Misogynistic Tweet Detection: Modelling CNN with Small Datasets
aa r X i v : . [ c s . C L ] A ug Misogynistic Tweet Detection: Modelling CNNwith Small Datasets
Md Abul Bashar , Richi Nayak , Nicolas Suzor , and Bridget Weir School of Electrical Engineering and Computer Science School of LawQueensland University of Technology, Brisbane, Australia { m1.bashar, r.nayak, n.suzor, bridget.weir } @qut.edu.au Abstract.
Online abuse directed towards women on the social mediaplatform Twitter has attracted considerable attention in recent years.An automated method to effectively identify misogynistic abuse couldimprove our understanding of the patterns, driving factors, and effective-ness of responses associated with abusive tweets over a sustained timeperiod. However, training a neural network (NN) model with a small setof labelled data to detect misogynistic tweets is difficult. This is partlydue to the complex nature of tweets which contain misogynistic content,and the vast number of parameters needed to be learned in a NN model.We have conducted a series of experiments to investigate how to train aNN model to detect misogynistic tweets effectively. In particular, we havecustomised and regularised a Convolutional Neural Network (CNN) ar-chitecture and shown that the word vectors pre-trained on a task-specificdomain can be used to train a CNN model effectively when a small setof labelled data is available. A CNN model trained in this way yields animproved accuracy over the state-of-the-art models.
Incidents of abuse, hate, harassment and misogyny have proliferated with thegrowing use of social media platforms (e.g. Twitter, Facebook, Instagram). Theseplatforms have generated new opportunities to spread online abuse [5]. Theexperience of online abuse is a common occurrence for women [12]. Often theseexperiences of online abuse can be categorised as sexist or misogynistic in nature,and can include name-calling and offensive language, threats of harm or sexualviolence, intimidation, shaming, and the silencing of women. While it is easy toidentify instances of abuse and major abusive campaigns on social media, it isdifficult to understand changes in levels of abuse over time, and almost impossibleto identify the effectiveness of interventions by platforms in combating abuse[24].An automated system to identify abusive tweets could help in ongoing efforts todevelop effective remedies.A key challenge in the automatic detection of misogynistic abusive tweets isunderstanding the context of an individual tweet. We focus here on misogynistictweets that are abusive towards an individual or group – a subset of the largercategory of tweets that include sexist or misogynistic words or concepts. Accord-ingly, this study sought to address the difficult task of separating abusive tweetsfrom tweets that were sarcastic, joking, or contained misogynistic keywords in
Bashar et al. a non-abusive context. A lexical detection approach tends to have low accuracy[4,28] because they classify all tweets containing particular keywords as misog-ynistic. Xiang et al. [28] reported that bag-of-words, part-of-speech (POS) andbelief propagation did not work well for the detection of profane tweets becauseof the significant noise in tweets. For example, tweets do not follow a standardlanguage format, words are often misspelled or altered, and tweets often includewords from local dialects or foreign languages. The automated algorithms shouldlook for patterns, sequences, and other complex features that are present, despitethe noise, and are correlated with misogynistic tweets. Traditional algorithms(e.g. Random Forest, Logistic Regression and Support Vector Machines) rely onmanual process to obtain these kinds of features and are limited by the kinds offeatures available. Neural Network (NN)-based models, on the other hand, canautomatically learn complex features and effectively use them to classify a giveninstance.Relying on a sufficiently large training dataset, CNN models have shown to beeffective in Natural Language Processing (NLP) tasks such as semantic parsing[29], document query matching [22], sentence classification [13], etc. A set ofconvolving filters are applied to local features (e.g. words) to learn patternssimilar to n Grams. Local features are commonly represented as word vectorswhere words are projected from a sparse representation onto a lower dimensionalvector space. Word vectors essentially encode semantic features of each word ina fixed number of abstract topics (or dimensions). The apparent success of CNNin NLP tasks can be credited to its capability to learn text patterns in semanticspace. However, given the requirement of setting the large number of parametersin CNN, often in millions, the CNN models are trained on a huge labelled dataset.In general, this is a limitation of any NN-based model [6]. Overall, curating alarge set of labelled tweets containing misogynistic abuse is