=Paper=
{{Paper
|id=Vol-3740/paper-114
|storemode=property
|title=Penta-nlp at EXIST 2024 Task 1-3: Sexism Identification, Source Intention, Sexism
Categorization In Tweets
|pdfUrl=https://ceur-ws.org/Vol-3740/paper-114.pdf
|volume=Vol-3740
|authors=Fariha Tanjim Shifat,Fabiha Haider,Md Sakib Ul Rahman Sourove,Deeparghya Dutta Barua,Md Farhan Ishmam,Md Fahim,Farhad Alam Bhuiyan
|dblpUrl=https://dblp.org/rec/conf/clef/ShifatHSBIFB24
}}
==Penta-nlp at EXIST 2024 Task 1-3: Sexism Identification, Source Intention, Sexism
Categorization In Tweets==
https://ceur-ws.org/Vol-3740/paper-114.pdf
Penta-nlp at EXIST 2024 Task 1-3: Sexism Identification,
Source Intention, Sexism Categorization In Tweets
Notebook for the EXIST Lab at CLEF 2024
Fariha Tanjim Shifat1,*,† , Fabiha Haider1,*,† , Md Sakib Ul Rahman Sourove1 ,
Deeparghya Dutta Barua1 , Md Farhan Ishmam1,3 , Md Fahim1,2,* and Farhad Alam Bhuiyan1,*
1
Research and Development, Penta Global Limited, Bangladesh
2
CCDS Lab, IUB, Bangladesh
3
Islamic University of Technology, Bangladesh
Abstract
Social media platforms contains a vast user base and offers ease with which information can be shared. This can
adversely facilitate the spread of sexist content which is infeasible for human monitoring and filtering. This paper
investigates the automated detection of sexism in tweets using Natural Language Processing (NLP) techniques.
Sexist tweets can create a hostile online environment and perpetuate harmful stereotypes. Manual identification
is impractical due to the vast amount of data. The research proposes a system utilizing machine learning models
to analyze text, identify bias and discriminatory language patterns, and flag tweets for moderation. The fourth
edition of EXIST shared task 2024 Tweets Dataset, containing labeled English and Spanish tweets, is used to train
and evaluate the models. The system explores various approaches, including TF-IDF with different classifiers
(SVM, XGB, RF), Long Short-Term Memory (LSTM) networks with and without attention mechanisms, and
pre-trained transformer models (XLM-Roberta, mBERT, BETO). The effectiveness of different preprocessing
techniques and the role of attention weights in identifying sexism are also explored. The paper outlines the
methodology, experimental setup, and analysis of results, paving the way for further discussion on error analysis
and conclusions in subsequent sections.
Keywords
Sexism identification, Tweets, Source intention detection, Sexism categorization, Multilingual Models, Natural
Language Processing
1. Introduction
This research paper discusses topics related to specific content that may be sensitive or offensive, which
some readers may find distressing. The intent is to analyze and understand the research work.
The commence and growth of internet have a profound impact on our social structure, the way we
communicate, our relationships through the development of various social platforms. Twitter, one
of such social platforms, has developed into a lively forum for sharing idea and discourse because of
its succinct format and emphasis on real-time updates, with attractive features like hashtags, tags,
etc. While such advancements of social platforms promotes connectivity and facilitates the spread of
information it also tempts people to gain fame thorough views and likes, and throw inappropriate
contents and comments in disguise of freedom of speech lacking empathy towards race, gender, religion
[1]. Evidently sexism exists in Twitter in the form of sexist tweets sometimes intentionally while at
other times unintentionally. These tweets and contents can range from blatant objectification and
insults to more implicit bias and prejudice. Sexist tweets can create a hostile environment online,
especially for women and other targeted groups. Identifying sexism in online platforms is crucial
for a various number of reasons. Recognizing sexist content is essential to advancing equality and
averting social harm. The propagation of detrimental stereotypes and biases through such information
CLEF 2024: Conference and Labs of the Evaluation Forum, September 09–12, 2024, Grenoble, France
*
Corresponding authors.
†
These authors contributed equally.
$ fariha.tanjim.shifat@gmail.com (F. T. Shifat); fabihahaider4@gmail.com (F. Haider); fahimcse381@gmail.com (M. Fahim)
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
contributes to gender inequality and has a detrimental effect on people’s mental health and self-esteem
[2]. We can promote an inclusive and respectful culture by addressing sexist content, making sure that
everyone feels respected and secure.
The sheer volume of such sexist tweets makes it difficult to identify sexism manually, requiring
the immediate need for automated solutions. Natural Language Processing (NLP) can emerge as a
successful tool in recognizing such harmful contents and filtering them out [3]. It can prevent people to
post contents that goes against community standards creating a safer and more acceptable platform.
Using natural language processing (NLP) techniques, we can create automated systems that can
recognize sexist tweets with accuracy. Text content can be analyzed by these technologies, which can
also spot bias and discriminatory language patterns and flag tweets for moderation or additional review.
