This paper describes our participation in the DEtection of TOXicity in comments In Spanish (DETOXIS) shared task 2021 at the 3rd Workshop on Iberian Languages Evaluation Forum. The shared task is divided into two related classi cation tasks: (i) Task 1: toxicity detection and; (ii) Task 2: toxicity level detection. They focus on the xenophobic problem exacerbated by the spread of toxic comments posted in di erent online news articles related to immigration. One of the necessary e orts towards mitigating this problem is to detect toxicity in the comments. Our main objective was to implement an accurate model to detect xenophobia in comments about web news articles within the DETOXIS shared task 2021, based on the competition's o cial metrics: the F1-score for Task 1 and the Closeness Evaluation Metric (CEM) for Task 2. To solve the tasks, we worked with two types of machine learning models: (i) statistical models and (ii) Deep Bidirectional Transformers for Language Understanding (BERT) models. We obtained our best results in both tasks using BETO, a BERT model trained on a big Spanish corpus. We obtained the 3rd place in Task 1 o cial ranking with the F1-score of 0.5996, and we achieved the 6th place in Task 2 o cial ranking with the CEM of 0.7142. Our results suggest: (i) BERT models obtain better results than statistical models for toxicity detection in text comments; (ii) Monolingual BERT models have an advantage over multilingual BERT models in toxicity detection in text comments in their pre-trained language.
The increase in the number of news pages where the reader can openly discuss
the articles has driven the dissemination of internet users' opinions through
social media [
The imbalance in the amount of data generated by in uencers and non-active
users creates a distorted reality where in uencers' opinions end up representing
the opinion of all internet users to society [
In recent years, the problem with xenophobia has been exacerbated by the
increase in the spread of toxic comments posted in di erent online news articles
related to immigration [
The DETOXIS shared task comprises Task 1 and Task 2, which are respectively toxicity detection and toxicity level detection. The two tasks are performed on comments posted in Spanish in response to di erent online news articles related to immigration. Task 1 is a binary classi cation problem where the objective is to classify a Spanish text comment as `toxic' or `not toxic'. Task 2 aims to classify the same comment but among four classes: `not toxic', `mildly toxic', `toxic', or `very toxic'. Table 1 displays examples of comments classi ed across all classes. asi me gusta, que se maten entre ellos y en alta mar. Mas inmigrantes asi porfavor
The detection of toxicity in comments is mostly done with Machine Learning
(ML) models, especially deep learning models, which require large amounts of
annotated datasets for robust predictions [
Our main goal was to implement an accurate model to detect xenophobic
comments on web news articles within the DETOXIS shared task 2021, using
the competition's o cial metrics. We decided to solve the problem by applying
models that can learn using only a small amount of data, which can be done
with statistical models and most advanced pre-trained deep learning models.
Roughly speaking, there are two types of statistical models: Generative and
Discriminative [
The work's main contribution is to help in the e ort to improve the results in the identi cation of toxic comments in news articles related to immigration. Unlike the vast majority of works [14], we use ML models that can tackle the xenophobia detection problem having only little data available. The second contribution is to build ML models and nd their best con guration to deal not only with the classi cation of news articles as `toxic' and `not toxic', but also to infer the toxicity level of the comments into `not toxic', `mildly toxic', `toxic', or `very toxic'. As far as we know, there are few works in the literature in which the solution model tries to infer the toxicity level of the comments posted in the news related to immigration. On the DETOXIS o cial ranking, we obtained the 3rd place in Task 1 with the F1-score of 0.5996, and we achieved the 6th place in Task 2 with the CEM of 0.7142.