difficult and costlyto achieve due to the large amount of data that needs to be manually examinedto rigorously identify abusive tweets.However, word vectors trained on a general-purpose corpus cannot capturethe task-specific semantics because the nature of general-purpose corpus andthe misogynistic tweets is completely different. For example, many words usedin these tweets are linguistically specific and unique to Twitter-based discussion,and are not covered in a general-purpose corpus. Consequently, a CNN classi-fier model built using these word vectors cannot adequately detect misogynisticabusive tweets.In this paper, we investigate the effectiveness of various corpus to generatepre-trained word vectors for building a CNN model when there is a small setof labelled data available. In particular, we trained a CNN using a small setof labelled data to detect misogynistic tweets. We pre-trained word vectors on0.2 billion unlabelled tweets that contain at least one misogynistic keyword (i.e. whore, slut, rape ). We customised and regularised the CNN architecture usedin [13]. On the test dataset, the trained CNN model achieves significantly bet-ter results than the state-of-the-art models. It is better by a large margin incomparison to the CNN models build on word vectors pre-trained on a size- isogynistic Tweet Detection 3 able general corpus. The experimental results show that a CNN classifier can betrained on a small labelled tweet data, provided that the word vectors are pre-trained in the context of the problem domain and a careful model customisationand regularisation is performed.This project investigates how to effectively apply data mining methods, witha focus on training a NN model, to detect misogynistic tweets. Three maincontributions of this paper are: (a) it shows that word vectors pre-trained ona task-specific domain can be used to effectively train CNN when a small setof labelled data is available; (b) it shows how to customise and regularise aCNN architecture to detect misogynistic tweets; and (c) finally, we present anautomated data mining method to detect misogynistic abusive tweets.
Misogynistic abusive tweet detection falls into the research area of text classifi-cation. Popular text classification algorithms used in hate speech and offensivelanguage detection are Naive Bayes [4], Logistic Regression [4], Support VectorMachine (SVM) [4,26,28] and Random Forest [4,28]. Performance of these algo-rithms depend on feature engineering and feature representation [4,28] . Therehave been some works where syntactic features are leveraged to identify the tar-gets and the intensity of hate speech. Examples of these features are relevantverb and noun occurrences (e.g. kill and
Jews ) [7], and the syntactic structures:I < intensity >< user intent >< hate target > (e.g. I f ∗ cking hate white people ) [23].Misogynistic tweet detection is challenging for text classification methodsbecause social media users very commonly use offensive words or expletives intheir online dialogue [25]. For example, the bag-of-words approach is straight-forward and usually has a high recall, but it results in higher number of falsepositives because the presence of misogynistic words causes these tweets to bemisclassified as abusive tweets [14].Recently, neural network-based classifiers have become popular as they auto-matically learn abstract features from the given input feature representation [1].Input to these algorithms can be various forms of feature encoding, includingmany of those used in the classic methods. Algorithm design in this categoryfocuses on the selection of the network topology to automatically extract use-ful abstract features. Popular network architectures are CNN, Recurrent NeuralNetworks (RNN) and Long Short-Term Memory network (LSTM). CNN is wellknown for extracting patterns similar to phrases and n Grams [1]. On the otherhand, RNN and LSTM are effective for sequence learning such as order infor-mation in text [1]. The CNN model has been successfully used for sentenceclassification [13] . To effectively identify patterns in the text, they used wordembedding pre-trained on Google News corpus while training a CNN model onthe labelled dataset.The neural network-based classifiers have not yet been applied in misog-ynistic tweet detection. It requires a rigorous investigation as to what extentpatterns and orderly information are present in misogynistic tweets, and howwe can optimise a Neural Network for classification accuracy. There are manyCNN architectures used in the current literature, but the design of an archi-tecture heavily depends on the problem at hand. Therefore, a customised CNN
Bashar et al. architecture is needed to classify misogynistic tweets. It also remains to be seenwhether the word embedding can be sensitive to the domain knowledge of thecorpus. Does the word embedding need to be trained on a similar tweet streamto capture contextual properties?