EXIST aims to capture sexism in a broad sense, from explicit misogyny to other subtle expressions
that involve implicit sexist behaviours. EXIST is a series of scientific events and shared tasks on sexism
identification in social networks [4], [5]. The EXIST 2024 Tweets Dataset contains more than 10,000
labeled tweets, both in English and Spanish. Based on the labeled data, the tasks were to identify sexist
tweets among them, the underlying intention of the author, and if it contained sexism at multiple
degrees. The challenge lies in the nuanced and context-dependent nature of language on social media.
Tweets often use slang, sarcasm, or coded language that can obscure sexist intent, while URLs can
lead to external content that may contain sexist material not immediately evident in the tweet itself.
Mentions and hashtags can complicate analysis by linking tweets to broader conversations or by being
used to target specific individuals. Emojis add another layer of complexity, as their meanings can vary
widely depending on context and cultural interpretation. These factors make automated detection
systems prone to errors, requiring sophisticated algorithms and often human oversight to accurately
identify and address sexism in tweets.
Our approach involved training different Machine Learning (ML) models to capture the sexist pattern
in the contents and choose the one that performs the best on the validation dataset. The machine learning
models included TF-IDF+SVM [6], TF-IDF+XGB [7], TF-IDF+RF [8], LSTM [9], LSTM+Attention [10],
XLM-Roberta [11], mBert [12], BETO [13]. We yielded results for different ways of preprocessing the
data and further finetuned the models based on best results. Additionally, we attempted to find out the
provoking words that contributed in the semantic meaning of sexism reflected through the attention
weights of those words. Taking into accounts the attention weights of the sentence representation
helped the same model to yield the best result for Task 1, Task 2 and Task 3. The results show that the
Bert based models are well trained to capture the pattern of sexism when considering the attention
layer. The performance metrics used were accuracy and F1 score and we could obtain an accuracy
of 84.80%, 72.06% and 88.25% and F1 score of 84.76%, 51.43% and 54.77% for Task 1, Task 2 and Task
3, respectively. To do better modeling and include different explainability results, we adopt different
training and explainable experiments from EDAL, ITPT, HateXplain [14, 15, 16] papers. In the remainder
of this paper, we present the Problem Description in Section 2, Methodology in Section 3, Experimental
Setup in Section 4, Result Analysis in Section 5, Error Analysis in Section 6 and Conclusion in Section 7.
2. Problem Description
2.1. Task Descriptions
In this work, we have addressed tasks 1, 2, and 3. The task descriptions are listed as per the guidelines.
1. Task 1 (Sexism Identification in Tweets): Given a tweet, this subtask aimed to classify whether
a tweet contains sexist expressions or behaviors. The tweet can be sexist itself, describe a sexist
situation or critic sexist behavior. It is a binary classification problem. We needed to label ’YES’
or ’NO’ to the tweets.
2. Task 2 (Source Intention in Tweets): This subtask aims to classify each tweet according to the
intention of the person who wrote it. This is a multi-class classification problem. The classes are
as follows:
• DIRECT: The intention is to write a message that is sexist itself.
• REPORTED: The author intends to report or describe a sexist situation or event suffered
by a woman or women in the first or third person.
• JUDGEMENTAL: The author intends to be judgemental since the tweet describes sexist
situations or behaviors to condemn them.
• NO: The tweet is detected as not sexist in subtask 1.
3. Task 3: This subtask categorizes the tweets according to the type of sexism. This is a multi-label
classification problem with 5 labels. So more than one class can be assigned to each tweet. The
labels are as follows:
• IDEOLOGICAL-INEQUALITY: The tweet discredits the feminist movement, rejects in-
equality between men and women, or presents men as victims of gender-based oppression.
• STEREOTYPING-DOMINANCE: The tweet expresses false ideas about women that
suggest they are more suitable to fulfill certain roles (mother, wife, family caregiver, faithful,
tender, loving, submissive, etc.), or inappropriate for certain tasks (driving, hard work, etc.),
or claims that men are somehow superior to women.
• OBJECTIFICATION: The tweet presents women as objects apart from their dignity and
personal aspects or assumes or describes certain physical qualities that women must have
to fulfill traditional gender roles (compliance with beauty standards, hypersexualization of
female attributes, women’s bodies at the disposal of men, etc.).
• SEXUAL-VIOLENCE: The tweet includes or describes sexual suggestions, requests for
sexual favors, or harassment of a sexual nature (rape or sexual assault).
• MISOGYNY-NON-SEXUAL-VIOLENCE: The tweet expresses hatred and violence towards
women, different from that with sexual connotations.
• NO: When none of the 5 labels are assigned to the tweet.
2.2. Dataset Statistics
The dataset includes over 10,000 tweets both in Spanish and English. The train, dev, and test sets contain
6064, 934, and 2076 tweets respectively. These numbers are after discarding the non-labeled samples in
the gold standard. The distribution between both languages has been somewhat balanced to tackle the
issue of biases. The exact figures are in Table 1.
