<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE article PUBLIC "-//NLM//DTD JATS (Z39.96) Journal Publishing DTD v1.3 20210610//EN" "JATS-journalpublishing1-3.dtd">
<article article-type="research-article" dtd-version="1.3" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xml:lang="ru"><front><journal-meta><journal-id journal-id-type="publisher-id">ntv</journal-id><journal-title-group><journal-title xml:lang="ru">Научно-технический вестник информационных технологий, механики и оптики</journal-title><trans-title-group xml:lang="en"><trans-title>Scientific and Technical Journal of Information Technologies, Mechanics and Optics</trans-title></trans-title-group></journal-title-group><issn pub-type="ppub">2226-1494</issn><issn pub-type="epub">2500-0373</issn><publisher><publisher-name>Университет ИТМО</publisher-name></publisher></journal-meta><article-meta><article-id custom-type="elpub" pub-id-type="custom">ntv-421</article-id><article-categories><subj-group subj-group-type="heading"><subject>Research Article</subject></subj-group><subj-group subj-group-type="section-heading" xml:lang="ru"><subject>КОМПЬЮТЕРНЫЕ СИСТЕМЫ И ИНФОРМАЦИОННЫЕ ТЕХНОЛОГИИ</subject></subj-group><subj-group subj-group-type="section-heading" xml:lang="en"><subject>COMPUTER SCIENCE</subject></subj-group></article-categories><title-group><article-title>Эффективный разреженный поиск с помощью построения инвертированного индекса на основе эмбеддингов</article-title><trans-title-group xml:lang="en"><trans-title>Efficient sparse retrieval through embedding-based inverted index construction</trans-title></trans-title-group></title-group><contrib-group><contrib contrib-type="author" corresp="yes"><contrib-id contrib-id-type="orcid">https://orcid.org/0009-0004-3056-8403</contrib-id><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Добрынин</surname><given-names>В. Ю.</given-names></name><name name-style="western" xml:lang="en"><surname>Dobrynin</surname><given-names>V. Yu.</given-names></name></name-alternatives><bio xml:lang="ru"><p>Добрынин Вячеслав Юрьевич — аспирант; старший разработчик</p><p>Санкт-Петербург, 197101</p><p>Лондон, W3 7XS</p></bio><bio xml:lang="en"><p>Viacheslav Yu. Dobrynin — PhD Student; Senior Developer</p><p>Saint Petersburg, 197101</p><p>London, W3 7XS</p></bio><email xlink:type="simple">vidobrynin@itmo.ru</email><xref ref-type="aff" rid="aff-1"/></contrib><contrib contrib-type="author" corresp="yes"><contrib-id contrib-id-type="orcid">https://orcid.org/0009-0005-5397-2772</contrib-id><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Абрамович</surname><given-names>Р. К.</given-names></name><name name-style="western" xml:lang="en"><surname>Abramovich</surname><given-names>R. K.</given-names></name></name-alternatives><bio xml:lang="ru"><p>Абрамович Роман Константинович — аспирант; старший разработчик</p><p>Санкт-Петербург, 197101</p><p>Лондон, E14 4QA</p></bio><bio xml:lang="en"><p>Roman K. Abramovich — PhD Student; Senior Developer</p><p>Saint Petersburg, 197101</p><p>London, E14 4QA</p></bio><email xlink:type="simple">asmetliness24237@gmail.com</email><xref ref-type="aff" rid="aff-2"/></contrib><contrib contrib-type="author" corresp="yes"><contrib-id contrib-id-type="orcid">https://orcid.org/0000-0002-8485-1296</contrib-id><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Платонов</surname><given-names>А. В.</given-names></name><name name-style="western" xml:lang="en"><surname>Platonov</surname><given-names>A. V.</given-names></name></name-alternatives><bio xml:lang="ru"><p>Платонов Алексей Владимирович — кандидат технических наук, доцент</p><p>Санкт-Петербург, 197101</p></bio><bio xml:lang="en"><p>Alexey V. Platonov — PhD, Associate Professor</p><p>Saint Petersburg, 197101</p></bio><email xlink:type="simple">avplatonov@itmo.ru</email><xref ref-type="aff" rid="aff-3"/></contrib></contrib-group><aff-alternatives id="aff-1"><aff xml:lang="ru"><institution>Университет ИТМО; Shift Lab LTD</institution><country>Россия</country></aff><aff xml:lang="en"><institution>ITMO University; Shift Lab LTD</institution><country>Russian Federation</country></aff></aff-alternatives><aff-alternatives id="aff-2"><aff xml:lang="ru"><institution>Университет ИТМО; Payler Ltd</institution><country>Россия</country></aff><aff xml:lang="en"><institution>ITMO University; Payler Ltd</institution><country>Russian Federation</country></aff></aff-alternatives><aff-alternatives id="aff-3"><aff xml:lang="ru"><institution>Университет ИТМО</institution><country>Россия</country></aff><aff xml:lang="en"><institution>ITMO University</institution><country>Russian Federation</country></aff></aff-alternatives><pub-date pub-type="collection"><year>2025</year></pub-date><pub-date pub-type="epub"><day>28</day><month>02</month><year>2025</year></pub-date><volume>25</volume><issue>1</issue><fpage>61</fpage><lpage>67</lpage><permissions><copyright-statement>Copyright &amp;#x00A9; Добрынин В.Ю., Абрамович Р.К., Платонов А.В., 2025</copyright-statement><copyright-year>2025</copyright-year><copyright-holder xml:lang="ru">Добрынин В.