Misogynistic abusive tweets may contain misogynistic keywords, but tweets canalso be misogynstic abuse without explicitly containing these slurs. Further, notall tweets that contain misogynistic keywords are abusive. Classifying misogy-nistic abuse in tweets requires close reading, and even humans can struggle toclassify these tweets accurately. The focus of this research is to detect abusivetweets that contain misogynistic words. A previous study has identified thatthree keywords – whore , slut and rape – are useful in identifying a substan-tial portion of misogynistic tweets [2]. However, these misogynistic words arecommonly used in tweets that are not abusive, and separating abusive tweetsfrom non-abusive tweets is difficult when we base our classification purely onthe occurrence of these words. We propose a two-step method to approach thisproblem: – Pre-filtering: We pre-filter tweets that contain any of the three main misog-ynistic keywords (slut, rape, whore) to find potentially-misogynistic tweets. – Training a CNN model: Using a small labelled data set, a CNN model istrained to classify the remaining potentially-abusive tweets. We propose sev-eral methods to accurately train the model.The research team used a systematic approach to generate the labelled datamanually. The following contextual information was used in assessing whethera tweet contains targeted misogynistic abuse, or not: (a) Is a specific person orgroup being targeted in this tweet? (b) Does this tweet contain a specific threator wish for violence? (c) Does this tweet encourage or promote self-harm orsuicide? (d) Is the tweet harassing a specific person, or inciting others to harassa specific person? (e) Does the tweet use misogynistic language in objectifyinga person, making sexual advances, or sending sexually explicit material? (f) Isthe tweet promoting hateful conduct by gender, sexual orientation, etc.?The labelled tweets reveal many challenges that need to be addressed to traina classifier effectively. These included: (a) The misogynistic words are not thediscriminatory words. Many keywords are overlapping between misogynistic andnon-misogynistic tweets, especially misogynistic keywords. (b) Words may bemisspelt and spelt in many ways. (c) Sometimes people mix words from localdialects or foreign languages. (d) The data is noisy and does not follow a standardlanguage sequence (format). (e) Effectively detecting misogynistic tweets needsaccess to semantics and context information that is often not available, e.g. itis difficult to use dictionary-based semantics for the nature of noise in tweetsand difficult to know the context because of the small length of a tweet. (f) Thelabelling process is time consuming and it is extremely difficult to generate alarge quantity of labelled data because only a very small portion of tweets canbe identified as misogynistic.Given these challenges, in this paper, we investigate how to effectively traina CNN model with a small set of labelled data to detect misogynistic abusive isogynistic Tweet Detection 5 tweets. We train a CNN on top of the pre-trained word vectors (a.k.a. wordembedding or distributed representation of words). The primary focus is to findout what kind of pre-trained word vectors is useful to train a CNN with asmall dataset. Another two important focuses are to find out what customisedarchitecture of CNN is effective in the given problem and to test the effectivenessof some simple data and feature augmentation.
