AI vs AI: Scientists Develop Neural Networks to Detect Generated Text Insertions
A research team, including Alexander Shirnin from HSE University, has developed two models designed to detect AI-generated insertions in scientific texts. The AIpom system integrates two types of models: a decoder and an encoder. The Papilusion system is designed to detect modifications through synonyms and summarisation by neural networks, using one type of models: encoders. In the future, these models will assist in verifying the originality and credibility of scientific publications. Articles describing the Papilusion and AIpom systems have been published in the ACL Anthology Digital Archive.
As language models like ChatGPT and GigaChat become more popular and widely used, it becomes increasingly challenging to distinguish original human-written text from AI-generated content. Artificial intelligence is already being used to write scientific publications and graduation papers. Therefore, it is crucial to develop tools capable of identifying AI-generated insertions in texts. A research team, including scientists from HSE University, presented their solutions at the SemEval 2024 and DAGPap24 international scientific competitions.
The AIpom model was used to identify the boundaries between original and generated fragments in scientific papers. In each paper, the proportion of machine-generated text to the author's text varied. To train the models, the organisers provided texts on the same topic. However, during the verification stage, the topics changed, making the task more challenging.
Alexander Shirnin
'Models perform well on familiar topics, but their performance declines when presented with new topics,' according to Alexander Shirnin, co-author of the paper and Research Assistant at the Laboratory for Models and Methods of Computational Pragmatics, HSE Faculty of Computer Science. 'It's like a student who, having learned how to solve one type of problem, struggles to solve a problem on an unfamiliar topic or from a different subject as easily or accurately.'
To improve the system's performance, the researchers combined two models: a decoder and an encoder. At the first stage, a neural network decoder was used, with the input consisting of an instruction and the source text, and the output being a text fragment presumably generated by AI. Next, in the original text, the area where the model predicted the beginning of a generated fragment was highlighted using a special <BREAK> token. The encoder then processed the text marked up in the first stage and refined the decoder's predictions. To do this, it categorised each token—the smallest unit of text, such as a word or part of a word—and identified whether it was written by a human or generated by AI. This approach improved accuracy compared to systems that used only one type of model: AIpom ranked second at the SemEval-2024 competition.
The Papilusion model also distinguished between written text and generated text. Using Papilusion, sections of the text were classified into four categories: written by a human, modified with synonyms, generated, or summarised by a model. The task was to accurately identify each category. The number of categories and the length of insertions in the texts varied.
In this case, the developers used three models, all of the same type: encoders. They were trained to predict one of the four categories for each token in the text, with each model trained independently of the others. When a model made an error, a cost was applied, and the model was retrained with the lower layers frozen.
'Each model has a different number of layers, depending on its architecture. When training a model, we can leave the first ten or so layers unchanged and adjust only the parameters in the last two layers. This is done to prevent losing important data embedded in the first layers during training,' explains Alexander Shirnin. 'It can be compared to an athlete who makes an error in the movement of their hand. We only need to explain this part to them, rather than resetting their entire learning and retraining them, as they might forget how to move correctly overall. The same logic applies here. The method is not universal and may not work with all models, but in our case, it was effective.'
The three encoders independently determined the category for each token (word). The system's final prediction was based on the category that received the most points. Papilusion ranked sixth out of 30 in the competition.
According to the researchers, current AI detection models perform reasonably well but still have limitations. Primarily, they struggle to process data beyond what they were trained on, and overall, there is a lack of diverse data to train the models effectively.
'To obtain more data, we need to focus on collecting it. Both companies and laboratories have been doing this. Specifically for this type of task, it is necessary to collect datasets that include texts modified using multiple AI models and modification methods,' the researcher comments. 'Instead of continuing a text using just one model, more realistic scenarios should be created, such as asking the model to add to the text, rewrite the beginning for better coherence, remove parts of it, or generate a portion of the text in a new style using a different prompt. Of course, it is also important to collect data in different languages and on a variety of topics.'
See also:
Scientists Examine Neurobiology of Pragmatic Reasoning
An international team including scientists from HSE University has investigated the brain's ability to comprehend hidden meanings in spoken messages. Using fMRI, the researchers found that unambiguous meanings activate brain regions involved in decision-making, whereas processing complex and ambiguous utterances engages regions responsible for analysing context and the speaker's intentions. The more complex the task, the greater the interaction between these regions, enabling the brain to decipher the meaning. The study has been published in NeuroImage.
