Why does python get so much hatred

The detection of inappropriate content on the Internet: AI process, evaluation and challenges

Very often aggressive and hateful posts can be found on social media. Their automatic detection is desirable and is currently being intensively researched. The quality of the corresponding algorithms must be evaluated on an ongoing basis. The article reports on evaluation results and the challenges within the framework of the HASOC Initiative 2019. Before that, the neural network architectures such as BERT, which currently deliver the best results, will be introduced.

In social media, aggressive and hateful posts are rather frequent. It is desirable to identify such content automatically and much research is dedicated to this area. The quality of recognition algorithms needs to be constantly monitored. This article reports evaluation results and the challenges within the HASOC initiative 2019. In the first instance, neural network architectures are introduced, as they currently deliver the best result for hate speech recognition. Within the HASOC dataset, the BERT system classified posts with an accuracy of around 80 per cent.

1 Hate Speech on Social Media

Social media and online communication have established themselves as a significant part of the reality of life for many citizens. This free and media-conveyed form of expression of opinion seems to seduce many people into making radical contributions. Hate speech and hate speech have been online for a long time and are already perceived by many as a threat to social peace (Hajok & Selg 2018). False reports, automatically generated amounts, political microtargeting and aggressive threats against individuals also disrupt a rational, public discourse and thus endanger the formation of democratic and pluralistic opinions. Several studies suggest that opinions on the extreme edges of the spectrum are more common. This emerged from the example of the discussion on the independence referendum in Catalonia (Ghanem et al. 2019). It is controversial to what extent this polarization is a consequence of the digitization of communication habits.

The controversial verdict in the case of the politician Künast in 2019 shows the difficulties in delimitation, the legal assessment when weighing against freedom of expression and the formation of public opinion on the subject. For this reason, politics has long been reflecting on possible countermeasures that, for example, result in the EU's measures against disinformation (Kettemann 2019).

In view of the continuously produced amounts of texts and comments, the use of automatic processes, which are often generally considered to be AI methods, has proven to be inevitable. The judgment of the European Court of Justice passed in October 2019 even considers the use of automated procedures to be necessary. The Court of Justice demands that expressions of hatred that have the same meaning as posts that have already been recognized as problematic are also automatically identified. Especially when evaluating human texts by computers, the ethical dimension and, above all, the fine line between freedom of expression and censorship come into focus. Society will only accept AI methods if there is confidence in their results (Kuhlen 1999). In his four theses on regulating the internet, Marc Zuckerberg called for a clear definition of hate speech (Lewanczik 2019). However, this poses a significant difficulty because ideas of what is harmful content vary widely between societies and individuals.

The automatic detection of hatred or hate speech is increasingly establishing itself as an active research topic. The area of ​​machine learning has made significant advances in recent years that make detection easier. The successful algorithms do not search for individual words or with comprehensible rules for linguistic patterns, but build classification procedures from many examples. These training data only contain examples of inappropriate and acceptable content. The amount of training in texts is what drives development. Data that deviate significantly from this (e.g. in other languages) cannot be recognized after the training. The recognition quality of the algorithms can be checked and optimized on the basis of these data sets. For this purpose, some amounts of data have been compiled in recent years. This article briefly introduces two of these scientific benchmarks and picks up on some relevant results. First, there is a brief introduction to the algorithms of deep learning, some of which were only established in recent years.

2 AI methods for text classification

The representation of texts for knowledge processing is traditionally done by linking them with words. Whether there are a few keywords in an intellectual indexing or in full-text searches, essentially all of the words or word forms present in the text. A single text document is transformed into a long vector of numbers. Each position in these stands for one word. In addition, some weighting methods such as the "inverse document frequency" and variants of length normalization have become established for ranking methods. These lead to a system not only having binary knowledge about the occurrence or non-occurrence of a word (Henrich 2014).

Classification or clustering methods then process these vectors in machine learning. Today these are called representations one hot encoding-Variants. This signals that from the entire extensive vocabulary of a text collection for a single document, only the existing words (i.e. the content-bearing features) are marked. All other positions of the vector are filled with zeros and thus indicate that the corresponding word is not contained in the document. This leads z. For example, two words always appear dissimilar to one another, even if they are closely related. Distributed semantic representations, which only require much shorter vectors, have been used for a long time. Unlike one hot encodings however, the dimensions cannot be assigned to a word or to any meaning at all. Systems move from symbolic representations to the sub-symbolic level at which internal structures can no longer be clearly interpreted (Smolensky 1988). So documents are represented by a series of numbers that cannot be interpreted. The values ​​are initially initialized randomly and then modified in the course of a learning process until similar documents also have similar vectors.