Table 1
Tweets Dataset, containing training, development and test splits. The dataset contained both English and Spanish
language.
Language Wise Samples
Data Splits Total Samples
EN ES
Train 6064 2870 3194
Dev 934 444 490
Test 2076 978 1098
Task 1 is a binary classification problem with ’YES’ and ’NO’ labels. Task 2 is a multi-class classification
problem with 3 sexism sources and one ’NO’ label. Task 3 is a multi-label classification problem with 5
different labels and a ’NO’ label when none of the labels are assigned. The distribution is not balanced.
Each tweet is labeled by six different annotators. The gold label is the average of the labels. The exact
figures of the distribution is in Table 2.
Table 2
Class or Label-wise distribution in the dataset for Task 1, Task 2 and Task 3.
Samples
Class/Labels
Train Dev
Task 1
NO 3376 479
YES 2697 455
Task 2
NO 3367 479
DIRECT 1294 204
REPORTED 459 83
JUDGEMENTAL 376 75
Task 3
IDEOLOGICAL-INEQUALITY 1113 212
STEREOTYPING-DOMINANCE 1423 241
OBJECTIFICATION 1103 183
SEXUAL-VIOLENCE 675 123
MISOGYNY-NON-SEXUAL-VIOLENCE 856 158
NO 3367 479
hback
back
Multilingual Tokenizer
Multilingual Model
Attention
I wanna go back
hgo
Aggregation
Pre-trained
Pre-trained
Classifier
go
Logits
hwanna
wanna
hI
MLP
h[CLS]
I
Figure 1: Model Architecture. Single sentence is passed through the tokenizer to separate the tokens.
They are then passed to the model which gives the encoded format of those tokens. Encoded tokens
except CLS are passed to the attention module whereas CLS is passed to the MLP layer which then is
aggregated using various methods and then this representation is used for classification. The logits are
yield by the classifier.
3. Methodology
In this section, we will outline our methodology. Given the multilingual nature of the dataset encom-
passing both English and Spanish texts, we employ a multilingual pretrained model. We further refine
this model through fine-tuning on the dataset. Our model architecture comprises five key components:
i) Pretrained model backbone, ii) Sentence Representation from CLS token, iii) Attention-based Context
Vector, iv) Feature Aggregation Module, and v) Classifier Head
3.1. Pretrained Multilingual Model as Backbone
An input sentence 𝑆 is passed into the pretrained multilingual tokenizer to obtain the tokens of the
sentence 𝑆 = {𝑡[𝐶𝐿𝑆] , 𝑡1 , 𝑡2 , . . . , 𝑡𝑛 , 𝑡[𝑆𝐸𝑃 ] }, where 𝑡𝑖 represents the 𝑖-th token and 𝑡[𝐶𝐿𝑆] & 𝑡[𝑆𝐸𝑃 ] are
the special tokens. The tokens are passed into the pretrained multilingual model to achieve contextual
representations for each token 𝑡𝑖 , denoted as 𝐻 = {ℎ[𝐶𝐿𝑆] , ℎ1 , ℎ2 , . . . , ℎ𝑛 , ℎ[𝑆𝐸𝑃 ] }, where ℎ𝑖 represents
the contextual representation of token 𝑡𝑖 . Specifically, we extract the last layer hidden representations
of the pretrained model, which are further fine-tuned during the dataset training.
3.2. Sentence Representation from CLS Token
To get the representation for the entire sentence, we use the representation of [CLS] token ℎ[𝐶𝐿𝑆] . This
representation is passed into a Single Layer Perception to get the enhanced representation.
′
ℎ𝐶𝐿𝑆 = 𝑊𝐶𝐿𝑆 · ℎCLS + 𝑏𝐶𝐿𝑆
3.3. Attention-based Context Vector
As the tasks are natural language understanding tasks where individual words hold distinct predictive
significance, we integrate an attention-based network to ascertain word importance. By accounting for
the significance of each word, we construct a context vector for the sentence. We only consider the
representations of the words and exclude the representations of the special tokens.
Once contextual representations 𝐻 are obtained for a sentence 𝑆, an additional attention layer is
added to compute learnable attention scores 𝛼𝑖 for each token 𝑡𝑖 in 𝐻, and its calculation is as follows:
𝛼𝑖 = softmax(𝑊 · ℎ𝑖 + 𝑏),
𝑖 = 1, 2, . . . , 𝑛
This results in a set of attention_scores = {𝛼1 , 𝛼2 , . . . , 𝛼𝑛 } corresponding to the tokens in sentence
𝑆. These attention scores collectively represent the overall attention distribution across the sentence,
indicating the relative importance or relevance of each token to the context of the entire sentence. After
finding attention scores for each token, we find the context vector for the sentence 𝑆 by multiplying
the contextual representations of token 𝑡𝑖 with its attention score 𝛼𝑖 .