Word embedding models map each word from the vocabulary to a vector ofreal numbers. They aim to quantify and categorise semantic similarities betweenwords based on their distributional property based on the premise that a wordis characterised by the company it keeps. Given a sizeable unlabelled corpus,these models can effectively learn a high-quality word embedding. Based on thefeed-forward neural network, Mikolov et al. [20] proposed two popular models:Skip-gram and Continuous Bag-of-Words as shown in Figure 1.Given the words within a sliding window, the continuous bag-of-words modelpredicts the current word w i from the surrounding context words C , i.e. p ( w i | C ).In contrast, the skip-gram model uses the current word w i to predict the sur-rounding context words C , i.e. p ( C | w i ). In Figure 1, for example, if the currentposition of a running sliding window contains the phrase she looks like a crackwhore . In continuous bag-of-words, the context words { she, looks, like, a, whore } can be used to predict the current word { crack } , whereas, in skip-gram, the cur-rent word { crack } can be used to predict the context words { she, looks, like, a,whore } . w i-2 w i-1 w i+1 w i+2 w i-2 w i-1 w i+1 w i+2 ∑ w i ∑ w i Input Projection Output Input Projection OutputContinuous bag-of-words Skip-gram S li d i ng W i ndo w S li d i ng W i ndo w Fig. 1: Word Embedding Models
The training objective is to find a word embedding that maximises p ( t i | C )or p ( C | t i ) over a training dataset. In each step of training, each word is either(a) pulled closer to the words that co-occur with it or (b) pushed away fromall the words that do not co-occur with it. A softmax or approximate softmax function can be used to achieve this objective [20]. At the end of the training,the embedding brings closer not only the words that are explicitly co-occurringin a training dataset, but also the words that implicitly co-occur. For example, if t explicitly co-occurs with t and t explicitly co-occurs with t , then the modelcan bring closer not only t to t , but also t to t . The continuous bag-of-words model is faster and has slightly better accuracy for the words that appearfrequently. Therefore, we use this model in this research. Bashar et al.
We empirically customise and regulate Kim’s [13] CNN architecture to detectmisogynistic tweets and reduce overfitting. Figure 2 shows the architecture. Weuse word embedding to represent each word w in an n -dimensional word vector w ∈ R n . A tweet t with m words is represented as a matrix t ∈ R m × n . Convolu-tion operation is applied to the tweet matrix with one stride. Each convolutionoperation applies a filter f i ∈ R h × n of size h . Empirically, based on the accuracyimprovement in ten-fold cross validation, we used 256 filters for h ∈ { , } and512 filters for h ∈ { } . The convolution is a function c ( f i , t ) = r ( f i · t k : k + h − ),where t k : k + h − is the k th vertical slice of the tweet matrix from position k to k + h − f i is the given filter and r is a ReLU function. The function c ( f i , t )produces a feature c k similar to n Grams or phrases for each slice k , resultingin m − n + 1 features. We apply the max-pooling operation over these featuresand take the maximum value, i.e. ˆ c i = max c ( f i , t ). Max-pooling is carried tocapture the most important feature for each filter. As there are a total of 1024filters (256+256+512) in the proposed model, the 1024 most important featuresare learned from the convolution layer.These features are passed to a fully connected hidden layer with 256 percep-trons that use the ReLU activation function. This fully connected hidden layerallows learning the complex non-linear interactions between the features fromthe convolution layer and generates 256 higher level new features. Finally these256 higher level features are passed to the output layer with single perceptronthat uses the sigmoid activation function. The perceptron in this layer generatesthe probability of the tweet being misogynistic.We randomly dropout a proportion of units from each layer except the outputlayer by setting them to zero. This is done to prevent co-adaptation of units ina layer and to reduce overfitting. We empirically dropout 50% units from theinput layer, the filters of size 3 and the fully connected hidden layer. We dropoutonly 20% units from the filters of size 4 and 5. shelookslikeacrackwhoreIhateher Tweet Matrix Convolution Max Pooling Concatenate Fully Connected Layers
Fig. 2: CNN Model Architecture
The primary objectives of the experiments are to show: (a) word vectors pre-trained on a task-specific domain is more effective than those pre-trained ona sizeable general corpus; (b) CNN trained on a small dataset and built onword vectors pre-trained on a task-specific domain can perform better than thestate-of-the-art models; and (c) the impact of some simple data and word aug-mentation techniques on training a CNN model. isogynistic Tweet Detection 7
We collected tweets using Twitter’s streaming API. For thelabelled dataset, we identified 10k tweets that contain any of the three mainmisogynistic keywords (i.e., whore, slut, rape). Following the misogynistic tweetdefinition in Section 3, the research team labelled a total of 5000 tweets with1800 misogynistic and 3200 non-misogynistic labels. A stratified data selectionwas made to reduce a trained models’ bias to a specific label, i.e. we kept 1800misogynistic and 1800 randomly selected nonmisogynistic tweets. We used 80%examples for training and 20% for testing. We used the ten-fold cross-validationto tune hyperparameters. We used the Porter’s suffix-stripping algorithm forpreprocessing.The tweet labelling method has the following limitations: (a) The coding isbased on a literal interpretation of the text; with limited context, we are likelyto include some sarcasm or humour. (b) We are only labelling tweets written inEnglish. (c) Identifying the tweets by keywords only, we will not catch abuse thatappears to be ordinary misogyny, e.g. get back in the kitchen . (d) Identifying thetweets by keywords only, we will not identify harassment that is targeted andorganised harassment, either ongoing over time or involving many participants,but does not use one of our keywords.
