| { | |
| "paper_id": "2020", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:52:26.806633Z" | |
| }, | |
| "title": "Towards Non-Toxic Landscapes: Automatic Toxic Comment Detection Using DNN", | |
| "authors": [ | |
| { | |
| "first": "Ashwin", | |
| "middle": [], | |
| "last": "Geet D'sa", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "CNRS", | |
| "location": { | |
| "postCode": "F-54000", | |
| "settlement": "Nancy", | |
| "region": "Inria, LORIA", | |
| "country": "France" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Irina", | |
| "middle": [], | |
| "last": "Illina", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "CNRS", | |
| "location": { | |
| "postCode": "F-54000", | |
| "settlement": "Nancy", | |
| "region": "Inria, LORIA", | |
| "country": "France" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Dominique", | |
| "middle": [], | |
| "last": "Fohr", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "CNRS", | |
| "location": { | |
| "postCode": "F-54000", | |
| "settlement": "Nancy", | |
| "region": "Inria, LORIA", | |
| "country": "France" | |
| } | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "The spectacular expansion of the Internet has led to the development of a new research problem in the field of natural language processing: automatic toxic comment detection, since many countries prohibit hate speech in public media. There is no clear and formal definition of hate, offensive, toxic and abusive speeches. In this article, we put all these terms under the umbrella of \"toxic speech\". The contribution of this paper is the design of binary classification and regression-based approaches aiming to predict whether a comment is toxic or not. We compare different unsupervised word representations and different DNN based classifiers. Moreover, we study the robustness of the proposed approaches to adversarial attacks by adding one (healthy or toxic) word. We evaluate the proposed methodology on the English Wikipedia Detox corpus. Our experiments show that using BERT fine-tuning outperforms feature-based BERT, Mikolov's and fastText representations with different DNN classifiers.", | |
| "pdf_parse": { | |
| "paper_id": "2020", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "The spectacular expansion of the Internet has led to the development of a new research problem in the field of natural language processing: automatic toxic comment detection, since many countries prohibit hate speech in public media. There is no clear and formal definition of hate, offensive, toxic and abusive speeches. In this article, we put all these terms under the umbrella of \"toxic speech\". The contribution of this paper is the design of binary classification and regression-based approaches aiming to predict whether a comment is toxic or not. We compare different unsupervised word representations and different DNN based classifiers. Moreover, we study the robustness of the proposed approaches to adversarial attacks by adding one (healthy or toxic) word. We evaluate the proposed methodology on the English Wikipedia Detox corpus. Our experiments show that using BERT fine-tuning outperforms feature-based BERT, Mikolov's and fastText representations with different DNN classifiers.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "The past few years have seen a tremendous rise in the usage of Internet and social networks. Unfortunately, the dark side of this growth is an increase in toxic speech. Toxic speech is a type of offensive communication mechanism. Toxic speech can target different societal characteristics such as gender, religion, race, disability, etc. (Delgado and Stefancic, 2014) and reflects a certain \"state of society\". There is no uniform definition of toxic speech in the scientific literature and there is no clear distinction between hate, offensive, toxic and abusive speech (Gr\u00f6ndahl et al., 2018; Waseem et al., 2017; Davidson et al., 2017) . We refer to these collectively with the generic term of toxic speech.", | |
| "cite_spans": [ | |
| { | |
| "start": 338, | |
| "end": 367, | |
| "text": "(Delgado and Stefancic, 2014)", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 571, | |
| "end": 594, | |
| "text": "(Gr\u00f6ndahl et al., 2018;", | |
| "ref_id": "BIBREF8" | |
| }, | |
| { | |
| "start": 595, | |
| "end": 615, | |
| "text": "Waseem et al., 2017;", | |
| "ref_id": "BIBREF19" | |
| }, | |
| { | |
| "start": 616, | |
| "end": 638, | |
| "text": "Davidson et al., 2017)", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "Manually monitoring and moderating the Internet and social media content to identify and remove toxic speech is extremely expensive. This article aims at designing methods for automatic toxic speech detection on the Internet. Despite the studies already published on this subject, the results show that the task remains very difficult (Nobata et al., 2016; Saleem et al., 2017) . In this paper, we use semantic content analysis methodologies from Natural Language Processing (NLP) and methodologies based on Deep Neural Networks (DNN).", | |
| "cite_spans": [ | |
| { | |
| "start": 335, | |
| "end": 356, | |
| "text": "(Nobata et al., 2016;", | |
| "ref_id": "BIBREF13" | |
| }, | |
| { | |
| "start": 357, | |
| "end": 377, | |
| "text": "Saleem et al., 2017)", | |
| "ref_id": "BIBREF16" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "Very recently, DNNs have become the state-of-the-art method for toxic speech detection. Badjatiya et al. (2017) investigated the application of DNNs for hate speech detection and compared it with various classical features like character n-grams, Term Frequency-Inverse Document Frequency (TF-IDF) values, Bag of Word Vectors (BoWV), and Global Vectors for Word Representation (GloVe) (Pennington et al., 2014) . They found DNN methods to significantly outperform the existing shallow methods. Zhang et al. (2018) combined Convolutional neural network (CNN) and Recurrent neural network (RNN) by giving the output of CNN to RNN with Gated Recurrent Unit (GRU). Van Aken et al. (2018) proposed a combination of shallow models and DNN methods that outperforms all the individual models. Several evaluations of a range of NLP features was performed by Nobata et al. (2016) . Stammbach et al. (2018) reported different pre-processing techniques and their impact on the final classification. Wulczyn et al. (2017) went beyond the simple classification task and developed a method that combines crowdsourcing and machine learning to analyse personal attacks.", | |