𝑛
∑︁
𝑐= 𝛼𝑖 · ℎ𝑖
𝑖=1
3.4. Feature Aggregation Module
In this section, we consolidate various representations of the sentence 𝑆. We obtain two distinct
′
representations: one based on the CLS token ℎ𝐶𝐿𝑆 and the other based on attention-based context
vector 𝑐. The aggregation of these representations is performed using two distinct techniques considering
the two following techniques:
′
• Concat-based Aggregation: For concat-based aggregation, we just simply concatenate ℎ𝐶𝐿𝑆
and 𝑐 to get the aggregated representation 𝑧 as follows:
′
𝑧_concat = concat[ℎ𝐶𝐿𝑆 , 𝑐]
Then 𝑧_concat is passed into single layer MLP layer to get the final combined representation 𝑧
where 𝑧 = MLP(𝑧_concat)
′
• Element wise Addition based Aggregation: In this aggregation, we combine ℎ𝐶𝐿𝑆 and 𝑐 by
summing them element wise as follows
′
𝑧𝑠𝑢𝑢𝑚𝑒𝑑 = ℎ𝐶𝐿𝑆 + 𝑐
Then 𝑧_suumed is passed into single layer MLP layer to get the final combined representation 𝑧
where 𝑧 = MLP(𝑧_suumed)
3.5. Classifier Head
After finding the aggregated feature representation 𝑧, it is fed into a classification layer. The representa-
tion is the logits 𝑧 is employed for the classification process by the following:
′
𝑧 =𝑊 ·𝑧+𝑏
′
Finally, we calculate the Cross-Entropy (CE) loss based on 𝑧 with the ground truth.
4. Experimental Setup
4.1. Model Selection
For the tasks, our initial approach involved a thorough exploration of various model types to determine
the most appropriate one. Our investigation led us to examine three distinct categories: Machine
Learning (ML) Models, Deep Learning Models, and Transformer-based Pretrained Models. For machine
learning models, we considered Support Vector Machine (SVM), Random Forest (RF), and XGBoost.
Employing TF-IDF as our feature extractor, we processed each sentence through this mechanism before
inputting it into the ML models for classification. For task 3, we consider Logistic Regression instead of
XGBoost.
In our deep learning methodologies, we leverage both LSTM [17] and LSTM + Attention models. As
for transformer-based approaches, we’ve explored XLM-RoBERTa [18], mBERT [19], and BETO [20].
The outcomes are detailed in Table 3 for the development set. These model selections were driven by the
presence of two distinct languages within the dataset. The total unique words for TF-IDF experiments
were 39613 after deleting the punctuations, auxiliaries, and spaces. For LSTM experiments, we have used
embedding dim = 50, hidden units = 64 of the LSTM layer, and output dim = 256 of a fully connected
layer with a learning rate of 0.001. The experiments were done in 20 epochs. For the transformer-based
models, we fine-tuned the pre-trained models. We use 10 epochs with a batch size of 32
Table 3
Comparing the Performance of Different Models in Validation Dataset for Task 1 and Task 2
Performance Metric
Model Task 1 Task 2
Dev Acc↑ Dev F1↑ Dev Acc↑ Dev F1↑
ML Models
TF-IDF + SVM 70.88 69.73 61.12 26.70
TF-IDF + XGB 76.55 76.26 66.11 37.24
TF-IDF + RF 68.09 66.40 61.12 26.52
Deep Learning Models
LSTM 71.73 73.14 64.44 31.92
LSTM + Attention 74.95 74.82 65.28 32.87
Transformer based Models
XLM-RoBERTa 83.94 83.94 73.13 56.55
mBERT 81.05 81.04 71.58 53.16
BETO 80.73 80.72 70.99 54.27
Table 3 presents the results of different models for Task 1 and Task 2 datasets, while Table 4 displays
the model performance specifically for Task 3 on the development set. Analysis of Table 3 reveals that
classical ML models exhibit competitiveness against deep learning counterparts. Notably, employing
TF+IDF with XGBoost yields superior results to deep learning models. Incorporating the attention
module leads to a notable 3% enhancement in accuracy for Task 1 and 1% improvement Task 2. Ad-
ditionally, fine-tuning pretrained models demonstrates substantial improvements ranging from 6-9%.
Table 4
Comparing the Performance of Different Models in Validation Dataset for Task 3
Performance Metric
Model Task 3
Dev Acc↑ Dev F1↑
ML Models
TF-IDF + LogisticRegression 85.89 60.85
TF-IDF + SVM 84.85 55.79
TF-IDF + RF 84.85 55.79
Transformer based Models
XLM-RoBERTa 87.69 47.76
mBERT 87.86 48.21
BETO 87.41 47.13
Among these, XLM-Roberta demonstrates optimal performance for both Task 1 and Task 2, thereby
being selected as the final model for further experimentation. Turning to Task 3, as indicated in Table 4,
mBERT marginally outperforms XLM-RoBERTa. Consequently, mBERT is chosen as the final model for
Task 3.
