This paper presents the system submitted to the DETOXIS 2021 challenge for detecting toxicity in Spanish social media texts. The chosen approach relies on an ensemble of di erent neural network architectures including thread and topic features as side information. For sub-task 1, we have also applied machine translation in order to reuse linguistic resources from other languages such as English. Our best submission scored 0.569 F1 in the test set, ranking 6th out of 31 competing teams.
News websites allow million of users to share and discuss their opinions publicly in near real-time every day. Such large reach and constantly increasing user base present challenges for content moderation teams, which not only need to ght a liate and cyber-crime operators but also less traditional forms of messaging abuse such as the spread of hate, propaganda and fake news.
While social media platforms are under increasingly pressure to swiftly deal
with the spread of toxic content, the use of over-aggressive ltering models and
the under-representation of certain user groups in the training data can also have
negative consequences if false positives happen at large scale [
Because of the aforementioned reasons, the automatic detection of toxic
language in social media has received growing attention from the NLP research
community in the last few years, which is also re ected in the number of public
evaluations and resources recently focused on this area: e.g. HASOC [
This paper evaluates our participation in the shared task DETOXIS [
Best performing approaches for toxicity detection follow the recent advances in
neural networks for NLP: Liu et al. [
Since the rst sub-task was only focused on determining if a comment is either toxic or not, we have treated it as a binary classi cation problem. 3.1
Social media texts usually contain informal lexical variants and out-of-vocabulary
words which can be di cult to understand not only for humans but also for
NLP tools and applications [
Data augmentation is a popular technique that can increase the volume and
diversity of the training data for many applications including NLP [
The list of models that our system comprises of is as follows:
{ capsule es Neural network with a capsule network architecture [
SBWC [
In addition to the actual comments, non-textual metadata was made available as part of the training dataset such as topic, thread id, comment id and reply to. These were used in order to engineer extra features for the stacking model as side information: { topic words max The maximum word-wise toxicity score in a comment after averaging all the capsule es model probabilities of the individual words across the training data by topic. { topic words avg The average word-wise toxicity score in a comment after averaging all the capsule es model probabilities of the individual words across the training data by topic. { avg group tox The average toxicity score determined by the capsule es model for all the comments with the same thread id.
Since there was no topic information in the test dataset, we have considered it as a separate topic when computing the features above. Although inaccurate (the test data had comments from the same set of topics as train) it did not impact negatively in the nal results. 3.5
Due the relatively small amount of training data in the NewsCom-TOX corpus
(less than 4000 samples, only 1147 positive) we went for an stacked generalization
strategy, where the soft probabilities calculated from several models are used
as features with the original labels against the stacking model. This not only
reduces the computing resources needed in order to tune hyperparameters and
perform cross-validation, but can also achieve competitive results even with just
pre-trained models [
Our stacking model was logistic regression with a custom threshold of 0.32, which was determined via cross-validation. The latter was required because of the class imbalance and the unusual evaluation metric used in this sub-task (F1 of the toxicity class rather than micro or macro averages) which favours aggressive models towards the positive class.
The most important features by considering the regression coe cients can be seen at Figure 1. From there we can determine that capsule networks and avg group tox are the strongest features for detecting the toxic and non-toxic class respectively. Our toxicity detection system obtained promising results as shown in Table 1: It ranked 6th/31, with a di erence in F1 of only 0.077 when compared against the winning system. It is also worth mentioning that only 16 systems (out of 31) achieved better F1 score than the AllToxic benchmark, which highlights the di culty of this sub-task for the chosen evaluation metric.
System F1 Toxic Model F1 Toxic Model F1 Toxic SINAI (best) 0.6461 capsule es 0.5040 capsule es 0.5168 Alejandro Mosquera 0.5691 capsule en 0.4872 capsule en 0.5299 AllToxic 0.4231 spacylr 0.4833 spacylr 0.5156 RandomClassi er 0.3760 detox unb 0.4117 detox unb 0.4671 ChainBOW 0.3746 detox multi 0.4053 detox multi 0.4430 BOWClassi er 0.1837 detox multi es 0.3887 detox multi es 0.4209 detox orig 0.3542 detox orig 0.4237
Alejandro Mosquera 0.5813
With regards to our individual models, we can observe that they are weaker, only 3 out of 7 would beat the AllToxic baseline, and exhibit higher variance between train and test scores than the nal stacking ensemble. However, a postworkshop analysis showed that removing the weakest models would have not improved the nal score. 5
In this paper we describe the system for detecting toxicity in Spanish social media texts engineered for DETOXIS 2021 sub-task 1. Since the amount of training data was relatively small, di erent strategies were applied in order to overcome this limitation, such as performing data augmentation through machine translation and leveraging pre-trained models using larger toxicity datasets. Our best submission was a logistic regression ensemble using neural network predictions and side information features extracted from thread and topic metadata.