| "cite_spans": [ | |
| { | |
| "start": 88, | |
| "end": 111, | |
| "text": "Badjatiya et al. (2017)", | |
| "ref_id": "BIBREF1" | |
| }, | |
| { | |
| "start": 385, | |
| "end": 410, | |
| "text": "(Pennington et al., 2014)", | |
| "ref_id": "BIBREF14" | |
| }, | |
| { | |
| "start": 494, | |
| "end": 513, | |
| "text": "Zhang et al. (2018)", | |
| "ref_id": "BIBREF24" | |
| }, | |
| { | |
| "start": 849, | |
| "end": 869, | |
| "text": "Nobata et al. (2016)", | |
| "ref_id": "BIBREF13" | |
| }, | |
| { | |
| "start": 872, | |
| "end": 895, | |
| "text": "Stammbach et al. (2018)", | |
| "ref_id": "BIBREF17" | |
| }, | |
| { | |
| "start": 987, | |
| "end": 1008, | |
| "text": "Wulczyn et al. (2017)", | |
| "ref_id": "BIBREF21" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "Currently, one of the most powerful semantic context representations are those obtained from BERT (Bidirectional Encoder Representations from Transformers) (Devlin et al., 2019; Young et al., 2018) . Compared to Mikolov's embedding (Mikolov et al., 2013) , BERT model takes into account large left and right semantic contexts of words and can generate different semantic representations for the same word based on its context. Furthermore, pretrained BERT model can be fine-tuned to a specific NLP task (Peters et al., 2019 ). The BERT model has resulted into new state-of-the-art for several NLP tasks.", | |
| "cite_spans": [ | |
| { | |
| "start": 156, | |
| "end": 177, | |
| "text": "(Devlin et al., 2019;", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 178, | |
| "end": 197, | |
| "text": "Young et al., 2018)", | |
| "ref_id": "BIBREF23" | |
| }, | |
| { | |
| "start": 232, | |
| "end": 254, | |
| "text": "(Mikolov et al., 2013)", | |
| "ref_id": "BIBREF12" | |
| }, | |
| { | |
| "start": 503, | |
| "end": 523, | |
| "text": "(Peters et al., 2019", | |
| "ref_id": "BIBREF15" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "In this article, we investigate several approaches based on different state-of-the-art DNN models and word representations for the task of automatic toxic comment detection. Among the classifiers, we used top performing DNNs in the field of NLP: CNN and RNN. CNN allows the extraction of local features in text, e.g. pertinent sequences of words. RNN is able to extract long-term dependencies that are definitely useful for toxic comment detection (Del Vigna et al., 2017) . To take into account the semantic context of the document, we propose to use different representations: Mikolov's, fastText and BERT embeddings. We compare these against transformers based BERT finetuning. The designed systems are evaluated on publicly available corpus of toxic comments from Wikipedia. The work of Bodapati et al. (2019) compares CNN based and fastText classifiers with various character and word based input representations to BERT fine-tuning. As compared to Bodapati et al. (2019) , we go beyond binary classification and propose a regression-based method. Furthermore, we analyse the robustness of these approaches with adversarial attacks by adding a toxic or healthy word to the comment. Additionally, we have compared CNN based architecture against RNN based Bi-LSTM and Bi-GRU classifiers. It should be noted that our results of binary classification are not directly comparable to Bodapati et al. (2019) due to differences in training and pre-processing setup.", | |
| "cite_spans": [ | |
| { | |
| "start": 448, | |
| "end": 472, | |
| "text": "(Del Vigna et al., 2017)", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 791, | |
| "end": 813, | |
| "text": "Bodapati et al. (2019)", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 954, | |
| "end": 976, | |
| "text": "Bodapati et al. (2019)", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 1383, | |
| "end": 1405, | |
| "text": "Bodapati et al. (2019)", | |
| "ref_id": "BIBREF2" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "The rest of the paper is organized as follows: Section 2 describes the approaches. The experiment protocol and the data are described in section 3. The classification results are discussed in section 4. Figure 1 presents a schema of our proposed methodology along with the different word representations. We first describe the different word representations and then discuss the DNN classifiers that we evaluate. In all our approaches, the DNN outputs represent the toxicity of a comment. ", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 203, | |
| "end": 211, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "2.1.1 Baseline approach: one-hot representation Our baseline is the classical one-hot input representation, wherein each input word is represented by a one-hot vector. Only the N most frequent words of training corpus are selected. The other words are represented as UNK. One-hot vectors are used as input to DNN classifier. The DNN will classify these sequences of one-hot vectors as toxic or nontoxic. The first hidden layer of the DNN computes the word embeddings. The weights of this embedding layer are trained together with the weights of the other layers of the network. The particularity here is that we do not exploit any pre-trained word embeddings and the entire training is performed using only the task specific corpus.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Comment representations", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "Feature-based approaches Embedding models entail vector-based word representations which are usually pre-trained on large datasets. In this work, pre-trained word representations are used as features in taskspecific DNN architectures. The DNN network classifies these sequences of word embeddings as toxic or non-toxic. We study and compare three state-of-the-art unsupervised word embedding models:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "2.1.2", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u2022 Mikolov's word embedding, which represent each word by taking into account a relatively small window of left and right context words (Mikolov et al., 2013) . \u2022 fastText subword embedding. It is an extension of Mikolov's embedding, which takes into account subword information and allows us to include rare and out-of-vocabulary words (Bojanowski et al., 2016; Mikolov et al., 2018) . \u2022 BERT WordPiece model (Devlin et al., 2019) . This model takes into account long left and right contexts of words. Thanks to this model, for each comment, embedding of each word-piece can be computed and used as input for DNN classifier. In the case of BERT model, the same word-piece can have different embeddings depending on the context. It is important to note that these representations are pretrained on corpora not specific to our task of toxic comment detection. Hence, will not be efficient to model the specificity of toxic speech (slang, affronts, abuse, etc.).", | |
