В работе рассматривается задача поиска почти-дубликатов текстов школьных сочинений в больших коллекциях данных. Предпосылками к решению данной задачи является возможность школьников применять для написания выпускных сочинений заранее заготовленные тексты, в том числе полученные из открытых коллекций школьных сочинений. Актуальность задачи подтверждается работами [1, 2], посвященными анализу нарушений при написании академических испытаний, а также частичному переходу школьного образования на удаленный режим. Задача поиска почти-дубликатов рассматривается как задача информационного поиска, где сочинению ставится в соответствие заимствованный текст из коллекции. В рамках рассматриваемой задачи сочинение представляется набором изображений рукописного текста, написанного автором, в то время как документы из коллекции представимы в виде машиночитаемых текстов.

The article considers the problem of multilingual summarization generation for scientific documents. To solve this, we propose a method based on the summarization-translation approach. The method decomposes the original task into two separate tasks: monolingual document summarization and multilingual summarization. In the first task monolingual summary is generated for the document in the document language. In the second task resulting monolingual summary is translated into the language of interest. The paper analyses different abstractive and extractive models to choose the optimal one as a monolingual summarization model. The best model is selected based on the rouge metric, as well as on the newly proposed metrics. The multilingual summarization model uses Moses statistical machine translation model and post-processing based on the mT5 transformer model. The proposed system was tested on the Wikipedia dataset for 15 different languages. It is shown that the proposed system of models can generate multilingual summaries in 15 languages.

The automatic generation of queries to a search engine based on the incoming text is important for question-answering, recommendation, and text reuse detection systems. Every such query requires resources from a user and a search engine itself. A method of ensembling query generation models that maximizes the search completeness metric for the minimum number of queries could be useful. The task of selecting the best model or an ensemble of models is trivial for the case of a fixed search engine. However, real search engines are constantly changing their behavior, learning on incoming data, changing their index of web pages and documents. They are black boxes for a user. In this paper we propose an approach to ensemble query generation models based on reinforcement learning. By reformulating the problem so that the agent selects a sequence of models rather than a single query generation model, we guarantee maximum retrieval recall even when the worst possible action is selected. As a reward, we introduce a discount recall metric that penalizes the agent for each extra step of a model request. We modify the UCB learning algorithm so that the re-initialization of the recidivism penalty matrix occurs independently of the engine index state. In this way, we ensure that the top 3 best actions (i.e. sequences of generation model requests) are found in just 5 epochs, each epoch contains 1050 documents. The model ensemble maintains a stable performance even when the index alters in a way that the ensemble was not informed about.

Численным моделированием продемонстрирован эффект теплоизоляции термоядерного топлива при имплозии топливных полостей в протяженной оболочке из тяжелых элементов. Найдены конечные условия имплозии, достижимые с учетом технологических ограничений.

Currently, more efforts are being made to improve the capabilities of Large Language Models than to address their implications. Modern language models are capable of generating texts that appear indistinguishable from those written by human experts. While providing a high quality of life, such breakthroughs at the same time pose new challenges in education, science and social media. In addition, existing approaches to detect texts created by artificial intelligence either require high computational cost or access to the internal computation of LLMs, which in turn hinders their public availability. Based on these considerations, this paper presents a new paradigm for detecting texts created by artificial intelligence based on on collecting preliminary token statistics and computing n-gram perplexity features. On the combination of HC3, M4GT and MAGE datasets it shows a speedup of 2x over existing approaches with a quality drop around 5%. Moreover, the combination of methods achieves the best quality. This strikes a balance between computational cost, accessibility and performance.