The
dataset consists of n = 1860 reactions to news stories
indicating that there is harm to a person, group, or
other. The labels for each record are mapped to seven
emotion categories, which include a neutral category
and Ekman’s basic emotion categories: anger, dis-
gust, fear, joy, sadness, and surprise. This label rep-
resents the dominant emotion for the text.
Table 1: GoEmotions emotion mapping to Ekman emo-
tions.
Emotion Association
anger anger, annoyance, disapproval
disgust disgust
fear fear, nervousness
joy joy, amusement, approval, ex-
citement, gratitude, love, opti-
mism, relief, pride, admiration,
desire, caring
sadness sadness, disappointment, em-
barrassment, grief, remorse
surprise surprise, realization, confusion,
curiosity
The second dataset used in our experiments is
the COVID-19 Real World Worry dataset (Kleinberg
Prompt Distillation for Emotion Analysis
331

Table 2: COVID-19 emotion mapping to Ekman emotions.
Emotion Association
anger anger
disgust disgust
fear fear, anxiety
joy happiness, relaxation
sadness sadness
surprise desire
et al., 2020). The dataset contains n = 2491 records
that were extracted by surveying participants and ask
them to express their emotional feelings towards the
COVID-19 pandemic. Participants were asked to con-
struct two different forms of text. The first document
they were asked to author included instructions to ex-
press their feelings towards the then current COVID-
19 situation with a minimum of 500 characters. The
second document expressed them to convey the same
feelings in the form of a social media post that had a
maximum of 240 characters. Participants were asked
to rate their emotions toward the situation and select
one of the following emotions that best represented
their feelings: anger, anxiety, desire, disgust, fear,
happiness, relaxation, and sadness. We used the emo-
tion definitions from (Demszky et al., 2020) as indi-
cated in Table 1 to map perform the emotion map-
pings as indicated in Table 2.
5.2 Baseline Experiments
To evaluate the efficacy of our proposed prompt-based
knowledge distillation model, we use PLMs as the
baseline for our experiments. We benchmark our
model using the BERT, RoBERTa, and XLNet PLMs
where the input will only be the document and tar-
get emotion(s). We evaluate the model performance
of each dataset and report the mean precision, recall,
and F
1
-scores after 3 runs using macro averaging.
5.3 Experimental Settings
Our model was constructed using the PyTorch frame-
work along with the HuggingFace transformers li-
brary for the pretrained language model implementa-
tions.
1
We followed similar experimental settings as
provided in (Demszky et al., 2020). Our model uses
the AdamW optimizer (Loshchilov and Hutter, 2017)
while setting the learning rate to 5e
−5
, batch size to
16, and maximum sequence length of 512. Since
previous research literature demonstrated overfitting
beyond four epochs, we limited our the number of
epochs during the fine-tuning step to four (Demszky
1
https://huggingface.co/docs/transformers/en/index
et al., 2020). For the large language model, we uti-
lized the GPT-4o model provided through the API.
5.4 Experimental Results
Table 3 reflects the results from the experiments con-
ducted in this paper. Each experiment was executed
independently of other datasets. The best results are
indicated in bold. As reflected in the results, our
method is able to demonstrate increased performance
above the baseline methods for the WASSA-21 and
RWW datasets. This demonstrates that the PLM ac-
quires additional knowledge through transfer learn-
ing and knowledge distillation through this technique
that it did not acquire through the data alone. Fur-
thermore, we also discover that the RoBERTa PLM
is able to achieve superior performance over the other
PLMs evaluated in the tests we conducted. Despite
the extreme differences in the distribution of the la-
bels between the datasets as evidenced in Figure 2, we
observe that the proposed technique is able to work
given the task-agnostic knowledge provided from the
teacher model. When RoBERTa was used as the
underlying PLM for our technique, we were able
to achieve a gain of ∆ = +2.18 increase in perfor-
mance for the F
1
score for the WASSA-21 dataset and
∆ = +1.86 for the RWW dataset.
It should also be noted that the largest gain in
performance was achieved through the prompt-based
knowledge distillation approach with the BERT PLM
in the RWW dataset. We observe an increase of
∆ = +2.37 in the F
1
score under these settings.