4.2. Preprocessing
The datasets retrieved from Twitter contain additional information such as usernames and URLs
alongside the original posts and comments. To gauge their impact, we conducted experiments using
various preprocessing methods. These techniques included removing usernames, URLs, punctuation,
and emojis. The outcomes of these experiments are presented in Table 5. From the table, we can see
that removing url from the tweets improves the model performance for task 1 and task 3. For task 2, no
preprocessing is helpful.
Table 5
Effect of different preprocessing in dev set.
Performance Metric
Task Model Preprocessing
Dev Acc↑ Dev F1↑
No preprocessing 83.94 83.94
Username Removed 83.73 83.73
Task 1 XLM-RoBERTa
URL Removed 84.80 84.76
Punctuation and Emoji Removed 84.37 84.36
No preprocessing 72.06 51.43
Username Removed 71.94 52.32
Task 2 XLM-RoBERTa
URL Removed 71.11 52.00
Punctuation and Emoji Removed 71.11 53.82
No preprocessing 87.86 38.21
Username Removed 87.76 40.94
Task 3 mBERT
URL Removed 88.25 54.77
Punctuation and Emoji removed 87.86 48.72
4.3. Settings
For the hyper parameter settings, we also investigated with different values of them. We did ablation
studies on learning rate with two different values [1e-5, 2e-5], on batch size with values [16, 32] and on
random seed with values [0, 42]. The experimented results are reported in the Table [6, 7, 8]. Considering
Table 6
Effect of different learning rates in dev set.
Performance Metric
Task Learning Rate
Dev Acc↑ Dev F1↑
2 × 10-5 84.90 83.90
Task 1
1 × 10-5 84.80 84.76
2 × 10-5 72.06 51.43
Task 2
1 × 10-5 70.87 51.93
2 × 10-5 88.25 54.77
Task 3
1 × 10-5 85.49 52.36
Table 7
Effect of different batch sizes in dev set.
Performance Metric
Task Batch Size
Dev Acc↑ Dev F1↑
32 84.90 83.90
Task 1
16 82.01 81.92
32 72.06 51.43
Task 2
16 73.25 54.42
32 88.25 54.77
Task 3
16 85.35 51.69
Table 8
Effect of different seed values for model perforamce in dev set.
Performance Metric
Task Random Seed
Dev Acc↑ Dev F1↑
42 84.90 83.90
Task 1
0 84.05 84.03
42 73.25 54.42
Task 2
0 71.70 52.32
42 88.25 54.77
Task 3
0 85.39 54.52
those results, we set a learning rate of 2e-5, random seed = 42 and batch size 32 = for tasks 1 & 3 and 16
for task 3 for our model. We use AdamW optimizer in our experiments with betas = (0.9, 0.99).
All experiments were conducted using Python (version 3.12) and PyTorch, leveraging the free NVIDIA
Tesla P100 GPU provided by Kaggle. For the transformer based models we consider Huggingface
transformers library. All the transformer based models were run for 10 epochs.
4.4. Evaluation Metrics
When assessing the effectiveness of the models, we consider different performance metrics. Mainly we
focus on Accuracy, Macro-F1, and ICM for our performance as our evaluation metircs.
4.4.1. Accuracy
Accuracy measures the proportion of correctly classified instances among all instances. It is calculated
by dividing the sum of true positives (correctly predicted positive instances) and true negatives (correctly
predicted negative instances) by the total number of instances.
𝑇𝑃 + 𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁
4.4.2. Macro F1 Score
The F1 score is the harmonic mean of precision and recall. In the macro F1 score, each class is given
equal weight, and the mean of these F1 scores across all classes is calculated. Macro F1 Score is calculated
as follows:
2
F1-Score𝑖 = 1 1
𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛𝑖 + 𝑅𝑒𝑐𝑎𝑙𝑙𝑖
𝑁
1 ∑︁
Macro-F1 = F1-Score𝑖
𝑁
𝑖=1
4.4.3. ICM
The ICM metric functions as a similarity measure employed to assess the likeness between model-
generated outputs and the actual ground truth in classification tasks. It does so by comaring the
Information Content of catrgories presented through its features with the help of a Similarity function.
It extends the principles of Pointwise Mutual Information, a common metric used for evaluating
relationships between words. When all parameters in the ICM formula are set to 1, it becomes equivalent
to PMI. This means ICM can capture similar relationships as PMI but with more flexibility. ICM helps
evaluate how well a system’s output aligns with the ground truth by comparing the information content
of the categories they represent. (ICM) is adopted for all tasks and evaluation types (hard-hard, hard-soft,
soft-soft). ICM-soft is an extenstion of ICM for evaluating hierarchical multi-label classification tasks
where there might be disagreements in the ground truth. It can can handle both soft system outputs
and soft ground truth.
5. Result Analysis
Table 9 exhibits the results of various sentence representation extractors employing different aggregation
techniques across three tasks on the validation dataset. Regarding task 1, employing the CLS token-
based representation yields an accuracy of 84.90% and an F1 score of 83.90%. Transitioning to context
vector representation through an attention feature extractor results in a slight increase in the F1 score.