Ю., Абрамович Р.К., Платонов А.В.</copyright-holder><copyright-holder xml:lang="en">Dobrynin V.Y., Abramovich R.K., Platonov A.V.</copyright-holder><license xml:lang="ru" license-type="creative-commons-attribution" xlink:href="https://creativecommons.org/licenses/by/4.0/" xlink:type="simple"><license-p>Данная работа распространяется под лицензией Creative Commons Attribution 4.0.</license-p></license><license xml:lang="en" license-type="creative-commons-attribution" xlink:href="https://creativecommons.org/licenses/by/4.0/" xlink:type="simple"><license-p>This work is licensed under a Creative Commons Attribution 4.0 License.</license-p></license></permissions><self-uri xlink:href="https://ntv.elpub.ru/jour/article/view/421">https://ntv.elpub.ru/jour/article/view/421</self-uri><abstract><p>Введение. Современные поисковые системы используют двухэтапную архитектуру для эффективного и качественного поиска по большим объемам данных. На первом этапе применяются простые и быстрые алгоритмы, такие как BM25, а на втором — более точные, но ресурсоемкие методы, например глубокие нейронные сети. Несмотря на то, что такой подход показывает хорошие результаты, он фундаментально ограничен по качеству из-за проблемы несовпадения словарей, что присуще простым алгоритмам первого этапа. Метод. Для решения проблемы ограничений качества поиска, в настоящей работе предлагается алгоритм построения инвертированного индекса с использованием векторных представлений. Представленный подход объединяет преимущества обоих этапов: эффективность инвертированного индекса и высокое качество поиска при использовании векторных моделей. Предложено создание векторного индекса, сохраняющего различные семантические значения токенов словаря. Для каждого токена определяются документы, в которых он используется, после чего его контекстуализированные эмбеддинги кластеризуются. Центроиды полученных кластеров представляют различные семантические значения токенов. Таким образом, формируется расширенный словарь, который применяется для построения инвертированного индекса. При построении индекса вычисляются оценки близости между каждым семантическим значением токена и документами, что затем используется в процессе поиска. Это позволяет сократить количество вычислений для оценки близости в режиме реального времени. Поиск по инвертированному индексу требует нахождения ключей в векторном индексе, что позволяет решить проблему несовпадения словарей. Основные результаты. Работа алгоритма продемонстрирована на задаче поиска в наборе данных SciFact. Показано, что предлагаемый метод обеспечивает высокое качество поиска при низких требованиях к объему памяти. Обсуждение. Разработанный алгоритм демонстрирует высокое качество поиска, при этом он поддерживает компактный векторный индекс, размер которого остается неизменным и определяется исключительно размерами словаря. Основным недостатком алгоритма является необходимость использования глубокой нейронной сети для генерации векторных представлений запроса в процессе поиска, что замедляет этот этап. Поиск путей для решения данной проблемы и сокращения времени поиска представляет собой направление дальнейших исследований.</p></abstract><trans-abstract xml:lang="en"><p>Modern search engines use a two-stage architecture for efficient and high-quality search over large volumes of data. In the first stage, simple and fast algorithms like BM25 are applied, while in the second stage, more precise but resourceintensive methods methods, such as deep neural networks, are employed. Although this approach yields good results, it is fundamentally limited in quality due to the vocabulary mismatch problem inherent in the simple algorithms of the first stage. To address this issue, we propose an algorithm for constructing an inverted index using vector representations combining the advantages of both stages: the efficiency of the inverted index and the high search quality of vector models. In our work, we suggest creating a vector index that preserves the various semantic meanings of vocabulary tokens. For each token, we identify the documents in which it is used, and then cluster its contextualized embeddings. The centroids of the resulting clusters represent different semantic meanings of the tokens. This process forms an extended vocabulary which is used to build the inverted index. During index construction, similarity scores between each semantic meaning of a token and documents are calculated which are then used in the search process. This approach reduces the number of computations required for similarity estimation in real-time. Searching the inverted index first requires finding keys in the vector index, helping to solve the vocabulary mismatch problem. The operation of the algorithm is demonstrated on a search task within the SciFact dataset. It is shown that the proposed method achieves high search quality with low memory requirements. The proposed algorithm demonstrates high search quality, while maintaining a compact vector index whose size remains constant and depends only on the size of the vocabulary. The main drawback of the algorithm is the need to use a deep neural network to generate vector representations of queries during the search process which slows down this stage. Finding ways to address this issue and accelerate the search process represents a direction for future research.