| "cite_spans": [ | |
| { | |
| "start": 135, | |
| "end": 157, | |
| "text": "(Mikolov et al., 2013)", | |
| "ref_id": "BIBREF12" | |
| }, | |
| { | |
| "start": 336, | |
| "end": 361, | |
| "text": "(Bojanowski et al., 2016;", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 362, | |
| "end": 383, | |
| "text": "Mikolov et al., 2018)", | |
| "ref_id": "BIBREF11" | |
| }, | |
| { | |
| "start": 409, | |
| "end": 430, | |
| "text": "(Devlin et al., 2019)", | |
| "ref_id": "BIBREF7" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "2.1.2", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "BERT fine-tuning approach The principle of fine-tuning consists in starting from a pretrained model and updating the model parameters on the task specific corpus. As our task of hate speech detection is an NLP task where context plays a critical role, the architecture of BERT will be very appropriate. We take the same BERT pre-trained model as in Section 2.1.2 and we fine-tune this model using our training data. For fine-tuning, the hyper-parameters: batch size, learning rate, and the number of the training epochs are varied.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "2.1.3", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The task of toxic comment detection can be viewed from two perspectives: \u2022 A binary classification task: The neural network is directly trained to decide if a comment is toxic or nontoxic. \u2022 A regression task: For each comment we compute a score between 0 and 1 as a normalized average of labels from different annotators. The neural network is trained for predicting these scores (regression task). A threshold on the predicted score can be used to decide if the comment is toxic or not. The threshold is adjusted on the development set to maximize the F1-score. We investigate three state-of-the-art DNN architectures for our tasks:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "DNN classifiers", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "\u2022 CNN to identify local patterns in the comments;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "DNN classifiers", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "\u2022 bi-directional Long Short-Term Memory (bi-LSTM) to capture long range dependencies in the comments; \u2022 bi-directional Gated Recurrent Unit (bi-GRU), to capture long range dependencies with lesser model parameters;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "DNN classifiers", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "3. Experimental setup 3.1 Data description", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "DNN classifiers", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "We used the data collected in the framework of Wikipedia Detox project (Wulczyn et al., 2017) , including user's talks. In our work we exploited only the toxicity part of the corpus. This part contains 160k comments from English Wikipedia talk pages, each labelled by approximately 10 annotators via crowd-sourcing, on a spectrum of how toxic/healthy the comment is with regard to the conversation.", | |
| "cite_spans": [ | |
| { | |
| "start": 71, | |
| "end": 93, | |
| "text": "(Wulczyn et al., 2017)", | |
| "ref_id": "BIBREF21" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Wikipedia Detox corpus", | |
| "sec_num": "3.1.1" | |
| }, | |
| { | |
| "text": "The following toxicity rates are used by annotators: very toxic, toxic, neither, healthy, very healthy. According to this label definition, toxic speech corresponds to very toxic and toxic labels.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Wikipedia Detox corpus", | |
| "sec_num": "3.1.1" | |
| }, | |
| { | |
| "text": "For many comments in the Wikipedia Detox corpus, there is a disagreement between annotators. Sometimes, it is difficult to define a dominant label for a comment. To perform the binary classification (toxic or not toxic), for each comment, we decided to use the following majority vote labelling: Some examples of the toxic comments are: \"You are a big fat idiot, stop spamming my userspace\", \"What the fuck is your problem?\", \"God damn it fuckers, i am using the god damn sand box\".", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Wikipedia Detox corpus", | |
| "sec_num": "3.1.1" | |
| }, | |
| { | |
| "text": "Train, development and test corpus We used the train/development/test partition provided with the Detox corpus (respectively 96k, 32k, 32k). Training data is used to train our classifiers and to fine-tune the BERT model. Development corpus is used to tune the hyperparameters. Test corpus is used to evaluate the performance of the system. We compared the classifier predictions in terms of F1-score.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "3.1.2", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "For many NLP tasks, training data pre-processing has an important impact on the performance of the system. Moreover, DNN approaches are data-driven. These two factors give a very high importance to the pre-processing. We decided to set the maximum length of a comment to 200 words for reducing the computation time and for avoiding the out-of-memory problems for BERT (because it is a very large model). For this, we keep the first 200 words of each comment of the training, development and test sets. We removed the toxic comments with more than 200 words per comment from the training set because it is possible that the toxic part of the comment is located after the 200 th word. We performed this removal only for training. This preprocessing removed about 5% of toxic comments from the training set. Table 1 shows that toxic comments represent only about 17% of all comments. So, our corpus has an unbalanced class distribution.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 805, | |