The article is organized as follows: Section 2 contains the methodology with fundamental concepts; Section 3 describes the experiments; Section 4 contains the results and discussions, and and Section 5 draws some conclusion and future work. 2
This section explains the data structure, the evaluation metrics, and the ML models applied to solve our classi cation problems. In addition, the text representation used to encode the text comments. 2.1
The DETOXIS shared task organization granted its participants the NewsComTOX dataset [17] divided into train set and test set where text data are in Spanish. The train set consists of 3463 instances, and the test set consists of 891 instances. Both sets have as main labels: (i) `Comment id' and (ii) `Comment'; but only the train set has the labels: (iii) `Toxicity' and (iv) `Toxicity level', respectively for Task 1 and Task 2. The `Comment id' is a unique reference number assigned to each instance within the NewsCom-TOX dataset. The `Comment' label is a text message posted in response to a Spanish online news article from di erent sources such as El Mundo, NIUS, ABC, etc., or discussion forums like 1 https://github.com/AngelFelipeMP/Machine-Learning-Tweets-Classification Meneame. Moreover, `Toxicity' labels the comment for a particular instance between `toxic' or `not toxic' and the `Toxicity level' label classi es the same comment as `not toxic', `mildly toxic', `toxic', or `very toxic'. Table 2 shows the label's distribution for `Toxicity' and `Toxicity level'. We can see that the labels are unbalanced in both cases.
The data annotation process was carried out by four annotators where two were linguists experts, and two were trained linguistic students. Three of them labeled all news article comments in parallel. Once they nished, an interannotator agreement test was executed. When a disagreement happens, the three annotators plus the senior annotator reviewed it in order to achieve accordance with the nal label [17]. In Table 1, we can see examples from the DETOXIS train set of comments and its labels attributed by the annotators for `Toxicity' and `Toxicity level'.
Next we explain how we used the data during the project development in both tasks. First, we applied 10-fold cross-validation in the train set to nd the best ML model. After that, we trained the selected model in the whole train set. Subsequently, we applied the selected model to make predictions on the o cial test set, as shown in Figure 1. These predictions were submitted to the DETOXIS shared task 2021. 2.2
Because the train set is imbalanced, as we can see in Table 2, we selected
evaluation metrics that are able to fairly evaluate ML models in this circumstance.
For Task 1, we adopted Accuracy, Recall, Precision, and F1-score, which was
the DETOXIS o cial evaluation metric for Task 1. For Task 2, we adopted
Accuracy, F1-macro, F1-weighted, Recall, Precision, and CEM [
To represent our text data in a way that the statistical models could handle it, we used two encode methods: (i) Bag of Words (BOW) [20]; and (ii) Term Frequency - Inverse Document Frequency (TF-IDF) [13]. The BOW represents a text comment by a unidimensional vector whose length is the size of the training vocabulary. In this case, each column of this vector contains the number of times a particular word from the vocabulary appears in the speci c comment. The TFIDF representation for each text comment is also a at vector with the size of the training vocabulary. However, the value for each word on the vector follows the well-known TF-IDF calculation [13].
This section explains the environment setup, the data preprocessing, and statistical models' feature extraction. Furthermore, the section also contains explanations for the 10-fold cross-validation process and how we selected the model to make predictions on the DETOXIS test set, which we submitted as our nal results to the competition. 3.1
For code purposes, we used python 3.7.10. As a code editor/machine, we used Google collaborator2. The main python libraries that we used were: (i) NumPy 1.19.5 to work with matrix, (ii) Pandas 1.1.5 to handle and visualize data, (iii) Spacy 2.2.4 and (iv) the Natural Language Toolkit (NLTK) 3.2.5 for natural language transformations, (v) Pytorch 1.6.0, and (vi) Transformers 3.0.0 to actually implement the BERT models. In addition, we used (vii) Sklearn 0.22.2 to implement the statistical models. 3.2
For both tasks, we only preprocess the data for the statistical models. The preprocessing step was carried out on the text data from the train and test sets. We used the built-in python model for Regular Expression (RegEx) and the NLTK python library. Applying RegEx, we removed stock market tickers, oldstyle retweet text, hashtags, hyperlinks and changed the numbers to the tag \<number>". We employed the NLTK on the text comments to remove stopwords, stem and tokenize the words. 3.3
The feature extraction process was executed to focus on achieving good results
with the statistical models. These models' performance is susceptible to their
input features [
The two proposed encode methods are based on word occurrences, and unfortunately, they completely ignore the relative position information of the words in the comments. Therefore, we lose the information about the local ordering of the words. In order to mitigate this problem and preserve some of the local word ordering information, we increase the vocabulary by extracting 2-grams and 3-grams from the text comments despite dimensional increasing. 2 https://colab.research.google.com/ 3.4
The cross-validation process was performed on the train set aiming to nd the best ML model to make the prediction on the DETOXIS test set. We can see the summary of our cross-validation process in Figure 2. During the cross-validation, each statistical model received di erent input features, and the BERT models tried di erent hyper-parameters.
