Doctoral thesis: Technology that's better at detecting hate speech may also increase security risks

Doctoral candidate Tommi Gröndahl will defend his doctoral dissertation on language technologies at Aalto University on 23 August. Language technology is one of the most common forms of artificial intelligence, and Gröndahl tested its methods for detecting lies and hate speech.

"Artificial intelligence is needed when massive amounts of data need to be processed and the targeted material is screened out. However, language technologies have been a black-box, which means that they have been utilised without actually understanding how they process language,“ Gröndahl says.

Gröndahl was particularly interested in the consequences of the methods from the perspective of information security and privacy protection.

"If language technology methods are developed, it improves the possibilities of both content moderation and censoring. The techniques are the same, regardless of the consequences the classification of the text has for the author,” says Gröndahl.

In the study, Gröndahl compared deep learning neural networks with more traditional rule-based methods. In the rule-based methods, a person creates a rule in the system, which is then utilised in automation. Major differences were found between the methods.

"Complex neural network models have a vast amount of numbers, which makes it difficult to know which feature of the text each number represents, and what happens to the text when one of the numbers is changed. When a text needs detailed structuring, the most commonly used neural network models are not always reliable. For example, a neural network model cannot distinguish between a sentence and its negative equivalent,” explains Gröndahl.

Then again, rule-based methods are not ideal for screening vast amounts of data. It is therefore essential to combine the best of the two methods.

Methods easily deceived

The classification of text is typically based on fairly simple features, such as specific words. For example, in the case of lie detection, the methods do not actually detect lies, but simple features in the data. Complex machine learning models catch such features similarly to simpler models. This makes models vulnerable to attacks; for instance, hate speech detectors are easily fooled by people, when spaces are removed from the text or the word ‘love’ is added to the text.

Gröndahl found that the assigned task and training data affect how well an AI-based classifier succeeds in the task.

"As machine learning models are massive, it is important that the classifier receives enough training data. It can also easily be caught out by undesirable features if the training data is distorted, i.e. in some way unrepresentative. Complex machine learning models do not know when to take into account a word or a character and when not, but they act on the basis of the training data provided,” says Gröndahl.

In addition to hate speech and lie detection, Gröndahl also studied the possibilities of language technology for automatic alteration of writing style and automatic text editing. The objective may be, for instance, to modify the text so that the author cannot be identified.

For example, when the writing style was automatically altered, the neural network produced a conversion similar to a machine translation that could result in repeating the same text or changing the meaning of the text content. In rule-based methods, it is possible to control in more detail, for example, the effect of a certain word, such as a synonym or a contradiction, on the conversion of a sentence.