WikiNews:
Word vectors of 300-dimension pre-trained on the Wikipedia 2017,UMBC webbase corpus and statmt.org news datasets containing a total of 16billion words using fastText (a library for learning word embeddings created byFacebook’s AI Research lab) [19].
GoogleNews:
Word vectors of 300-dimension pre-trained on Google News cor-pus containing a total of three billion words using the Continuous Bag-of-WordsWord2vec model [18]. Potentially Misogynistic Tweets:
Word vectors of 200-dimension pre-trainedon 0.2 billion tweets that contain any of the three main misogynistic keywords.A Continuous Bag-of-Words Word2vec model is used in pre-training while min-imum count for word is set to 100.
We used six standard evaluation measures of classification performance: Accu-racy, Precision, Recall, F Score, Cohen Kappa (CK) and Area Under Curve(AUC). We also report True Positive (TP), True Negative (TN), False Positive(FP) and False Negative (FN) values.
We have implemented eight baseline models to compare the performance withthe proposed CNN model. – Long Short-Term Memory Network (LSTM) [10]. We have implementedLSTM with 100 units, 50% dropout, binary cross-entropy loss function,Adam optimiser and sigmoid activation. https://bit.ly/2esteWf Bashar et al. – Feedforward Deep Neural Network (DNN) [8]. We have implemented DNNwith five hidden layers, each layer containing eight units, 50% dropout ap-plied to the input layer and the first two hidden layers, softmax activa-tion and 0.04 learning rate. For all neural network based models (CNN,LSTM, DNN), hyperparameters are manually tuned based on ten-fold cross-validation. – Non NN models including Support Vector Machines (SVM) [9], RandomForest [17], XGBoost (XGB) [3], Multinomial Naive Bayes (MNB) [16], k-Nearest Neighbours (kNN) [27] and Ridge Classifier (RC) [11]. Hyperpa-rameters of all these models are automatically tuned using ten-fold cross-validation and GridSearch from sklearn.
We conducted experiments to see the ef-fects of different word embeddings in training the CNN model. A summary of theembeddings is given in Table 3 and the experimental results are given in Table 4.Three main observations from the results are: (a) Word vectors pre-trained on alarge dataset (e.g., WE1, WE2, WE4, WE5) always improves performance. Theconvolution layer, that captures n Gram-like patterns in the tweets while usingword vectors to represent the tweets, allows the model to find these patternsin semantic space. The pre-trained word embedding can provide the semanticsof words that have fewer appearances in the training dataset. This reinforcesthe prior finding [21] that the features obtained from a pre-trained deep learningmodel perform well on a variety of tasks. (b) Updating the word vectors with thelabelled data while training the classifier improves the performance (e.g., WE1over WE2). This allows the semantics of words to be more focused over thetraining set. (c) Word vectors pre-trained on potentially misogynistic tweets andupdated with labelled data performs the best. It improves the CNN accuracy byaround 12% compared with word vectors pre-trained on a standard corpus (e.g.Google News corpus). This observation challenges the previous findings [13] thatgeneral pre-trained word vectors (e.g. word vectors pre-trained on Google News)are universal feature extractors. Due to the small labelled dataset used in train-ing the CNN model, it was not enough to update the necessary word vectors forthe problem domain, given that tweets are very noisy and mostly different fromstandard corpora like Google News or Wikipedia. The word vectors pre-trainedon unlabelled datasets in the task-specific domain can address this problem.The apparent performance correlation of CNN and word vector can be re-lated to the similar goal that CNN and word vector have. In the word vectorrepresentation, semantically similar words are represented with similar vectorsand semantically dissimilar words are represented with dissimilar vectors. Thisis obtained through training the word vectors on a corpus where it searches forco-occurring words through a filter called sliding window. To train CNN with alabelled tweet, the words in the tweet are represented with word vectors. CNNdiscovers patterns in these word vectors through varying length filters, where apattern identifies something similar to a set of words that co-occur in the la-belled tweets. It can be ascertained that both CNN and word vector capture the isogynistic Tweet Detection 9 patterns of co-occurring words. Because CNN learns the patterns in the vectorspace, it harnesses the patterns (or semantic relations) already learned in thevector space. Thus, pre-trained word vectors, especially trained on a corpus fromthe similar nature domain, may significantly improve the performance of CNNmodel when only a small labeled dataset is available for training. Pre-trainedword vectors acts as the smoothing used in many language models [30].