In Figure 3, we can see the cross-validation process that focuses on the BERT models. We tried di erent combinations for the Output BERT, Learning Rate, Batch Size, and Epochs. The BERT models are composed of a pre-trained model plus a linear layer at the top which receives the output of the pre-trained BERT model. We have two di erent options for the Output BERT: (i) the sequence of hidden states at the output of the last layer which we performed a mean pooling and max pooling operation and concatenated them into a uni ed unidimensional vector that we called `hidden'; (ii) the pooler of the last layer's hidden state of the rst token of the sequence further processed by a linear layer and a tanh activation function that we called `pooler'. For the Learning Rate, we tried 1E-5, 3E-5, and 5E-5. For the Batch Size, we tried 8, 16, 32, and 64. The number of Epochs was from 1 to 20.
Figure 4 illustrates the cross-validation process for the statistical models. We can see that we tried four algorithm versions of the Naive Bayes (NB) model: the Multinomial, the Bernoulli, the Gaussian, and the Complement ones. On the other hand, we tried only the original version of the Maximum Entropy (ME) model but with di erent solvers: the liblinear, newton, sag, saga, and lbfgs. We call solvers the algorithms used in the optimization problem. We tried di erent vocabulary sizes for all statistical models using di erent n-grams combinations and the two encode methods: BOW and TF-IDF. parameters. For Task 1, the evaluation metrics are Accuracy, F1-score, Recall, and Precision, and for Task 2, the evaluation metrics are Accuracy, F1-macro, F1-weighted, Recall, Precision, and CEM.
Model OBuEtRpTut LeRaranteing BSaitzceh Epochs Accuracy F1-score Recall Precision Model OBuEtRpTut LeRaranteing BSaitzceh Epochs Accuracy F1-score Recall Precision mBERT pooler hidden hidden hidden pooler 3E-05 5E-05 5E-05 3E-05 3E-05 BETO pooler pooler pooler pooler pooler 1E-05 1E-05 1E-05 5E-05 At the end of the cross-validation, we selected the best model for each task accordingly with the DETOXIS o cial metric for the speci ed task, as shown in Figure 5.
OBuEtRpTut LeRaranteing BSaitzceh Epochs Accuracy mFac1ro weiFgh1ted Recall Precision CEM
After the cross-validation, we chose the best model for Task 1, which following Table 12 is BETO with the respective parameters: (i) pooler as Output BERT; (ii) 1E-05 Learning Rate; (iii) Batch Size equal 32; and (iv) 4 training Epochs. We also selected the best model for Task 2 that following Table 14 is BETO with the respective parameters: (i) hidden Output BERT; (ii) 1E-05 Learning Rate; (iii) Batch Size equal 16; and (iv) 4 training Epochs. Having the best models and their parameters, we trained the models on the train set.
Once the best models are trained, we use those models to make the predictions on the DETOXIS test set. These predictions afterward were submitted to the DETOXIS shared task organization as our nal results. 4
We discovered important information on the cross-validation results. Looking at Table 3, we can see that the ME model achieves its best results on Task 1 with the BOW encode based on the F1-score evaluation metric, which is 0.4679. The highest Accuracy 0.7126 and Recall 0.4019 are also performed with the BOW encode. The only performance metric in which the TF-IDF encode obtains a higher score is Precision that is 0.8928. Thus, we can conclude that BOW is the best encoding for the ME model on Task 1 in the DETOXIS training set. Moreover, employing the Sag solver, the ME model achieved a higher F1-score, Recall, and Precision. Hence, it seems to us that Sag was the best solver for the ME model on Task 1 in the DETOXIS training set. We do not have a de nitive conclusion about the vocabulary size because the ME model achieved its highest results with di erent numbers of n-grams for each metric.
