Recent work shows that offensive language in social media is a serious problem that affects especially vulnerable groups. Therefore, systems designed to detect offensive language automatically have been the focus of attention of several works. Various Machine Learning approaches have been utilised for the classification of offensive text data. Within the scope of this research we aim to develop a neural network system that will effectively classify offensive text considering different aspects of it. In addition, multilingual and multi-task learning experiments are planned.
Social media has become one of the most
important environments for communication
among people. As user-generated content on
social media increases significantly, so does
the harmful content such as offensive
language. Aggressiveness in social media is a
problem that especially affects vulnerable groups
Traditional methods to detect offensive language include use of blacklisted keywords and phrases based on profane words, regular expressions, guidelines and human moderators to manually review and detect unwanted content. However, these methods are not sufficient, particularly considering the users that tend to use more obfuscated and implicit expressions.
Automatic identification of offensive
language is essentially considered as a
classification task. Previous research on the topic
include approaches from different
perspectives, utilizing different data sets and focusing
on various contents such as abusive language
Where machine learning approaches are
of concern,
There are several previous shared tasks
similar to offensive language detection. The
shared task on Aggression Identification
called ’TRAC’ provided participants a
dataset containing annotated Facebook posts and
comments in English and Hindi
After an extensive literature review, collection of additional previous datasets related to the topic and preliminary experiments; we started to experiments through shared tasks as described below. 2.1
A bi-LSTM neural network model that has
been developed
This model consists of a bidirectional Long Short-Term Memory Networks (biLSTM) model with an Attention layer on top. The model captures the most important semantic information in a tweet, including emojis and hashtags. A simplified schema of our model can be seen in the following figure.
Figura 1: Schema of the model
First, the tweets were tokenized removing punctuation marks and keeping emojis and full hashtags because can contribute to define the meaning of a tweet. Second, the embedding layer transforms each element in the tokenized tweet (such as words, emojis and hashtags) into a low-dimension vector. The embedding layer, composed of the vocabulary of the task, was randomly initialized from a uniform distribution (between -0.8 and 0.8 values and with 300 dimensions). Recent studies have reported that pre-trained word embeddings are far more satisfactory than the randomly initialized embeddings (Erhan et al., 2010; Kim, 2014). For that reason, the initialized embedding layer was updated with the word vectors included in a pre-trained model based on all the tokens, emojis and hashtags from 20M English tweets (Barbieri et al., 2016), which were updated during the training.
Then, a biLSTM layer gets high-level features from previous embeddings. The LSTM were introduced by Hochreiter and Schmidhuber (1997) and were explicitly designed to avoid the longterm dependency problem. LSTM systems keep relevant information of inputs by incorporating a loop enabling data to flow from one step to the following. LSTM gets a word embedding sequentially, left to right, at each time step, produces a hidden step and keeps its hidden state through time. Whereas, biLSTM does the same process as standard LSTM, but processes the text in a left to right as well as right-to-left order in parallel. Therefore, gives two hidden state as output at each step and is able to capture backwards and longrange dependencies.
A critical and apparent disadvantage of seq2seq models (such as LSTM) is that they compress all information into a fixed-length vector, causing the incapability of remembering long tweets. Attention mechanism aims to overcome the limitation of fixed-length vector keeping relevant information from long tweet sequences. In addition, attention techniques have been recently demonstrated success in multiple areas of the Natural Language Processing such as question answering, machine translations, speech recognition and relation extraction (Bahdanau et al., 2014; Hermann et al., 2015; Chorowski et al., 2015; Zhou et al., 2016). For that reason, we added an attention layer, which produces a weight vector and merge word-level features from each time step into a tweet-level feature vector, by multiplying the weight vector. Finally, the tweet-level feature vector produced by the previous layers is used for classification task by a fully-connected layer. Furthermore, we applied dropout regularization in order to alleviate overfitting. Dropout operation sets randomly to zero a proportion of the hidden units during forward propagation, creating more generalizable representations of data. As in Zhou et al. (2016), we employ dropout on the embedding layer, biLSTM layer and before the output layer. The dropout rate was set to 0.5 in all cases. 2.2
In this work, we presented a bi-LSTM model with two dense layers at the end. We have developed a system in the context of the shared task: MEX-A3T: Authorship and aggressiveness analysis in twitter. Specifically, the Aggressiveness Identification track, which focuses on the detection of aggressive comments in tweets from Mexican users and the other related IberLEF 2019 shared tasks.