Fig. 3: Summary of Word Embeddings
Model DescriptionWE1 W2V pre-trained on potentially abusivetweets and updated with labelled dataWE2 W2V pre-trained on potentially abusivetweets but not updated with labelled dataWE3 W2V Trained with only labelled dataWE4 W2V pre-trained on google news and updatedwith labelled dataWE5 fastText pre-trained on Wikipedia pages andupdated with labelled data
Fig. 4: Performance of CNN appliedon different Word Embeddings
WE1 WE2 WE3 WE4 WE5TP
264 194 217 199TN
279 273 274 281FP
82 88 87 80FN
97 167 144 162Accuracy Score
Classifier Models Comparison
We implemented the proposed CNN modeland the eight baseline models to detect misogynistic tweets. Guided by the ex-perimental results in previous section, both CNN and LSTM models were builton word vectors that are pre-trained on potentially abusive tweets and updatedwith the labelled dataset during the classifier training. Performances of the mod-els are summarised in Table 1.Result shows that CNN outperforms all other models. For example, the im-provement in precision, accuracy, Cohen Kappa score and AUR of CNN over thesecond best performing model LSTM are 6.120%, 4.364%, 13.855% and 4.364%respectively. LSTM is known to be effective in text datasets and the results re-flect this. The reason for CNN outperforming LSTM and other baseline modelsmight be the nature of tweets. Tweets are super condensed texts, full of noise andoften do not follow the standard sequence of the language. Traditional models(e.g. RF, SVC, kNN, etc.) are based on bag-of-words representation that can behighly impacted by the significant noise in tweets [28]. Besides, the bag-of-wordsrepresentation cannot capture sequences and patterns that are very important toidentify a misogynistic tweet. For example, if a tweet contains a sequence if youknow what I mean , there is a high chance that this tweet might be misogynistic,even though individual keywords are innocent.The performance of LSTM is better than traditional models as it can cap-ture sequences. However, sequences in tweets often get altered by noises (e.g.misspelled or intentionally altered by the author); therefore LSTM might strug-gles to detect misogynistic tweets. CNN models are well known for effectivelydiscovering a large number of patterns and sub-patterns through many filterswith varying size. If a few words of a given tweet are altered by noise it can stillmatch a sub-pattern. This means CNN is less affected by noise. As a result CNNout performs LSTM.CNN is popularly used in Computer Vision and is known to be effective onlyif the model is trained on massive datasets. However, in this research, we trained a simple CNN with only three thousand labelled tweets. This simple CNN usesonly one layer of convolutions on top of word vectors, and it achieves significantlybetter results than state-of-the-art models. These results ascertain that a CNNcan be trained on a small labelled dataset, provided that word vectors are pre-trained in the context of the problem domain, and a careful model customisationand some regularisations are performed.