3 word embeddings

First, the development of distributed semantic representations for words is shown. It is based on artificial neural networks, which are almost synonymous with artificial intelligence. Neural networks are adaptive systems that process information in a tolerant and robust manner and are based on very simple basic principles. Like their natural model - the nervous systems of living beings - they consist of numerous simple, interconnected processors. Via their connections, these processors or neurons exchange numerical information in the form of activation (Kruse et al. 2015). The links have values ​​that determine their permeability. The neurons send out activation, but only as much activation arrives at the next neuron as the link allows. The connections become interesting for learning because their value changes.

Fig. 1:

Schematic representation of an artificial neuron.

Values ​​initially set randomly for the connections are changed depending on the desired behavior until the mapping of input patterns in output classes for the training data is successful. Then the learning process is considered complete and a neural network has learned a function from input to output. The “backpropagation algorithm” is used in particular (Kruse et al. 2015). The activation propagation runs in it from an input layer to an output layer with the desired result. In between there are hidden layers with neurons, which each transmit the activation and are highly networked with the other layers. The learning error in the output is determined and then, depending on this, the weights in the entire network are adjusted so that the result for this learning example is achieved somewhat better. The process is repeated very often.

Neural networks were proposed for language processing systems and information retrieval at an early stage (Mandl 2000), but it has only been around 2013 that they have benefited from the computing capacity that is now available. The so-called Word embeddings. A classification algorithm tries to learn the sequential sequence of words. For this purpose, several training examples are generated from each sentence. For example, a neural network could attempt to predict the middle of three words. For Word Embeddings, a word means a sequence of mostly 300 numbers. The mapping onto a vector of 300 numbers is to be learned from two words as input with two times 300 dimensions. After randomly assigning values ​​for all words, numerous training examples are obtained from very large amounts of text. With them, the vector of each word is slowly changed until the error is as small as possible. If this succeeds, the system can determine which word is most likely between the other two. If different words often appear in the context of the two input words, these vectors will be very similar.

While the similarity of words in the vector space model described above is only recorded by their common occurrence in documents, Word Embeddings model the similarity from the sequence of texts. Words that usually appear in similar neighborhoods converge on similar vectors. Several Word Embeddings that have already been trained are available for download and can easily be used with just a few lines in Python (e.g. GLOVE).

Tab. 1:

Eight dimensions of the word embeddings for some words (from the model at https://github.com/devmount/GermanWordEmbeddings).

stupid

0,16

0,111

0,084

0,043

-0,095

0,147

-0,098

-0,016

Smart

0,329

-0,043

-0,022

0,162

-0,171

0,028

0,11

-0,232

Well

0,172

-0,427

0,104

0,162

-0,21

0,098

-0,027

-0,195

bad

0,219

-0,18

0,121

0,161

-0,015

-0,053

-0,132

-0,346

old

0,044

-0,135

-0,089

-0,079

-0,275

0,272

0,154

-0,374

New

0,181

-0,142

0,462

0,028

0,138

-0,07

-0,057

-0,349

Fire

0,211

0,115

0,23

0,032

0,115

0,18

-0,19

0,298

Tab. 2:

Similarity comparison for some words based on Word Embeddings (for the model at https://github.com/devmount/GermanWordEmbeddings).

stupid

Smart

Well

bad

old

New

Fire

young

stupid

1

0,612

0,456

0,609

0,458

0,262

0,192

0,533

Smart

0,612

1

0,47

0,443

0,414

0,388

0,177

0,522

Well

0,456

0,47

1

0,781

0,427

0,41

0,217

0,429

bad

0,609

0,443

0,781

1

0,385

0,386

0,174

0,399

old

0,458

0,414

0,427

0,385

1

0,382

0,205

0,723

New

0,262

0,388

0,41

0,386

0,382

1

0,246

0,352

Fire

0,192

0,177

0,217

0,174

0,205

0,246

1

0,167

young

0,533

0,522

0,429

0,399

0,723

0,352

0,167

1

Do not necessarily indicate opposing terms such as old and New very little resemblance. This is plausible, since they often occur in similar contexts and the word embeddings are trained from the word sequences.