| "end": 812, | |
| "text": "Table 1", | |
| "ref_id": "TABREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Data pre-processing", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "We converted all words to lowercase and used uncased BERT, fastText and Mikolov's pre-trained models. We removed the punctuations for the Mikolov's, fastText and one-hot approach. We kept the punctuation for the BERT model.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data pre-processing", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "As Detox corpus is limited in size, we used pre-trained models: Mikolov's word embedding: provided by Google 1 and pretrained on a wide corpus of 100G words from Google news corpus. Embedding dimension is 300 for 3M words. fastText subword embedding: provided by Facebook 2 and pre-trained on Wikipedia 2017, UMBC webbase and 1 https://github.com/mmihaltz/word2vecGoogleNews-vectors statmt.org news datasets with total 16B tokens. Embedding dimension is 300, the vocabulary is 1M words. BERT-base WordPiece model: English (uncased) model provided by Google, pre-trained on BookCorpus and Wikipedia, with 12 transformer layers and 12 self-attention heads. The embedding size is 768, the number of WordPieces is 30k (including the punctuations). The total number of parameters is 110 million. WordPiece BERT model and fastText models succeed to represent all words in our corpus. Mikolov's embedding is a word based model. Some words from our corpus are not included in its vocabulary i.e, Out-Of-Vocabulary (OOV) words. Our training set has 86.5k occurrences of OOVs (2%), development set has 45.8k OOV occurrences (2.4%) and the test set has 45.3k (2.4%). To obtain an embedding for these OOV, we compute an average of the embeddings of all the words in the vocabulary.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Embedding models", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "The evaluated configurations are presented in the following: for one-hot approach we keep the 75K or 100K most frequent words. For CNN based model we explored one or two convolutional layers (filter size between 3 and 5), followed by two dense layers (with 64-256, 16-64 dense units), with or without dropout. For bi-LSTM and bi-GRU, we explored one or two layers (with 50, 128 units), followed by one or two dense layers (with 64-256, 16-64 dense units), with or without dropout. We use L2 regularization and adam optimizer. For fine-tuning BERT we used maximum sequence length of 256, batch size of 32, learning rate of 2\u202210 -5 and 2 epochs. Table 2 , part A, shows the results for baseline methods for one-hot approach: using words or using the same wordpieces as in BERT. Part B focuses on pre-trained embeddings for feature-based approaches. Moreover, we concatenate Mikolov's and BERT embeddings together and use it as input features to DNN (indicated as 'Mikolov's+BERT word embedding' in Table 2 ). In this model, words split into word-pieces by BERT tokenizer are averaged and concatenated with corresponding Mikolov's word embedding. The embeddings obtained by averaging the word-piece tokens are indicated as 'BERT word embedding'. For 'Mikolov's+BERT fine-tun. word emb.' configuration we concatenate Mikolov's and BERT finetuned embeddings. The results of Part C are obtained by BERT fine-tuning. For the two parts (A, B), we have experimented with three different classifiers: CNN, bi-LSTM and bi-GRU.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 644, | |
| "end": 651, | |
| "text": "Table 2", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 996, | |
| "end": 1003, | |
| "text": "Table 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "DNN model configurations", | |
| "sec_num": "3.4" | |
| }, | |
| { | |
| "text": "As shown in the table, our proposed methods in part B and C show better performance than the baseline methods in part A. Among the classifiers, bi-LSTM and bi-GRU performs slightly better than the CNN. Mikolov's embedding of part B performs worse than one-hot approach. This can be due to the presence of OOV words: the one-hot approach models N most frequent words of training corpus, while Mikolov's embeddings is trained on non-toxic corpus and it is possible that some important toxic words (slang) of our corpus are missing in the Mikolov's pre-trained model. BERT with words (BERT word embedding) slightly underperforms compared to BERT word-piece embeddings. This can be due to some loss of information while averaging the embeddings. BERT embedding performed better than one-hot approach. Joint embedding (Mikolov's+BERT) give slightly better performance than BERT embedding alone. The best method is BERT fine-tuning which achieves 78.2% F1-score. Joint embedding Mikolov's+BERT fine-tuned word embedding achieves the performance close to BERT fine-tuning. Table 2 exhibits that BERT is effective for both the fine-tuning and feature-based approaches. It is worth noting that the evaluated models have a different numbers of learned parameters: DNN based classifier models have about 1M parameters, whereas BERT fine-tuned model has 110M parameters. BERT embedding is a good trade-off between performance and number of model parameters. Table 2 : Binary classification F1-score for different classifiers and different input representations. A preliminary error analysis shows that sometimes nontoxic speech can be misclassified as toxic speech in the presence of words like bullies, anti-semitism. For example, the comment \"You're a nice guy Irishpunktom. It takes guts to speak against bullies.\" is misclassified as toxic. Likewise, toxic speech is misclassified as non-toxic speech due to sarcasm, irony, rhetoric question, etc. For example, \"Thats fine. Thank your extreme rudeness. That front page looks so unwelcoming.\" is misclassified as non-toxic.", | |
| "cite_spans": [ | |
| { | |
| "start": 813, | |
| "end": 829, | |
| "text": "(Mikolov's+BERT)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 1066, | |
| "end": 1073, | |
| "text": "Table 2", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 1446, | |
| "end": 1453, | |