Table 1: Performances of Classification Models
CNN LSTM DNN SVC RF XGB MNB kNN RCTP
264 275 257 279 286 272 95 263TN
263 171 244 229 223 251 302 245FP
98 190 117 132 138 110 59 116FN
97 86 104 82 75 89 266 98Accuracy Score
Data and Word Augmentation Performances
Data augmentation anddocument expansion is popularly used in computer vision and information re-trieval respectively to artificially inflating a small labelled dataset and/or inputvectors. In this paper, we augmented/expanded the data multiple ways andstudied their impact on training the CNN model. We used two sources of datato generate augmented data: (1) the word vectors pre-trained on the potentialmisogynistic tweets; and (2) topics identified by Non-Negative Matrix Factori-sation (NMF) [15] on the tweet training dataset, performed separately on eachclass. A total of six policies were followed. AT1: Words in a labelled tweet are ran-domly replaced by semantically similar words from word vector space to createan artificial tweet. AT2: Discriminative Words in a labelled tweet are randomlyreplaced by semantically similar words from word vector space to create an ar-tificial tweet. A discriminative word is a word that more frequently appears inthe tweet of a specific label. AT3: A tweet is expanded by adding its semanti-cally similar words found from word vector space. AT4: A tweet is expanded byadding its semantically similar words found from NMF. AT5: Use the discoveredtopics in NMF as artificial tweets. AT6: A set of words from word vector spacethat is semantically similar to a tweet is used as an artificial tweet.Table 2 reports the performance of model trained with each of these aug-mentation policies and the CNN model trained with the original labelled datasetbefore any augmentation (labelled as AT0). The experimental results show thatthese ways of augmentation do not improve the accuracy. We conjecture thatadditional external features (i.e. words) may distort the patterns exist in theoriginal tweets, since the CNN classifier largely depends on learning these pat-terns, the performance degrades.
This paper presents a novel method of misogynistic tweet detection using wordembedding and the CNN model when only a small amount of labelled data isavailable. We report the results of a series of experiments conducted to investi-gate the effectiveness of training a model with a small dataset. We customised isogynistic Tweet Detection 11Table 2: CNN results from Data Augmentation Policies
AT0 AT1 AT2 AT3 AT4 AT5 AT6TP 267 270 242 301 281 292 288TN 283 275 287 224 203 238 241FP 78 86 74 137 158 123 120FN 94 91 119 60 80 69 73Accuracy 0.762 0.755 0.733 0.727 0.670 0.734 0.733Precision 0.774 0.758 0.766 0.687 0.640 0.704 0.706Recall 0.740 0.748 0.670 0.834 0.778 0.809 0.798F Score 0.756 0.753 0.715 0.753 0.703 0.753 0.749CK 0.524 0.510 0.465 0.454 0.341 0.468 0.465AUC 0.762 0.755 0.733 0.727 0.670 0.734 0.733 and regularised a CNN architecture, and it performs better than the state-of-the-art models, provided that the CNN is built on word vectors pre-trained onthe task-specific domain. Experimental results show that a CNN model built onword vectors pre-trained on the task-specific unlabelled dataset is more effectivethan built on word vectors pre-trained on a sizeable general corpus. Experimen-tal results also show that simple data augmentation policies are not adequate toimprove misogynistic tweet detection performance in the CNN model.
This research was fully supported by the QUT IFE Catapult fund. Suzor is therecipient of an Australian Research Council DECRA Fellowship (project numberDE160101542).