Word Embeddings achieve very good results for many applications. They belong to the paradigm of deep learning, in which the descriptive properties (the dimensions of the vectors) are learned by the system itself and not developed by an expert, for example. However, one step is still missing for the classification or retrieval of documents. The representations of several words must be combined into one document. As a simple approach, the values ​​of all word vectors at each position were simply averaged.

4 neural language models for sentences

Among other things, recursive language models provide a better variant. The basic idea of ​​recurrent neural networks has been known for a long time. With modern hardware and large amounts of data, however, they have only been able to develop their potential for a few years. The neural network learns the mapping from the current word to the next. In addition, it receives the previous context of the sentence as input, in that the hidden layer enters the network as the second input. The target output is always the following word or its word embedding vector. Every word becomes a training example. A representation of the semantics of the entire sentence is built up in the hidden layer (Kruse et al. 2015).

In essence, it becomes a so-called Sentence encoding reached. So instead of just converting one word into a vector, the networks transform sentences into a vector that usually encompass significantly more dimensions than those of the word embeddings. Today the basic idea of ​​recursion is further developed in more complex architectures. These include, for example, the Long-Short Term Memory Systems (LSTM). Inside, however, there is always the simple transmission of activation between neurons. In addition, the sequences of letters and sentences are modeled in multiple layered systems.

The new BERT procedure uses so-called Transformerthat process sequential data more flexibly. BERT stands for "Bidirectional Encoder Representations from Transformers". Since its introduction in 2018, it has been considered a magic bullet in language processing that delivers very good results. BERT was developed by Google and has already achieved good results for retrieval evaluations (Craswell et al. 2019). The search engines Bing and Google are probably already using BERT productively. It plays an important role in the results for hate speech detection.

5 evaluation benchmarks for hate speech

The detection of hate speech and other inappropriate content is a major challenge. The topics associated with hateful content can be very heterogeneous (De Smedt et al. 2018). This means that systems can not only classify the texts thematically. Marc Zuckerberg said in an interview: "Determining if something is hate speech is very linguistically nuanced"[1]. Clear rules therefore do not provide a solution in this application.

The training resources with which the neural networks presented above or other classification systems are trained are of particular importance for the development of systems. Increasingly, resources are developed and then in so-called Shared tasks made available to the community. This allows the results of various algorithms to be compared fairly and the results to be discussed in the specialist public. For this purpose, real tweets or posts are systematically collected and categorized into two or more classes. A data set was developed for German as part of the GermEval initiative (Struß et al. 2019). It was based on Twitter data, which the organizers had annotated. GermEval defines the ABUSE, INSULT and PROFANITY classes and the primary task of the systems is to divide all tweets into these. In total, the set includes over 7,000 tweets. According to the organizers, the extreme right-wing spectrum clearly predominates (90% of the content falls into this category). The agreement of four annotators at GermEval was measured for a sample of 300 tweets and achieved a kappa value of 0.59, which is considered moderate (Struß et al. 2019). This shows how difficult it is to find common standards.

In addition to classifying them as hate speech, the problematic tweets should be sorted into the classes implicitly and explicitly. The classification accuracy of the systems is measured using the F1 measure, which combines recall and precision. The best system achieved an F1 measure of 0.76 for the binary task (Struß et al. 2019). A total of 12 teams submitted results for GermEval in 2019. The best six systems are shown in Table 3.

Tab. 3:

Best systems at GermEval 2019 in the binary task.

position

team

run

F1

approach

1

Politehnica Univ. of Bucharest

1

0,7695

Self trained BERT

2

Politehnica Univ. of Bucharest

3

0,7695

Self trained BERT

3

Politehnica Univ. of Bucharest

2

0,7686

Self trained BERT

4

TU Vienna

1

0,7680

N-gram, ensemble

5

TU Vienna

2

0,7675

N-gram, ensemble

6

Mittweida University of Applied Sciences

2

0,7663

fastText Embeddings, SVM

It is noticeable how close the best systems are. Their difference is less than one percent. Nevertheless, they use very heterogeneous systems. The BERT, which was only introduced in 2018, delivers the best performance, but far less complex systems come very close to the front runner.