Observing Table 4, we see that the NB model achieves its best results on Task 1 with the BOW encode based on the F1-score evaluation metric, which is 0.5355. The NB model also obtained the higest Recall 0.8004 with the BOW encode, but its highest results for Accuracy 0.6933 and Precision 0.7282 were with the TF-IDF encode. Therefore, we can not conclude which encode method is the best for the NB model on Task 1 in the DETOXIS training set. A similar case occurs with the vocabulary size where the NB model that employed 1-grams, 2-grams, and 3-grams achieved the highest F1-score and Recall. However, the NB model with a 1-grams vocabulary size obtained the highest Accuracy and Precision. The di erent NB algorithms obtained a similar performance based on the F1-score. In most cases, they achieved their best results with the BOW encode.
We can see in Table 5 that the ME model achieved its best results on Task 2 with the BOW encode based on the CEM evaluation metric, which is 0.7080. The highest F1-macro 0.3587, F1-weighted 0.6125, Recall 0.3367, and Precision 0.4942 are also obtained with the BOW encode. The only performance metric in which the TF-IDF encode obtains a higher score is the Accuracy, which is 0.6846. Thus, we can conclude that BOW is the best encoding for the ME model on Task 2 in the DETOXIS training set. Moreover, employing the Newton solver, the ME model achieved a higher Accuracy, F1-macro, F1-weighted, Recall, and Precision. Hence, it seems to us that Newton was the best solver for the ME model on Task 2 in the DETOXIS training set. We concluded that the vocabulary size of 1-grams is the best for the ME model on Task 2 in the DETOXIS training set because the ME model achieved its highest Accuracy, F1-macro, F1-weighted, Recall, and Precision.
Table 6 shows that the NB model achieved its best results on Task 2 with the TF-IDF encode based on the CEM evaluation metric, which is 0.6882. The NB model also obtained its highest Accuracy 0.6769, F1-weighted 0.5845, and Precision 0.3171, with the TF-IDF encode, but its highest results for F1-macro 0.2916 and Recall 0.3365 were obtained with the BOW encode. Therefore, we can conclude that the TF-IDF encode best suits the NB model on Task 2 in the DETOXIS training set. We see indications in Table 6 that the ideal vocabulary for the NB model on Task 2 in the DETOXIS training set is composed of 1-grams and 2 grams. Once with this vocabulary, the model obtained its highest Accuracy, Recall, Precision, and CEM results. The di erent NB algorithms obtained similar performance based on the CEM ranged from 0.49 to 0.68.
Based on the F1-score, the mBERT model achieved its best performance on Task 1 with a value of 0.6010, as we can see in Table 7. The model parameters are the following: (i) pooler as Output BERT; (ii) 3E-05 Learning Rate; (iii) Batch Size equal 32; and (iv) 11 training Epochs. Table 8 shows that the BETO model obtained its best performance on Task 1 also based on the F1-score with the following parameters: (i) pooler as Output BERT; (ii) 1E-05 Learning Rate; (iii) Batch Size equal 32; and (iv) 4 training Epochs. The BETO model obtained a F1score value of 0.6314, which was also the highest among all the ML models in the cross-validation process. For this reason, the BETO model with the mentioned parameters was used for our Task 1 o cial prediction on the DETOXIS test set. These predictions afterward were submitted as our o cial Task 1 results.
Observing Table 9, we can conclude that based on the CEM, the mBERT model achieved its best performance on Task 1 with the following parameters: (i) pooler as Output BERT; (ii) 1E-05 Learning Rate; (iii) Batch Size equal to 16; and (iv) 12 training Epochs. This model achieved the CEM of 0.7599. Table 10 shows that the BETO model obtained its best performance on Task 2 also based on the CEM with the following parameters: (i) hidden as Output BERT; (ii) 1E05 Learning Rate; (iii) Batch Size equal to 16; and (iv) 4 training Epochs. The BETO model obtained CEM value of 0.7769, which was also the highest among all the ML models in the cross-validation process. For this reason, the BETO model with the mentioned parameters was used for our Task 2 o cial prediction on the DETOXIS test set. These predictions afterward were submitted as our o cial Task 2 results.