6 CONCLUSIONS
Throughout our work in this paper, we investigated
the task of emotion analysis under a prompt-based
knowledge distillation setting where we trained a stu-
dent model by aligning it with a teacher model which
provides instruction on how to generate similar proba-
bility distributions in a task-specific objective. Future
directions for this work can involve exploring other
techniques, such as chain-of-thought or other reason-
ing approaches, or augmented LLM approaches to
improve the teacher model through prompting strate-
gies. The proposed methodology can be extended to
consider additional modalities of data.
REFERENCES
Agrawal, A. and An, A. (2012). Unsupervised emotion de-
tection from text using semantic and syntactic rela-
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
332

Table 3: Comparison of baselines with experimental settings. Our proposed prompt-based knowledge distillation models
outperform the baseline models.
Type Model
WASSA-21 RWW
Precision Recall F1 Precision Recall F1
Baseline
BERT 68.52 68.67 67.70 18.81 19.33 18.80
RoBERTa 72.39 73.84 71.74 23.20 21.97 20.61
XLNet 60.74 63.18 60.92 20.40 20.26 18.52
Experiment
BERT+PKD 69.02 71.15 68.58 23.52 23.32 21.17
RoBERTa+PKD 73.85 75.16 73.92 24.63 22.69 22.47
XLNet+PKD 62.55 64.29 61.75 22.52 21.41 19.28
∆ Change +1.46 +1.32 +2.18 +1.43 +0.72 +1.86
tions. In 2012 IEEE/WIC/ACM International Confer-
ences on Web Intelligence and Intelligent Agent Tech-
nology, volume 1, pages 346–353.
Alhuzali, H. and Ananiadou, S. (2021). SpanEmo: Casting
multi-label emotion classification as span-prediction.
In Merlo, P., Tiedemann, J., and Tsarfaty, R., edi-
tors, Proceedings of the 16th Conference of the Eu-
ropean Chapter of the Association for Computational
Linguistics: Main Volume, pages 1573–1584, Online.
Association for Computational Linguistics.
Baccianella, S., Esuli, A., and Sebastiani, F. (2010). Senti-
WordNet 3.0: An enhanced lexical resource for senti-
ment analysis and opinion mining. In Calzolari, N.,
Choukri, K., Maegaard, B., Mariani, J., Odijk, J.,
Piperidis, S., Rosner, M., and Tapias, D., editors, Pro-
ceedings of the Seventh International Conference on
Language Resources and Evaluation (LREC’10), Val-
letta, Malta. European Language Resources Associa-
tion (ELRA).
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.,
Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G.,
Askell, A., Agarwal, S., Herbert-Voss, A., Krueger,
G., Henighan, T., Child, R., Ramesh, A., Ziegler,
D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler,
E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner,
C., McCandlish, S., Radford, A., Sutskever, I., and
Amodei, D. (2020). Language models are few-shot
learners.
Buechel, S., Buffone, A., Slaff, B., Ungar, L., and Sedoc, J.
(2018). Modeling empathy and distress in reaction to
news stories. In Proceedings of the 2018 Conference
on Empirical Methods in Natural Language Process-
ing. Association for Computational Linguistics.
Buechel, S., Modersohn, L., and Hahn, U. (2020). Towards
label-agnostic emotion embeddings.
Calefato, F., Lanubile, F., and Novielli, N. (2018). Emotxt:
A toolkit for emotion recognition from text.
Demszky, D., Movshovitz-Attias, D., Ko, J., Cowen, A.,
Nemade, G., and Ravi, S. (2020). Goemotions: A
dataset of fine-grained emotions. In Proceedings of
the 58th Annual Meeting of the Association for Com-
putational Linguistics. Association for Computational
Linguistics.
Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.
(2019). Bert: Pre-training of deep bidirectional trans-
formers for language understanding. In Proceedings
of the 2019 Conference of the North. Association for
Computational Linguistics.
Ekman, P. and Friesen, W. V. (1971). Constants across cul-
tures in the face and emotion. Journal of personality
and social psychology, 17(2):124.
Farquhar, S., Kossen, J., Kuhn, L., and Gal, Y. (2024). De-
tecting hallucinations in large language models using
semantic entropy. Nature, 630(8017):625–630.
Halder, K., Akbik, A., Krapac, J., and Vollgraf, R. (2020).
Task-aware representation of sentences for generic
text classification. In Scott, D., Bel, N., and Zong, C.,
editors, Proceedings of the 28th International Confer-
ence on Computational Linguistics, pages 3202–3213,
Barcelona, Spain (Online). International Committee
on Computational Linguistics.
