Neural network architecture for subtle hate speech detection
Additional background information is used with a trained neural network based model to help classify whether text is a subtly and/or ambiguously offensive. This additional background information can come from different sources such as the article on which the comment was made, world knowledge about the external entities (e.g., Wikipedia, Urban Dictionary), phrases referenced in the text being classified, and, the context of the previous comments/text in the thread. The background information is retrieved based on key entities (e.g., people, places things) and/or key phrases in the comment. Sentence matrix encodings are built for both the comment and the background information. The background information encoding is used to condition the comment encoding. The background information encoding, and the conditioned comment encoding are fed to a trained multi-level perceptron to classify the comment as hate speech or non-hate speech.
1. A method, comprising:
providing a neural network model which has been trained to identify hate speech for an input text sequence, the input text sequence extracted from a comment on a website, the neural network model comprising:
a first encoder that generates a first representation based on the input text sequence, the first encoder including a first embedding module that generates a sentence matrix based on the input text sequence, the first encoder including a first neural network encoder that generates the first representation based on the sentence matrix;
a second encoder that generates a second representation based on background information, the background information extracted from a source external to the comment on the website, the second encoder including a second embedding module that generates a background information matrix based on the background information, the second encoder including a second neural network encoder that generates the second representation based on the background information matrix;
a cross-text interaction module that generates a third representation based on the first representation conditioned on the second representation; and,
a neural network classifier that receives the second representation and the third representation;
receiving the input text sequence containing a first sequence of words;
receiving the background information containing a second sequence of words;
with a processor, determining a classification for the input text sequence using the neural network model; and,
outputting information based on the classification.
2. The method of claim 1 , wherein receiving the background information is performed by a background information retriever module.
3. The method of claim 1 , wherein the input text sequence comprises a comment posted in association with a news article.
4. The method of claim 3 , wherein the background information comprises a summary associated with the news article.
5. The method of claim 3 , wherein the background information comprises at least one previous comment posted in association with the news article.
6. The method of claim 3 , wherein the comment is posted to a first website and the background information comprises information retrieved from a second website external to the first website, the information retrieved from the second website being associated with an entity referenced in the comment by the second website.
7. The method of claim 3 , wherein the comment is posted to a first website and the background information comprises information retrieved from a second website external to the first website, the information retrieved from the second website being associated, by the second website, with a key phrase extracted from the comment.
8. A method, comprising:
receiving an input text sequence containing a first sequence of words, the input text sequence extracted from a comment on a website;
receiving background information containing a second sequence of words, the background information extracted from a source external to the comment on the website;
generating a sentence matrix based on the first sequence of words;
encoding the sentence matrix using a recurrent neural network based encoder to generate a first representation of the first sequence of words;
generating a background information matrix based on the second sequence of words;
encoding the background information matrix using the recurrent neural network based encoder to generate a second representation of the second sequence of words;
generating a third representation by conditioning the first representation using the second representation; and,
generating an output classification by inputting the third representation and the second representation to a trained multilevel perceptron classifier.
9. The method of claim 8 , further comprising:
outputting information based on the output classification.
10. The method of claim 8 , wherein conditioning the first representation using the second representation captures cross-text interactions to generate the third representation.
11. The method of claim 8 , wherein the sentence matrix is generated using word embedding.
12. The method of claim 8 wherein the background information matrix is generated using word embedding.
13. The method of claim 8 wherein the sentence matrix and the background information matrix are generated using word embedding.
14. The method of claim 8 wherein the recurrent neural network uses long short-term memory units as encoders.
15. A system, comprising:
memory that stores a neural network model which has been trained to identify hate speech for an input text sequence, the input text sequence extracted from a comment on a website, the neural network model comprising:
a first encoder that generates a first representation based on the input text sequence, the first encoder including a first embedding module that generates a sentence matrix based on the input text sequence, the first encoder including a first neural network encoder that generates the first representation based on the sentence matrix;
a second encoder that generates a second representation based on background information, the background information extracted from a source external to the comment on the website, the second encoder including a second embedding module that generates a background information matrix based on the background information, the second encoder including a second neural network encoder that generates the second representation based on the background information matrix;
a cross-text interaction module that generates a third representation based on the first representation conditioned on the second representation; and,
a neural network classifier that receives the second representation and the third representation;
an input component that receives the input text sequence;
a background information component that receives the background information;
a processor that determines a classification for the input text sequence using the neural network model; and,
an output component that outputs information based on the classification.
16. The system of claim 15 , wherein the input text sequence comprises a comment posted in association with a news article.
17. The system of claim 16 , wherein the background information comprises a summary associated with the news article.
18. The system of claim 17 , wherein the background information comprises at least one previous comment posted in association with the news article.
19. The system of claim 16 , wherein the comment is posted to a first website and the background information comprises information retrieved from a second website external to first website, the information retrieved from the second website being associated with an entity referenced in the comment by the second website.
20. The system of claim 16 , wherein the comment is posted to a first website and the background information comprises information retrieved from a second website external to first website, the information retrieved from the second website being associated, by the second website, with a key phrase extracted from the comment.