Furthermore, combining both types of representation through addition-based aggregation shows a 2%
enhancement in performance across the board. If we combine them using concat based aggregation
techniques we see a small improvement in F1 score but not like the addition based aggregation one
Table 9
Effect of different aggregation processes of the CLS token in dev set.
Performance Metric
Task Model Sentence Repr Extractor Aggregation Method
Dev Acc↑ Dev F1↑
XLM-RoBERTa CLS - 84.90 83.90
XLM-RoBERTa Attention - 84.15 83.94
Task 1
XLM-RoBERTa Attention + CLS Addition 86.30 86.28
XLM-RoBERTa Attention + CLS Concat 84.90 84.88
XLM-RoBERTa CLS - 72.06 51.43
XLM-RoBERTa Attention - 72.77 52.65
Task 2
XLM-RoBERTa Attention + CLS Addition 73.13 54.65
XLM-RoBERTa Attention + CLS Concat 74.20 57.87
mBERT CLS - 88.12 54.77
mBERT Attention - 87.97 51.50
Task 3
mBERT Attention + CLS Addition 84.40 53.50
mBERT Attention + CLS Concat 88.24 54.82
For task 2, we experience similar trend in the performance where attention based sentence rep-
resentation performs better than CLS token based sentence extractor. If we combine both CLS and
context vector getting from attention through addition based aggregation techniques, 1% improvement
in accuracy and 2% improvement in F1 score. While we consider concat based aggregation techniques
for combining both sentence level representation, the performance is further improved by 1.2% in
accuracy and around 3% improvement in F1 score. For task 3, we also get better result than the CLS
based baseline while we aggregate both sentence representations using concatenation.
Table 10
EXIST test results for submitted predictions with the best performing configuration along with CLS token
aggregation on the dev set.
Task Model [Agg. Type] Hard-Hard Rank ICM-Hard ICM-Hard Norm Macro F1
XLM-RoBERTa + ALL 29 47.79 74.02 75.08
Task 1 Attention + EN 29 46.01 73.48 72.09
CLS [Addition] ES 29 48.04 74.02 77.33
XLM-RoBERTa + ALL 9 20.89 56.79 48.56
Task 2 Attention + EN 13 12.03 54.16 44.31
CLS [Concat] ES 10 27.34 58.54 51.96
mBERT + ALL 14 -25.97 43.97 43.79
Task 3 Attention + EN 17 -26.92 43.40 41.56
CLS [Concat] ES 16 -27.41 43.88 44.60
The top-performing model is chosen to generate predictions for the test dataset across all three
tasks. The performance results on the test dataset are detailed in Table 10, showcasing the ICM-Hard,
ICM-Hard Norm, and Macro-F1 scores as our performance metrics. Our best-performing model achieves
an F1 score of 75.01 for all tweets, with 72.09 and 77.33 F1 scores recorded for English and Spanish
tweets, respectively. For task 2, the score is 48.56, while task 3 reaches 43.79 across all tweet data. The
table shows that our model can predict Spanish tweets more effectively than English tweets.
6. Error Analysis
6.1. Confusion Matrix
The evaluation of our approach for Task 1 is done on the development dataset. The confusion matrix
shown in 2(a) represents the performance of the classification. The number of True Positive is the
number of correctly predicted positive cases. The model correctly predicted 388 cases as "Yes" (True Yes).
True Negative cases is the number of correctly predicted negative cases. The model correctly predicted
418 cases as "NO" (True No). False Positive is the number of incorrectly predicted positive cases. The
model incorrectly predicted 61 cases as "YES" when they were actually "NO" (False YES). False Negative
(FN) is the number of incorrectly predicted negative cases. The model incorrectly predicted 67 cases
as "NO" when they were actually "YES" (False NO). The model produced 806 correct classification as
opposed to 128 misclassification.
Similarly, the confusion matrix is shown in 2(b) shows the result of the evaluation done on the
development dataset for Task 2. As shown in the figure, out of 479 true NO labels, 420 cases were
correctly predicted as NO, 150 were correctly predicted as DIRECT out of 204 true DIRECT cases, 36
were correctly predicted as REPORTED out of 65 true REPORTED cases and 18 were correctly predicted
as JUDGEMENTAL out of 83 true JUDGEMENTAL cases. The model produced 624 correct classification
as opposed to 217 misclassification.
6.2. Attention Heatmap
We have calculated the Layer Integrated Gradient attributions for few specific input tweets using
Captum [21]. The attributions explain how each input element of a tweet contributes to the model’s
prediction for the target class. It means that the attention provided by the model to each tokens. So we
name it attention heatmap. The darker the color in the cell of a token, the higher its attribution value
(a) Task 1 (b) Task 2
Figure 2: Confusion matrix with all classes for dev set with best performing configuration of each task.
and the higher its contribution to the target. We have considered both English and Spanish tweet for
the experiment and creating the heatmap. The corresponding labels are also listed in the Figure 3. We
can infer that for ’NO’ label the most attentive tokens among the 4 tweets are last, abuse, ada, in and for
’YES’ label the most attentive tokens among the 4 tweets are economy, s, a, LAS. This experiment was
done considering the best performing model in the task 1.