| "text": "Table 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Binary classification", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "These experiments compare the performances based on the regression model. A threshold is applied to the regression score to decide if the comment is toxic or not. We use bi-LSTM classifier as it gives the best performance according to Table 2 .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 235, | |
| "end": 242, | |
| "text": "Table 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Classification using regression model", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "We observe that BERT model is more powerful than other models. As for binary classification, BERT fine-tuning gives the best results. Mikolov's+BERT word embedding shows the results close to BERT fine-tuning. We obtained the following results in terms of RMSE (Root Mean Square Error) and MAE (Mean Absolute Error): Word-based one-hot 0.065 and 0.050; Word-piece based one-hot 0.065 and 0.050; Mikolov's 0.066 and 0.049; fastText 0.062 and 0.047; BERT 0.062 and 0.047; Mikolov's+BERT word emb. 0.06 and 0.047; BERT fine-tuning 0.06 and 0.047.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Classification using regression model", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "These measures further confirm our conclusions.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Classification using regression model", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "In order to evaluate the robustness of our classification systems, we added a toxic word ('fuck') to each comment of the test set and a healthy word ('love') to each comment of the test set. Table 4 shows the percentage of correctly classified comments that change from predicted non-toxic to toxic comments when a toxic word is appended, and from toxic to non-toxic when a healthy word is appended. In these experiments, we use bi-LSTM (the best DNN according to Table 2 ) and the threshold of 0.6 with the regression model. We perform the tests only on feature-based models. We observe that all models are susceptible to the word appending attacks, as also observed in (Gr\u00f6ndahl et al., 2018) . Classifiers using Mikolov's and fastText embeddings are more sensitive to appending of a single word. Classifier using BERT embedding is more robust.", | |
| "cite_spans": [ | |
| { | |
| "start": 671, | |
| "end": 694, | |
| "text": "(Gr\u00f6ndahl et al., 2018)", | |
| "ref_id": "BIBREF8" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 191, | |
| "end": 198, | |
| "text": "Table 4", | |
| "ref_id": "TABREF6" | |
| }, | |
| { | |
| "start": 464, | |
| "end": 471, | |
| "text": "Table 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Robustness evaluation", | |
| "sec_num": "4.3" | |
| }, | |
| { | |
| "text": "In this article, we have investigated several approaches for toxic comment classification using DNNs. We explored feature-based unsupervised comment representations using Mikolov's, fastText and BERT pre-trained models. These representations are used as input for DNN networks. These approaches are compared to the BERT fine-tuning. We designed binary classification and regression-based approaches. On Wikipedia Detox corpus, our analysis has shown that BERT fine-tuning is the most efficient at this task. Moreover, BERT embedding is the most robust to word attacks. Among DNN based classifiers, bi-LSTM performs better than CNN and bi-GRU at classifying toxic speech.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "In the future, we would like to study the impact of data bias on toxic speech detection (Wiegand et al., 2019) and to perform depth study of the multi-class classification (Vaswani et al., 2017) . A detailed error analysis to evaluate the linguistic phenomena will also be performed. Moreover, models like XLNet pre-trained model (Yang et al., 2019) or ULMFiT pre-trained language model (Howard and Ruder, 2018) can be studied.", | |
| "cite_spans": [ | |
| { | |
| "start": 88, | |
| "end": 110, | |
| "text": "(Wiegand et al., 2019)", | |
| "ref_id": "BIBREF20" | |
| }, | |
| { | |
| "start": 172, | |
| "end": 194, | |
| "text": "(Vaswani et al., 2017)", | |
| "ref_id": "BIBREF18" | |
| }, | |
| { | |
| "start": 330, | |
| "end": 349, | |
| "text": "(Yang et al., 2019)", | |
| "ref_id": "BIBREF22" | |
| }, | |
| { | |
| "start": 387, | |
| "end": 411, | |
| "text": "(Howard and Ruder, 2018)", | |
| "ref_id": "BIBREF9" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "https://fasttext.cc/docs/en/english-vectors.html", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "This work was funded by the M-PHASIS project supported by the French National Research Agency (ANR) and German National Research Agency (DFG) under contract ANR-18-FRAL-0005.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Acknowledgements", | |
| "sec_num": "6." | |
| } | |
| ], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Challenges for Toxic Comment Classification: An In-Depth Error Analysis", | |
| "authors": [ | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Aken Van", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Risch", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Krestel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "L\u00f6ser", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the 2nd Workshop on Abusive Language Online (ALW2)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "33--42", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Aken van, B., Risch, J., Krestel, R., and L\u00f6ser, A. (2018). Challenges for Toxic Comment Classification: An In- Depth Error Analysis. In Proceedings of the 2nd Workshop on Abusive Language Online (ALW2), pp. 33- 42.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Deep Learning for Hate Speech Detection in Tweets", | |