The second benchmark with texts for German started in 2019. HASOC is a multilingual resource for hate speech recognition and offers data for German, Hindi and English. HASOC stands for Hate Speech and Offensive Content Identification in Indo-European Languages. The track was established as part of the FIRE conference (Mandl et al. 2019). HASOC 2019 received a great response from the research community and 37 teams gave results to the organizers. 28 experiments were submitted for German.

Fig. 2:

WordCloud from the HATE and OFFENSIVE tweets for HASOC English (generated with Voyant Tools).

HASOC initially models the task as a binary classification. The posts identified as problematic should then be divided more precisely into the following classes in the second step: HATE, OFFENSIVE and PROFANE. The organizers have annotated 4,600 Twitter tweets and Facebook posts for the German. As an example, some results for the binary task for English and German are presented here. The top teams for German mostly used the BERT system and are very close together (Mandl et al. 2019).

Tab. 4:

Best systems at HASOC 2019 for German (Sub Task A).

position

team

run

Marco F1

approach

1

IIT Kharagpur

1

0,616

BERT, boosting

2

Univ. of the Saarland

1

0,606

BERT

3

Univ. of the Saarland

2

0,595

BERT

4

Univ. of Illinois & IIT Kanpur

1

0,577

BERT

5

NITK Surathkal

1

0,574

No description

6

Vellore Institute of Technology, Chennai

2

0,552

TF / IDF, Random Forest

Overall, better F1 values ​​were achieved for English and Hindi. The distribution of all participants for two different F1 dimensions is shown in Figure 3.

Fig. 3:

Box plot of the distribution of the system results for English (Sub Task A).

Even at HASOC, the best systems deliver a similarly good quality. The top quartile of the systems is closely related. Overall, the F1 values ​​of the best systems are very similar to those of GermEval. And again, even without BERT, traditional systems achieve a quality that is close to that of the top systems. To illustrate the F1 values, the following table shows detailed results for the two most important classes for the best experiment.

Tab. 5:

Number of hits and errors of the best systems.

English

COURT

NOT

correctly identified

210

696

wrongly recognized

89

158

German

COURT

NOT

correctly identified

50

621

wrongly recognized

86

93

It is currently unclear whether an F1 value around 80% represents an upper limit for recognition, as people often cannot make a clear decision as to what constitutes hate speech. The next few years will show whether significant increases in quality are still possible.

Despite all the similarities, there are some differences between HASOC and GermEval. Hashtags were not rated as problematic at GermEval, only the text was decisive (e.g. #crimigrants or #rapefugees). HASOC classified this as hate speech in order to better appreciate the overall intention of a post.

Advanced language models such as Word Embeddings and BERT lead to the best results when it comes to identifying problematic content. These deep learning processes will certainly be improved in the coming years and adapted more strongly for languages ​​and tasks. For example, there is currently no specific variant for Hindi.

From the point of view of benchmark development, the question arises of how reliable the evaluation is at the moment and to what extent the results can be transferred to real environments. In the coming years, more in-depth analyzes will be necessary for a better understanding of the collections and their representativeness.

6 Conclusion

The development of procedures for the detection of problematic content represents a social challenge. Be it terrorist propaganda or misogyny: platforms and citizens should be enabled to identify problematic texts. This is a contribution to the confident handling of information. Research in this field shows their social responsibility. It must not only take place behind the closed doors of the Internet giants, but must take place in open forums in a comprehensible manner. Only in this way can the performance of the algorithms be discussed appropriately and social acceptance guaranteed. Transparency in particular seems to make an important contribution to acceptance (Brunk et al. 2019). The challenge for democratic societies is how to deal with problematic content without endangering freedom of expression.

The identification of hate speech is an example of the increasing performance of algorithms that have long been able to generate texts. Other examples of complex classification services are the recognition of irony, the finding of authors or stylistic inconsistencies in texts (Daelemans et al. 2019). These AI processes will leave their mark in all areas in which people work professionally with texts and knowledge representation.

bibliography

Al-Hassan, Areej; Al-Dossari, Hmood: Detection of hate speech in social networks: A Survey on multilingual corpus. In: Computer Science & Information Technology, 9.2 (2019), pp. 83-100. Search in Google Scholar