‘HSE’s Industry Ties Are Invaluable’
Pan Zhengwu has spent the last seven years at HSE University—first as a student of the Bachelor’s in Software Engineering and now in the Master’s in System and Software Engineering at the Faculty of Computer Science. In addition to his busy academic schedule, he works as a mobile software engineer at Yandex and is an avid urban photographer. In his interview with the HSE News Service, Zhengwu talks about the challenges he faced when he first moved to Russia, shares his thoughts on ‘collaborating’ with AI, and reveals one of his top spots for taking photos in Moscow.
Scientists Present New Solution to Imbalanced Learning Problem
Specialists at the HSE Faculty of Computer Science and Sber AI Lab have developed a geometric oversampling technique known as Simplicial SMOTE. Tests on various datasets have shown that it significantly improves classification performance. This technique is particularly valuable in scenarios where rare cases are crucial, such as fraud detection or the diagnosis of rare diseases. The study's results are available on ArXiv.org, an open-access archive, and will be presented at the International Conference on Knowledge Discovery and Data Mining (KDD) in summer 2025 in Toronto, Canada.
Hi-Tech Grief: HSE Researchers Explore the Pros and Cons of Digital Commemoration
Researchers at HSE University in Nizhny Novgorod have explored how technological advancements are transforming the ways in which people preserve the memory of the deceased and significant events. Digital technologies enable the creation of virtual memorials, the preservation of personal stories and belongings of the deceased, interaction with their digital footprint, and even the development of interactive avatars based on their online activity. However, these technologies not only evoke nostalgia and provide a sense of relief but can also heighten anxiety and fear, and delay the process of accepting loss. The study has been published in Chelovek (The Human Being).
Scientists Find Out Why Aphasia Patients Lose the Ability to Talk about the Past and Future
An international team of researchers, including scientists from the HSE Centre for Language and Brain, has identified the causes of impairments in expressing grammatical tense in people with aphasia. They discovered that individuals with speech disorders struggle with both forming the concept of time and selecting the correct verb tense. However, which of these processes proves more challenging depends on the speaker's language. The findings have been published in the journal Aphasiology.
Implementation of Principles of Sustainable Development Attracts More Investments
Economists from HSE and RUDN University have analysed issues related to corporate digital transformation processes. The introduction of digital solutions into corporate operations reduces the number of patents in the field of green technologies by 4% and creates additional financial difficulties. However, if a company focuses on sustainable development and increases its rating in environmental, social, and governance performance (ESG), the negative effects decrease. Moreover, when the ESG rating is high, digitalisation can even increase the number of patents by 2%. The article was published in Sustainability.
Russian Scientists Develop New Compound for Treating Aggressive Tumours
A team of Russian researchers has synthesised a novel compound for boron neutron capture therapy (BNCT), a treatment for advanced cancer that uses the boron-10 isotope. The compound exhibits low toxicity, excellent water solubility, and eliminates the need for administering large volumes. Most importantly, the active substance reaches the tumour with minimal impact on healthy tissues. The study was published in the International Journal of Molecular Sciences shortly before World Cancer Day, observed annually on February 4.
Scientists Discover Link Between Brain's Structural Features and Autistic Traits in Children
Scientists have discovered significant structural differences in the brain's pathways, tracts, and thalamus between children with autism and their neurotypical peers, despite finding no functional differences. The most significant alterations were found in the pathways connecting the thalamus—the brain's sensory information processing centre—to the temporal lobe. Moreover, the severity of these alterations positively correlated with the intensity of the child's autistic traits. The study findings have been published in Behavioural Brain Research.
Earnings Inequality Declining in Russia
Earnings inequality in Russia has nearly halved over the past 25 years. The primary factors driving this trend are rising minimum wages, regional economic convergence, and shifts in the returns on education. Since 2019, a new phase of this process has been observed, with inequality continuing to decline but driven by entirely different mechanisms. These are the findings made by Anna Lukyanova, Assistant Professor at the HSE Faculty of Economic Sciences, in her new study. The results have been published in the Journal of the New Economic Association.
Russian Physicists Discover Method to Increase Number of Atoms in Quantum Sensors
Physicists from the Institute of Spectroscopy of the Russian Academy of Sciences and HSE University have successfully trapped rubidium-87 atoms for over four seconds. Their method can help improve the accuracy of quantum sensors, where both the number of trapped atoms and the trapping time are crucial. Such quantum systems are used to study dark matter, refine navigation systems, and aid in mineral exploration. The study findings have been published in the Journal of Experimental and Theoretical Physics Letters.