| "authors": [ | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Badjatiya", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Gupta", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Gupta", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Varma", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Proceedings of the 26th International Conference on World Wide Web Companion", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "759--760", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Badjatiya, P., Gupta, S., Gupta, M., and Varma, V. (2017). Deep Learning for Hate Speech Detection in Tweets. In Proceedings of the 26th International Conference on World Wide Web Companion, pp. 759-760.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Neural Word Decomposition Models for Abusive Language Detection", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Bodapati", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Gella", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Bhattacharjee", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Al-Onaizan", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the Third Workshop on Abusive Language Online", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "135--145", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Bodapati, S., Gella, S., Bhattacharjee, K., and Al-Onaizan, Y. (2019). Neural Word Decomposition Models for Abusive Language Detection. In Proceedings of the Third Workshop on Abusive Language Online, pp. 135-145.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Enriching Word Vectors with Subword Information", | |
| "authors": [ | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Bojanowski", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "\u00c9", | |
| "middle": [], | |
| "last": "Grave", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Joulin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Mikolov", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Transactions of the Association for Computational Linguistics", | |
| "volume": "5", | |
| "issue": "", | |
| "pages": "135--146", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Bojanowski, P., Grave, \u00c9., Joulin, A., and Mikolov, T. (2017). Enriching Word Vectors with Subword Information. Transactions of the Association for Computational Linguistics, 5, 135-146.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Automated Hate Speech Detection and the Problem of Offensive Language", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Davidson", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Warmsley", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Macy", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "I", | |
| "middle": [], | |
| "last": "Weber", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Proceedings of the Eleventh International AAAI Conference on Web and Social Media", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Davidson, T., Warmsley, D., Macy, M., and Weber, I. (2017). Automated Hate Speech Detection and the Problem of Offensive Language. In Proceedings of the Eleventh International AAAI Conference on Web and Social Media.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Hate me, Hate me not: Hate Speech Detection on Facebook", | |
| "authors": [ | |
| { | |
| "first": "Del", | |
| "middle": [], | |
| "last": "Vigna", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Cimino", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Dell'orletta", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Petrocchi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Tesconi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Proceedings of the First Italian Conference on Cybersecurity (ITASEC17)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "86--95", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Del Vigna, F., Cimino, A., Dell'Orletta, F., Petrocchi, M., and Tesconi, M. (2017). Hate me, Hate me not: Hate Speech Detection on Facebook. In Proceedings of the First Italian Conference on Cybersecurity (ITASEC17), pp. 86-95.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Hate Speech in Cyberspace", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Delgado", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Stefancic", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Wake Forest L. Rev", | |
| "volume": "49", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Delgado, R., and Stefancic, J. (2014). Hate Speech in Cyberspace. Wake Forest L. Rev., 49, 319.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Devlin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [ | |
| "W" | |
| ], | |
| "last": "Chang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Lee", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Toutanova", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", | |
| "volume": "1", | |
| "issue": "", | |
| "pages": "4171--4186", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Devlin, J., Chang, M. W., Lee, K., and Toutanova, K. (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1, pp. 4171-4186.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "All You Need is\" Love\" Evading Hate Speech Detection", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Gr\u00f6ndahl", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Pajola", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Juuti", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Conti", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Asokan", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the 11th ACM Workshop on Artificial Intelligence and Security", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "2--12", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Gr\u00f6ndahl, T., Pajola, L., Juuti, M., Conti, M., and Asokan, N. (2018). All You Need is\" Love\" Evading Hate Speech Detection. In Proceedings of the 11th ACM Workshop on Artificial Intelligence and Security, pp. 2-12.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Universal Language Model Fine-tuning for Text Classification", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Howard", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Ruder", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics", | |
| "volume": "1", | |
| "issue": "", | |
| "pages": "328--339", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Howard, J., and Ruder, S. (2018). Universal Language Model Fine-tuning for Text Classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, Volume 1: Long Papers, pp. 328-339.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Multi-Task Deep Neural Networks for Natural Language Understanding", | |