Brunk, Jens; Mattern, Jana; Riehle, Dennis M .: Effect of Transparency and Trust on Acceptance of Automatic Online Comment Moderation Systems. In: IEEE 21st Conference on Business Informatics, 1 (2019), pp. 429-435. Search in Google Scholar

Craswell, Nick et al .: TREC Overview Deep Learning Track. In: Proceedings 28th Text REtrieval Conference (2019), https://trec.nist.gov/proceedings/proceedings.html [Access: January 9, 2020]. Search in Google Scholar

Daelemans, Walter et al .: Overview of PAN 2019: Bots and gender profiling, celebrity profiling, cross-domain authorship attribution and style change detection. In: International Conference of the Cross-Language Evaluation Forum for European Languages. Basel 2019, pp. 402-416. Search in Google Scholar

Devlin, Jacob et al .: BERT: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings Conference North American Chapter of the Association for Computational Linguistics. Minneapolis, Minnesota 2019, pp. 4171-4186, https://www.aclweb.org/anthology/N19-1423 [Access: January 9, 2020]. Search in Google Scholar

Ghanem, Bilal; Rosso, Paolo; Rangel, Francisco: An Emotional Analysis of False Information in Social Media and News Articles (2019), https://arxiv.org/abs/1908.09951 [Access: January 9, 2020]. Search in Google Scholar

Hajok, Daniel; Selg, Olaf: Communication on the wrong track? A critical look at fake news and hate speech. In: JMS Jugend Medien Schutz-Report 41.4 (2018), pp. 2–6, https://www.nomos-elibrary.de/10.5771/0170-5067-2018-4-2.pdf [Access: 09.01.2020 ]. Search in Google Scholar

Henrich, Andreas: Information Retrieval 1 (Basics, Models and Applications), https://fis.uni-bamberg.de/handle/uniba/17569 [Access: January 9, 2020]. Search in Google Scholar

Kettemann, Matthias C .: International rules for social media: Protecting human rights and fighting disinformation. In: Global Governance Spotlight 2 (2019), http://hdl.handle.net/10419/203137 [Access: January 9, 2020]. Search in Google Scholar

Kruse, Rudolf et al .: Computational Intelligence: A Methodical Introduction to Artificial Neural Networks, Evolutionary Algorithms, Fuzzy Systems and Bayesian Networks. Basel 2015.Search in Google Scholar

Kuhlen, Rainer: The consequences of information assistants: What does informational autonomy mean or how can trust in electronic services be secured in open information markets? Frankfurt am Main 1999, Search in Google Scholar

Lewanczik, Niklas: Data protection by third parties? Zuckerberg's idea of ​​the globally regulated Internet (2019), https://onlinemarketing.de/news/datenschutz-dritte-zuckerbergs-global-regulierter-internet [Access: January 9, 2020]. Search in Google Scholar

Mandl, Thomas: Tolerant information retrieval with backpropagation networks. In: Neural Computing & Applications, 9.4 (2000), pp. 280-289. Search in Google Scholar

Mandl, Thomas et al .: Overview of the HASOC Track at FIRE 2019: Hate Speech and offensive content Identification in Indo-European Languages. In: Proceedings of the 11th Forum for Information Retrieval Evaluation (2019), ACM, pp. 14-17. Search in Google Scholar

De Smedt, Tom et al .: Multilingual Cross-domain Perspectives on Online Hate Speech. In: CLiPS Technical Report Series 8 (2018), pp. 1–24, https://arxiv.org/abs/1809.03944 [Access: January 9, 2020]. Search in Google Scholar

Smolensky, Paul: On the Proper Treatment of Connectionism. In: Behavioral and Brain Sciences 11.1 (1988), pp. 1-23. Search in Google Scholar

Struß, Julia Maria et al .: Overview of GermEval Task 2, 2019 Shared Task on the Identification of Offensive Language. Proceedings of the 15th Conference on Natural Language Processing (KONVENS 2019), https://ids-pub.bsz-bw.de/files/9319/Struss_etal._Overview_of_GermEval_task_2_2019.pdf [Access: 09.01.2020]. Search in Google Scholar

Vosoughi, Soroush; Roy, Deb; Aral, Sinan: The spread of true and false news online. In: Science 359.6380 (2018), pp. 1146-1151. Search in Google Scholar

Published Online: 2020-03-03
Published in Print: 2020-03-03

© 2020 Walter de Gruyter GmbH, Berlin / Boston