Attention heatmap for tokens of a sample tweet Lang Label
EN NO
EN NO
ES NO
ES NO
EN YES
EN YES
ES YES
ES YES
Figure 3: Attention heatmap for task 1 of a few samples of the dev set with XLM-RoBERTa and the
best performing configuration.
6.3. Most Attentive Tokens
We also extracted the most attentive tokens for each category in the validation dataset for further
analysis. We determine these tokens by extracting the highest average self-attention scores for each
token across the entire category-wise validation dataset. Transformer-based models compute attention
scores to gauge the relationship between 𝑤𝑜𝑟𝑑𝑖 and 𝑤𝑜𝑟𝑑𝑗 across various layers and heads. Considering
a transformer-based model with 𝐿 layers and 𝐻 heads per layer, and a sequence length of 𝑛 for a given
sentence 𝑆, the resulting attention matrix has dimensions of 𝐿 × 𝐻 × 𝑛 × 𝑛. To compute the average
attention score for token𝑖 within sentence 𝑆, we follow this calculation:
𝐿 ∑︁
𝐻 𝑛
1 ∑︁ ∑︁
Avg_Attn_Score𝑖𝑆 = Attention𝑖,𝑡,ℎ,𝑙
𝐿·𝐻 ·𝑛
𝑙=1 ℎ=1 𝑡=1,𝑡̸=𝑖
To find category-wise average attention score for token𝑖 on the validation dataset, we take the average
of Avg_Attn_Score𝑖𝑆 for those sentences where token𝑖 appears in. After calculating the token attention
scores, we arrange them in descending order and select the top K words. In Table 11, we present the
category-wise most attentive tokens predicted by the top-performing model for both Task 1 and Task 2.
Table 11
List of most attentive tokens for tasks 1 and 2 for each class with the best performing configuration for each task.
Task Label List of Most Attentive Tokens [Top 10]
volant, slut, cum, hot , summer
NO
peligro, constante, staff, development , practic
Task 1
penis, forever, death, economy , ika
YES
GPS, tomorrow, exam, gangbang, room
pun, wall, question, falso, auténtico
NO
modo, concert, syon, yes, making
where, GPS, tomorrow, exam, room
DIRECT
attention, foot, cla, without, uni
Task 2
wenr, school, bath, dom, GAL
REPORTED
ILE, ingu, ebla, dia, ACIÓN
saben, ándose, cocina, make, sense
JUDGEMENTAL
want, mother, ape, ici, gas
The table presents a detailed breakdown of the most attentive tokens categorized by the top-
performing model across two distinct tasks. In Task 1, where the classification pertains to identifying
sexist content, the model identifies tokens such as "volant," "slut," and "cum" as prominent indicators
for non-sexist content, while tokens like "penis," "forever," and "gangbang" are highlighted for identify-
ing sexist content. Task 2, which involves various categories like "NO," "DIRECT," "REPORTED," and
"JUDGEMENTAL," exhibits a diverse range of most attentive tokens. For instance, in the "NO" category,
tokens like "pun," "wall," and "question" stand out, while "DIRECT" category tokens include "where,"
"GPS," and "tomorrow." Additionally, tokens such as "wenr," "school," and "saben" are emphasized in the
"REPORTED" and "JUDGEMENTAL" categories, respectively. This comprehensive analysis sheds light
on the model’s attention mechanism and its discernment of different types of content within each task.
7. Conclusion
The EXIST challenge is designed to promote research on automated sexism detection and modeling in
online environments, with a particular focus on Twitter. In this paper, we performed extensive research
and conducted thorough experiments to achieve this objective, employing advanced techniques in
natural language understanding and machine learning. Specifically, we enhanced existing machine
learning models by incorporating an attention layer and a CLS token, which emphasize the words that
contribute to the context of sexism. Our findings demonstrate the effectiveness of our approach and
models on both the training and validation datasets.
However, as the training is done on Spanish and English language the model might not be proficient
at identifying sexist content in other languages. This can lead to misclassification of sexist Tweets.
Acknowledgements
This project has been sponsored by Penta Global Limited Bangladesh. We would like to express our
deepest gratitude to Penta Global for for their financial support.
References
[1] J. B. Walther, Social media and online hate, Current Opinion in Psychology 45 (2022) 101298.
[2] A. Elbarazi, How social media affects people’s ideas on sexist behaviours and gender-based
violence (2023). doi:10.19080/GJIDD.2023.12.555838.
[3] Z. Al-Makhadmeh, A. Tolba, Automatic hate speech detection using killer natural language
processing optimizing ensemble deep learning approach, Computing 102 (2020) 501–522.