| "authors": [ | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "He", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Gao", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "4487--4496", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Liu, X., He, P., Chen, W., and Gao, J. (2019). Multi-Task Deep Neural Networks for Natural Language Understanding. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 4487-4496.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Advances in Pre-Training Distributed Word Representations", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Mikolov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "\u00c9", | |
| "middle": [], | |
| "last": "Grave", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Bojanowski", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Puhrsch", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Joulin", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the Eleventh International Conference on Language Resources and Evaluation", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mikolov, T., Grave, \u00c9., Bojanowski, P., Puhrsch, C., and Joulin, A. (2018). Advances in Pre-Training Distributed Word Representations. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018).", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Distributed Representations of Words and Phrases and their Compositionality", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Mikolov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "I", | |
| "middle": [], | |
| "last": "Sutskever", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Corrado", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Dean", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceedings of the Advances in Neural Information Processing Systems", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "3111--3119", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., and Dean, J. (2013). Distributed Representations of Words and Phrases and their Compositionality. In Proceedings of the Advances in Neural Information Processing Systems, pp. 3111-3119.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Abusive Language Detection in Online User Content", | |
| "authors": [ | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Nobata", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Tetreault", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Thomas", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Mehdad", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Chang", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "Proceedings of the 25th International Conference on World Wide Web", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "145--153", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nobata, C., Tetreault, J., Thomas, A., Mehdad, Y., and Chang, Y. (2016). Abusive Language Detection in Online User Content. In Proceedings of the 25th International Conference on World Wide Web, pp. 145-153.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "Glove: Global Vectors for Word Representation", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Pennington", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Socher", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [ | |
| "D" | |
| ], | |
| "last": "Manning", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1532--1543", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Pennington, J., Socher, R., and Manning, C. D. (2014). Glove: Global Vectors for Word Representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1532-1543.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "To Tune or Not to Tune? Adapting Pretrained Representations to Diverse Tasks", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Peters", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Ruder", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [ | |
| "A" | |
| ], | |
| "last": "Smith", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the 4th Workshop on Representation Learning for NLP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "7--14", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Peters, M. E., Ruder, S., and Smith, N. A. (2019). To Tune or Not to Tune? Adapting Pretrained Representations to Diverse Tasks. In Proceedings of the 4th Workshop on Representation Learning for NLP (RepL4NLP-2019), pp. 7-14.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "A Web of Hate: Tackling Hateful Speech in Online Social Spaces", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [ | |
| "M" | |
| ], | |
| "last": "Saleem", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [ | |
| "P" | |
| ], | |
| "last": "Dillon", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Benesch", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Ruths", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:1709.10159" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Saleem, H. M., Dillon, K. P., Benesch, S., and Ruths, D. (2017). A Web of Hate: Tackling Hateful Speech in Online Social Spaces. arXiv preprint arXiv:1709.10159.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Offensive Language Detection with Neural Networks for Germeval Task", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Stammbach", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Zahraei", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Stadnikova", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Klakow", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the 14th Conference on Natural Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Stammbach, D., Zahraei, A., Stadnikova, P., and Klakow, D. (2018). Offensive Language Detection with Neural Networks for Germeval Task 2018. In Proceedings of the 14th Conference on Natural Language Processing, 2018, p. 58.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "Attention is all you need", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Vaswani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Shazeer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Parmar", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Uszkoreit", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Jones", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [ | |
| "N" | |
| ], | |
| "last": "Gomez", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "I", | |
| "middle": [], | |
| "last": "Polosukhin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Proceedings of the Advances in Neural Information Proc. Systems", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., and Polosukhin, I. (2017). Attention is all you need. In Proceedings of the Advances in Neural Information Proc. Systems.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Understanding Abuse: A Typology of Abusive Language Detection Subtasks", | |