Hasan, M., Rundensteiner, E., and Agu, E. (2019). Auto-
matic emotion detection in text streams by analyzing
twitter data. International Journal of Data Science
and Analytics, 7.
Hinton, G., Vinyals, O., and Dean, J. (2015). Distilling
the knowledge in a neural network. In NIPS Deep
Learning and Representation Learning Workshop.
Jiang, Y., Chan, C., Chen, M., and Wang, W. (2023). Lion:
Adversarial distillation of proprietary large language
models. In The 2023 Conference on Empirical Meth-
ods in Natural Language Processing.
Kleinberg, B., van der Vegt, I., and Mozes, M. (2020).
Measuring emotions in the covid-19 real world worry
dataset.
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D.,
Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov,
V. (2019). Roberta: A robustly optimized bert pre-
training approach.
Loshchilov, I. and Hutter, F. (2017). Decoupled weight
decay regularization. In International Conference on
Learning Representations.
Lukasik, M., Bhojanapalli, S., Menon, A. K., and Kumar, S.
(2022). Teacher’s pet: understanding and mitigating
biases in distillation. Transactions on Machine Learn-
ing Research.
Prompt Distillation for Emotion Analysis
333

Mackey, A., Gauch, S., and Labille, K. (2021). Detecting
fake news through emotion analysis.
Plutchik, R. (1982). A psychoevolutionary theory of emo-
tions.
Rahman, A. B. S., Ta, H.-T., Najjar, L., Azadmanesh,
A., and G
¨
on
¨
ul, A. S. (2024). Depressionemo: A
novel dataset for multilabel classification of depres-
sion emotions.
Russell, J. A. and Mehrabian, A. (1977). Evidence for a
three-factor theory of emotions. Journal of research
in Personality, 11(3):273–294.
Staiano, J. and Guerini, M. (2014). Depeche mood: a lexi-
con for emotion analysis from crowd annotated news.
In Proceedings of the 52nd Annual Meeting of the As-
sociation for Computational Linguistics (Volume 2:
Short Papers), volume 2, pages 427–433.
Tafreshi, S., De Clercq, O., Barriere, V., Buechel, S., Sedoc,
J., and Balahur, A. (2021). WASSA 2021 shared task:
Predicting empathy and emotion in reaction to news
stories. In De Clercq, O., Balahur, A., Sedoc, J., Bar-
riere, V., Tafreshi, S., Buechel, S., and Hoste, V., ed-
itors, Proceedings of the Eleventh Workshop on Com-
putational Approaches to Subjectivity, Sentiment and
Social Media Analysis, pages 92–104, Online. Asso-
ciation for Computational Linguistics.
Toraman, C., S¸ ahinuc¸, F., and Yilmaz, E. (2022). Large-
scale hate speech detection with cross-domain trans-
fer. In Calzolari, N., B
´
echet, F., Blache, P., Choukri,
K., Cieri, C., Declerck, T., Goggi, S., Isahara, H.,
Maegaard, B., Mariani, J., Mazo, H., Odijk, J.,
and Piperidis, S., editors, Proceedings of the Thir-
teenth Language Resources and Evaluation Confer-
ence, pages 2215–2225, Marseille, France. European
Language Resources Association.
Turcan, E., Muresan, S., and McKeown, K. (2021).
Emotion-infused models for explainable psychologi-
cal stress detection. In Proceedings of the 2021 Con-
ference of the North American Chapter of the Associa-
tion for Computational Linguistics: Human Language
Technologies. Association for Computational Linguis-
tics.
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B.,
Xia, F., Chi, E., Le, Q., and Zhou, D. (2022). Chain-
of-thought prompting elicits reasoning in large lan-
guage models.
Wullach, T., Adler, A., and Minkov, E. (2021). Fight fire
with fire: Fine-tuning hate detectors using large sam-
ples of generated hate speech. In Findings of the Asso-
ciation for Computational Linguistics: EMNLP 2021.
Association for Computational Linguistics.
Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov,
R. R., and Le, Q. V. (2019). Xlnet: Generalized
autoregressive pretraining for language understand-
ing. In Wallach, H., Larochelle, H., Beygelzimer,
A., d'Alch
´
e-Buc, F., Fox, E., and Garnett, R., editors,
Advances in Neural Information Processing Systems,
volume 32. Curran Associates, Inc.
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
334