References
1. Badjatiya, P., Gupta, S., Gupta, M., Varma, V.: Deep learning for hate speechdetection in tweets. In: Proceedings of the 26th International Conference on WorldWide Web Companion, pp. 759–760. International World Wide Web ConferencesSteering Committee (2017)2. Bartlett, J., Norrie, R., Patel, S., Rumpel, R., Wibberley, S.: Misogyny on twitter.Demos (2014)3. Chen, T., Guestrin, C.: Xgboost: A scalable tree boosting system. In: Proceedingsof the 22nd acm sigkdd international conference on knowledge discovery and datamining, pp. 785–794. ACM (2016)4. Davidson, T., Warmsley, D., Macy, M., Weber, I.: Automated hate speech detectionand the problem of offensive language. arXiv preprint arXiv:1703.04009 (2017)5. Dragiewicz, M., Burgess, J., Matamoros-Fern´andez, A., Salter, M., Suzor, N.P.,Woodlock, D., Harris, B.: Technology facilitated coercive control: domestic violenceand the competing roles of digital media platforms. Feminist Media Studies pp.1–17 (2018)6. Fadaee, M., Bisazza, A., Monz, C.: Data augmentation for low-resource neuralmachine translation. In: Proceedings of the 55th Annual Meeting of the Associationfor Computational Linguistics (Volume 2: Short Papers), vol. 2, pp. 567–573 (2017)7. Gitari, N.D., Zuping, Z., Damien, H., Long, J.: A lexicon-based approach for hatespeech detection. International Journal of Multimedia and Ubiquitous Engineering (4), 215–230 (2015)8. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforwardneural networks. In: Proceedings of the thirteenth international conference onartificial intelligence and statistics, pp. 249–256 (2010)9. Hearst, M.A., Dumais, S.T., Osuna, E., Platt, J., Scholkopf, B.: Support vectormachines. IEEE Intelligent Systems and their applications (4), 18–28 (1998)2 Bashar et al.10. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation (8), 1735–1780 (1997)11. Hoerl, A.E., Kennard, R.W.: Ridge regression: applications to nonorthogonal prob-lems. Technometrics (1), 69–82 (1970)12. International, A.: Toxic twitter - a toxic place for women (2018). URLhttps://bit.ly/2FZYQhV13. Kim, Y.: Convolutional neural networks for sentence classification. arXiv preprintarXiv:1408.5882 (2014)14. Kwok, I., Wang, Y.: Locate the hate: Detecting tweets against blacks. In: AAAI(2013)15. Lee, D.D., Seung, H.S.: Algorithms for non-negative matrix factorization. In: Ad-vances in neural information processing systems, pp. 556–562 (2001)16. Lewis, D.D.: Naive (bayes) at forty: The independence assumption in informationretrieval. In: European conference on machine learning, pp. 4–15. Springer (1998)17. Liaw, A., Wiener, M., et al.: Classification and regression by randomforest. R news (3), 18–22 (2002)18. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word repre-sentations in vector space. International Conference on Learning Representations(ICLR) Workshop (2013)19. Mikolov, T., Grave, E., Bojanowski, P., Puhrsch, C., Joulin, A.: Advances in pre-training distributed word representations. In: Proceedings of the InternationalConference on Language Resources and Evaluation (LREC 2018) (2018)20. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed repre-sentations of words and phrases and their compositionality. In: Advances in neuralinformation processing systems, pp. 3111–3119 (2013)21. Sharif Razavian, A., Azizpour, H., Sullivan, J., Carlsson, S.: Cnn features off-the-shelf: an astounding baseline for recognition. In: Proceedings of the IEEEconference on computer vision and pattern recognition workshops, pp. 806–813(2014)22. Shen, Y., He, X., Gao, J., Deng, L., Mesnil, G.: Learning semantic representationsusing convolutional neural networks for web search. In: Proceedings of the 23rdInternational Conference on World Wide Web, pp. 373–374. ACM (2014)23. Silva, L.A., Mondal, M., Correa, D., Benevenuto, F., Weber, I.: Analyzing thetargets of hate in online social media. In: ICWSM, pp. 687–690 (2016)24. Suzor, N., Van Geelen, T., Myers West, S.: Evaluating the legitimacy of platformgovernance: A review of research and a shared research agenda. InternationalCommunication Gazette (4), 385–400 (2018)25. Wang, W., Chen, L., Thirunarayan, K., Sheth, A.P.: Cursing in english on twitter.In: Proceedings of the 17th ACM conference on Computer supported cooperativework & social computing, pp. 415–425. ACM (2014)26. Warner, W., Hirschberg, J.: Detecting hate speech on the world wide web. In:Proceedings of the Second Workshop on Language in Social Media, pp. 19–26.Association for Computational Linguistics (2012)27. Weinberger, K.Q., Saul, L.K.: Distance metric learning for large margin nearestneighbor classification. Journal of Machine Learning Research10