| "authors": [ | |
| { | |
| "first": "Z", | |
| "middle": [], | |
| "last": "Waseem", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Davidson", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Warmsley", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "I", | |
| "middle": [], | |
| "last": "Weber", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Proceedings of the First Workshop on Abusive Language Online", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "78--84", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Waseem, Z., Davidson, T., Warmsley, D., & Weber, I. (2017). Understanding Abuse: A Typology of Abusive Language Detection Subtasks. In Proceedings of the First Workshop on Abusive Language Online, pp. 78-84.", | |
| "links": null | |
| }, | |
| "BIBREF20": { | |
| "ref_id": "b20", | |
| "title": "Detection of Abusive Language: the Problem of Biased Datasets", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Wiegand", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Ruppenhofer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Kleinbauer", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", | |
| "volume": "1", | |
| "issue": "", | |
| "pages": "602--608", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Wiegand, M., Ruppenhofer, J., and Kleinbauer, T. (2019). Detection of Abusive Language: the Problem of Biased Datasets. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1, pp. 602-608.", | |
| "links": null | |
| }, | |
| "BIBREF21": { | |
| "ref_id": "b21", | |
| "title": "Ex machina: Personal attacks seen at scale", | |
| "authors": [ | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Wulczyn", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Thain", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Dixon", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Proceedings of the 26th International Conference on World Wide Web", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1391--1399", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Wulczyn, E., Thain, N., and Dixon, L. (2017). Ex machina: Personal attacks seen at scale. In Proceedings of the 26th International Conference on World Wide Web, pp. 1391- 1399.", | |
| "links": null | |
| }, | |
| "BIBREF22": { | |
| "ref_id": "b22", | |
| "title": "XLNET: Generalized Autoregressive Pretraining for Language Understanding", | |
| "authors": [ | |
| { | |
| "first": "Z", | |
| "middle": [], | |
| "last": "Yang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Z", | |
| "middle": [], | |
| "last": "Dai", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Yang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Carbonell", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Salakhutdinov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Le", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Q", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the Advances in Neural Information Processing Systems", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "5754--5764", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov, R., and Le, Q. (2019). XLNET: Generalized Autoregressive Pretraining for Language Understanding. In Proceedings of the Advances in Neural Information Processing Systems, pp. 5754-5764.", | |
| "links": null | |
| }, | |
| "BIBREF23": { | |
| "ref_id": "b23", | |
| "title": "Recent Trends in Deep Learning based Natural Language Processing", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Young", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Hazarika", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Poria", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Cambria", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "IEEE Computational Intelligence Magazine", | |
| "volume": "13", | |
| "issue": "3", | |
| "pages": "55--75", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Young, T., Hazarika, D., Poria, S., and Cambria, E. (2018). Recent Trends in Deep Learning based Natural Language Processing. IEEE Computational Intelligence Magazine, 13(3), pp. 55-75.", | |
| "links": null | |
| }, | |
| "BIBREF24": { | |
| "ref_id": "b24", | |
| "title": "Detecting Hate Speech on Twitter using a Convolution-GRU based Deep Neural Network", | |
| "authors": [ | |
| { | |
| "first": "Z", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Robinson", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Tepper", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the European Semantic Web Conf", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "745--760", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Zhang, Z., Robinson, D., and Tepper, J. (2018, June). Detecting Hate Speech on Twitter using a Convolution- GRU based Deep Neural Network. In Proceedings of the European Semantic Web Conf., pp.745-760. Springer, Cham.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "type_str": "figure", | |
| "text": "Proposed system architecture for toxic comment detection.", | |
| "uris": null, | |
| "num": null | |
| }, | |
| "TABREF2": { | |
| "html": null, | |
| "num": null, | |
| "type_str": "table", | |
| "content": "<table/>", | |
| "text": "Statistics on Wikipedia Detox data after preprocessing. 'K' denotes thousand, 'M' denotes million." | |
| }, | |
| "TABREF5": { | |
| "html": null, | |
| "num": null, | |
| "type_str": "table", | |
| "content": "<table><tr><td colspan=\"3\">Binary classification</td><td colspan=\"3\">Regression model</td></tr><tr><td colspan=\"2\">Mikolov fast</td><td colspan=\"3\">BERT Mikolov fast</td><td>BERT</td></tr><tr><td/><td>Text</td><td/><td/><td>Text</td></tr><tr><td>non-toxic to toxic 88.0</td><td>78.1</td><td>37.5</td><td>71.9</td><td colspan=\"2\">78.0 34.1</td></tr><tr><td>toxic to non-toxic 6.5</td><td>4.8</td><td>4.1</td><td>10.9</td><td colspan=\"2\">10.0 7.6</td></tr></table>", | |
| "text": "F1-score for Bi-LSTM classifier and different input representations using a threshold on regression model." | |
| }, | |
| "TABREF6": { | |
| "html": null, | |
| "num": null, | |
| "type_str": "table", | |
| "content": "<table/>", | |
| "text": "Percentage of correctly classified comments, a new word is appended. Bi-LSTM and different models." | |
| } | |
| } | |
| } | |
| } |