We have used data from different tasks in order to train more examples in the model. As we believe that the tasks of humor and sentiment analysis could help in detecting aggressive language, we have selected three additional task to train with MEX-A3T at the same time. The other tasks were IroSva, that aims investigating the recognition of irony in Twitter messages in three different Spanish variants (from Spain, Mexico, and Cuba); HAHA which we used the classification task related to identify if a Spanish tweet is a joke or not and TASS 2019 that focuses on the evaluation of polarity classification systems of tweets written in Spanish. We used the data related to this task, tweets written in the Spanish language spoken in Spain, Peru, Costa Rica, Uruguay and Mexico, which were annotated with 4 different levels of opinion intensity (Positive, Negative, Neutral and Nothing).
Figura 2: Simplified schema of the multi- task model
In this scenario, we defined an Embedding layer for each Spanish variant in IroSva task. Classification tasks with the same Spanish variant used the same Embedding layer during the training process. Furthermore, all task shared the biLSTM layer during training. For the moment this approach was not very successful; however this may be due to lack of data to train the different models. 3
Despite the progress in this shared task, there are potential issues for the future work. Future experiments were planned mainly in 2 groups:
First, improvement areas will be
investigated for the efficiency of the classification
model developed for SemEval 2019 - Task 6
shared task, with the same dataset that is
called Offensive Language Identification
Dataset (OLID)
Initial experiments have been done taking only the words into account. Using additional features such as WordNet synsets, Part of Speech (POS) tags, frequencies, offensive word dictionaries and so on, is expected to improve the precision of the results.
Furthermore, changes in the methodology
such as applying ’Bidirectional Encoder
Representations from Transformers’
Secondly, in a later phase of the study, it is planned to obtain a new dataset using Twitter’s streaming API and crowdannotation and using the new dataset for the experiments including the metadata such as usersession time, whether it is a reply or a retweet.
For this purpose, first of all, a set of specific hashtags will be decided with a high potential of being associated with offensive tweets.
After pulling the data and deciding the annotation scheme, the data will be presented for crowd annotation.
After compilation of a corpus, model training will be carried out with the system given the most promising results for the OLID dataset.
Additional improvements for the system design and other potential features will be experimented considering the performance of the preliminary tests. 4
The main research questions that are intended to answer with this work are as the following:
•What algorithms are those that provide us with greater accuracy to identify offensive language in a text?
•What characteristics should be taken into account in the process of analysing text in terms of aggressiveness?
•What type of metadata would be useful to increase the accuracy while analysing the text?
•Finally, how would be the overall system for this classification task that will bring the highest accuracy? 5
The main objectives of the research can be listed as follows:
I.Executing preliminary experiments to classify offensive messages in social media (particularly tweets and short messages) datasets.
There are several published datasets belonging previous researches that is annotated as Offensive or within the similar context such as cyberbullying, hate speech related, misogyny 1,2,3.According to the specific annotation scheme and the content, handcrafted features might have an important parameter for the performance. Experimenting on these previous datasets will help understanding the strengths and weaknesses of different design specifications and features and eventually help optimization of them.
II.Experiments to improve the performance of the current system with fine-tuned system design and feature engineering.
The neural network system for the initial experiments took only words into account. However, there is a potential to improve the results of this system with additional feature extraction. Furthermore, detailed analysis on integration of linguistic annotations into neural network and other models like convolution can be considered to improve the performance.
III.Creating a new dataset with crowd annotation. There are several crowd annotation platforms such as: Mechanical Turk4 , crowdflower5 , crowdtruth6 . By uploading the data and deciding the rules of annotation these platforms help annotating the data by human annotators.
To crowd-annotate tweet data, first of all, the data will be pulled from Twitter API according to certain hashtags. Hashtags will be decided for certain contexts such as political debate hashtags or hashtags related to sportive rivalry. After that, annotation schema will be decided. Annotation schema of previous datasets are usually in hierarchical order and contains additional information such as target or for instance if it contains aggression whether it is cyberbullying or not.
IV.Experiments on the new dataset with various approaches on the system and features.
A new dataset can give the opportunity to 1https://www.kaggle.com/alternacx/hateoffensivespeechdetection 2https://www.amnesty.org/en/ 3https://zenodo.org/record/1184178.XTBv2pMzaRt 4https://www.mturk.com/ 5https://www.figure-eight.com/ 6http://crowdtruth.org/ reproduce previous well-performed systems designed. Moreover, majority of the related datasets published do not include metadata. With the new dataset collected through Twitter API it will be possible to obtain metadata, as well. Therefore, user-related features such as the frequency of profanity in previous messages can be obtained and it would help understand the importance of metadata on the performance.