To sum up the comments about the cross-validation results, looking at Tables 12 and 14, we can see that the BETO model with di erent combinations of parameters obtained the ve rst positions on the ranking for the best ML model for Task 1 and Task 2.
The DETOXIS organization provided us with the results of the test set. Table 15 shows our result on Task 1 plus the three o cial DETOXIS baselines: Random Classi er, Chain BOW, and BOW Classi er. Our model obtained an F1-score around 59% greater than the results obtained by the best baseline on Task 1.
Table 16 shows the results of our model and the three DETOXIS baselines on Task 2. Our BETO model was able to achieve a CEM of 9% higher than the best DETOXIS baseline result obtained by the Random Classi er.
On the DETOXIS o cial ranking, we obtained 3rd place on Task 1 with F1-score 0.5996, and we achieved 6th place on Task 2 with CEM 0.7142. 5
Xenophobia is a problem which is aggravated by the increase in the spread of toxic comments posted in di erent online news articles related to immigration. In this paper, to address this problem within the DETOXIS 2021 shared task, we tried two types of ML models: (i) statistical models and (ii) BERT models. We obtained the best results in both tasks using BETO, a BERT model pretrained with a big Spanish corpus. Our contributions are as follows: (i) help in the e ort to improve the results in the identi cation of toxic comments in news articles related to immigration. Unlike the vast majority of works, we use ML models that can tackle the xenophobia detection problem having only little data available; (ii) We build an ML model and nd its best con guration to deal not only with the classi cation of news articles as `toxic' and `not toxic', but also to infer the toxicity level of the comments into `not toxic', `mildly toxic', `toxic', or `very toxic'.
Based on the DETOXIS o cial metrics, we concluded that our results indicate that: (i) BERT models obtain better results than statistical models for toxicity and toxicity level detection in text comments; and (ii) Monolingual BERT models achieve higher results in comparison with the multilingual BERT models in toxicity detection and toxicity level detection in their pre-trained language.
After all, our BETO model obtained the 3rd position on Task 1 o cial ranking with the F1-score of 0.5996, and it achieved the 6th position on Task 2 o cial ranking with the CEM of 0.7142. As future work, we aim to include sentiment lexicons on the model's input to boost its performance. 13. Pimpalkar, A.P., Raj, R.J.R.: In uence of pre-processing strategies on the performance of ml classi ers exploiting tf-idf and bow features. ADCAIJ: Advances in Distributed Computing and Arti cial Intelligence Journal 9(2), 49{68 (2020) 14. Plaza-Del-Arco, F.M., Molina-Gonzalez, M.D., Uren~a Lopez, L.A., Mart nValdivia, M.T.: Detecting misogyny and xenophobia in spanish tweets using language technologies. ACM Trans. Internet Technol. 20(2) (Mar 2020). https://doi.org/10.1145/3369869 15. Risch, J., Krestel, R.: Delete or not delete? semi-automatic comment moderation for the newsroom. In: Proceedings of the rst workshop on trolling, aggression and cyberbullying (TRAC-2018). pp. 166{176 (2018) 16. Stroud, N.J., Van Duyn, E., Peacock, C.: News commenters and news comment readers. Engaging News Project pp. 1{21 (2016) 17. Taule, M., Ariza, A., Nofre, M., Amigo, E., Rosso, P.: Overview of the detoxis task at iberlef-2021: Detection of toxicity in comments in spanish. Procesamiento del Lenguaje Natural 67 (2021) 18. Winter, S., Bruckner, C., Kramer, N.C.: They came, they liked, they commented: Social in uence on facebook news channels. Cyberpsychology, Behavior, and Social Networking 18(8), 431{436 (2015) 19. Xenophobia: Retrieved from https://en.oxforddictionaries.com/de nition/money.
Oxford Online Dictionary (2021) 20. Zhang, Y., Jin, R., Zhou, Z.H.: Understanding bag-of-words model: a statistical framework. International Journal of Machine Learning and Cybernetics 1(1-4), 43{52 (2010)