[4] L. Plaza, J. Carrillo-de-Albornoz, V. Ruiz, A. Maeso, B. Chulvi, P. Rosso, E. Amigó, J. Gonzalo,
R. Morante, D. Spina, Overview of EXIST 2024 – Learning with Disagreement for Sexism Identifi-
cation and Characterization in Social Networks and Memes, in: Experimental IR Meets Multilin-
guality, Multimodality, and Interaction. Proceedings of the Fifteenth International Conference of
the CLEF Association (CLEF 2024), 2024.
[5] L. Plaza, J. Carrillo-de-Albornoz, V. Ruiz, A. Maeso, B. Chulvi, P. Rosso, E. Amigó, J. Gonzalo,
R. Morante, D. Spina, Overview of EXIST 2024 – Learning with Disagreement for Sexism Identifi-
cation and Characterization in Social Networks and Memes (Extended Overview), in: G. Faggioli,
N. Ferro, P. Galuščáková, A. G. S. de Herrera (Eds.), Working Notes of CLEF 2024 – Conference
and Labs of the Evaluation Forum, 2024.
[6] S. M. H. Dadgar, M. S. Araghi, M. M. Farahani, A novel text mining approach based on tf-idf
and support vector machine for news classification, in: 2016 IEEE International Conference on
Engineering and Technology (ICETECH), IEEE, 2016, pp. 112–116.
[7] G. Ozogur, M. A. Erturk, Z. Gurkas Aydin, M. A. Aydin, Android malware detection in bytecode
level using tf-idf and xgboost, The Computer Journal 66 (2023) 2317–2328.
[8] V. Sundaram, S. Ahmed, S. A. Muqtadeer, R. R. Reddy, Emotion analysis in text using tf-idf, in: 2021
11th International Conference on Cloud Computing, Data Science & Engineering (Confluence),
IEEE, 2021, pp. 292–297.
[9] J. Nowak, A. Taspinar, R. Scherer, Lstm recurrent neural networks for short text and sentiment
classification, in: Artificial Intelligence and Soft Computing: 16th International Conference,
ICAISC 2017, Zakopane, Poland, June 11-15, 2017, Proceedings, Part II 16, Springer, 2017, pp.
553–562.
[10] X. Bai, Text classification based on lstm and attention, in: 2018 Thirteenth International Conference
on Digital Information Management (ICDIM), IEEE, 2018, pp. 29–32.
[11] X. Ou, H. Li, Ynu@ dravidian-codemix-fire2020: Xlm-roberta for multi-language sentiment
analysis., in: FIRE (Working Notes), 2020, pp. 560–565.
[12] L. Khan, A. Amjad, N. Ashraf, H.-T. Chang, Multi-class sentiment analysis of urdu text using
multilingual bert, Scientific Reports 12 (2022) 5436.
[13] M.-I. Limaylla-Lunarejo, N. Condori-Fernandez, M. R. Luaces, Requirements classification using
fasttext and beto in spanish documents, in: International Working Conference on Requirements
Engineering: Foundation for Software Quality, Springer, 2023, pp. 159–176.
[14] M. Fahim, M. S. Shahriar, M. R. Amin, Hatexplain space model: Fusing robustness with explain-
ability in hate speech analysis (????).
[15] M. Fahim, Aambela at blp-2023 task 2: Enhancing banglabert performance for bangla sentiment
analysis task with in task pretraining and adversarial weight perturbation, in: Proceedings of the
First Workshop on Bangla Language Processing (BLP-2023), 2023, pp. 317–323.
[16] M. Fahim, A. A. Ali, M. A. Amin, A. M. Rahman, Edal: Entropy based dynamic attention loss
for hatespeech classification, in: Proceedings of the 37th Pacific Asia Conference on Language,
Information and Computation, 2023, pp. 775–785.
[17] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural computation 9 (1997) 1735–1780.
[18] A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, F. Guzmán, E. Grave, M. Ott,
L. Zettlemoyer, V. Stoyanov, Unsupervised cross-lingual representation learning at scale, arXiv
preprint arXiv:1911.02116 (2019).
[19] J. Libovickỳ, R. Rosa, A. Fraser, How language-neutral is multilingual bert?, arXiv preprint
arXiv:1911.03310 (2019).
[20] J. Cañete, G. Chaperon, R. Fuentes, J.-H. Ho, H. Kang, J. Pérez, Spanish pre-trained bert model and
evaluation data, in: PML4DC at ICLR 2020, 2020.
[21] N. Kokhlikyan, V. Miglani, M. Martin, E. Wang, B. Alsallakh, J. Reynolds, A. Melnikov, N. Kliushk-
ina, C. Araya, S. Yan, O. Reblitz-Richardson, Captum: A unified and generic model inter-
pretability library for pytorch, CoRR abs/2009.07896 (2020). URL: https://arxiv.org/abs/2009.07896.
arXiv:2009.07896.