</p></trans-abstract><kwd-group xml:lang="ru"><kwd>инвертированный индекс</kwd><kwd>проблема несоответствия словарей</kwd><kwd>нейронные сети</kwd><kwd>векторные представления</kwd><kwd>кластеризация</kwd></kwd-group><kwd-group xml:lang="en"><kwd>inverted index</kwd><kwd>vocabulary mismatch problem</kwd><kwd>neural networks</kwd><kwd>vector representations</kwd><kwd>clusterization</kwd></kwd-group></article-meta></front><back><ref-list><title>References</title><ref id="cit1"><label>1</label><citation-alternatives><mixed-citation xml:lang="ru">Khattab O., Zaharia M. ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT // Proc. of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 2020. P. 39–48. https://doi.org/10.1145/3397271.3401075</mixed-citation><mixed-citation xml:lang="en">Khattab O., Zaharia M. ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT // Proc. of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 2020. P. 39–48. https://doi.org/10.1145/3397271.3401075</mixed-citation></citation-alternatives></ref><ref id="cit2"><label>2</label><citation-alternatives><mixed-citation xml:lang="ru">Zamani H., Dehghani M., Croft W.B., Learned-Miller E., Kamps J. From neural re-ranking to neural ranking: Learning a sparse representation for inverted indexing // Proc. of the 27th ACM International Conference on Information and Knowledge Management. 2018. P. 497–506. https://doi.org/10.1145/3269206.3271800</mixed-citation><mixed-citation xml:lang="en">Zamani H., Dehghani M., Croft W.B., Learned-Miller E., Kamps J. From neural re-ranking to neural ranking: Learning a sparse representation for inverted indexing // Proc. of the 27th ACM International Conference on Information and Knowledge Management. 2018. P. 497–506. https://doi.org/10.1145/3269206.3271800</mixed-citation></citation-alternatives></ref><ref id="cit3"><label>3</label><citation-alternatives><mixed-citation xml:lang="ru">Bai Y., Li X., Wang G., Zhang C., Shang L., Xu J., Wang Z., Wang F., Liu Q. SparTerm: Learning term-based sparse representation for fast text retrieval // arXiv. 2020. arXiv:2010.00768. https://doi.org/10.48550/arXiv.2010.00768</mixed-citation><mixed-citation xml:lang="en">Bai Y., Li X., Wang G., Zhang C., Shang L., Xu J., Wang Z., Wang F., Liu Q. SparTerm: Learning term-based sparse representation for fast text retrieval // arXiv. 2020. arXiv:2010.00768. https://doi.org/10.48550/arXiv.2010.00768</mixed-citation></citation-alternatives></ref><ref id="cit4"><label>4</label><citation-alternatives><mixed-citation xml:lang="ru">Devlin J., Chang M.W., Lee K., Toutanova K. BERT: Pre-training of deep bidirectional transformers for language understanding // arXiv. 2018. arXiv:1810.04805v2. https://doi.org/10.48550/arXiv.1810.04805</mixed-citation><mixed-citation xml:lang="en">Devlin J., Chang M.W., Lee K., Toutanova K. BERT: Pre-training of deep bidirectional transformers for language understanding // arXiv. 2018. arXiv:1810.04805v2. https://doi.org/10.48550/arXiv.1810.04805</mixed-citation></citation-alternatives></ref><ref id="cit5"><label>5</label><citation-alternatives><mixed-citation xml:lang="ru">Formal T., Piwowarski B., Clinchant S. SPLADE: Sparse lexical and expansion model for first stage ranking // Proc. of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2021. P. 2288–2292. https://doi.org/10.1145/3404835.3463098</mixed-citation><mixed-citation xml:lang="en">Formal T., Piwowarski B., Clinchant S. SPLADE: Sparse lexical and expansion model for first stage ranking // Proc. of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2021. P. 2288–2292. https://doi.org/10.1145/3404835.3463098</mixed-citation></citation-alternatives></ref><ref id="cit6"><label>6</label><citation-alternatives><mixed-citation xml:lang="ru">Santhanam K., Khattab O., Saad-Falcon J., Potts C., Zaharia M. ColBERTv2: Effective and efficient retrieval via lightweight late interaction // Proc. of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022. P. 3715–3734. https://doi.org/10.18653/v1/2022.naacl-main.272</mixed-citation><mixed-citation xml:lang="en">Santhanam K., Khattab O., Saad-Falcon J., Potts C., Zaharia M. ColBERTv2: Effective and efficient retrieval via lightweight late interaction // Proc. of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022. P. 3715–3734. https://doi.org/10.18653/v1/2022.naacl-main.272</mixed-citation></citation-alternatives></ref><ref id="cit7"><label>7</label><citation-alternatives><mixed-citation xml:lang="ru">Johnson J., Douze M., Jégou H. Billion-scale similarity search with GPUs // IEEE Transactions on Big Data. 2021. V. 7. N 3. P. 535–547. https://doi.org/10.1109/tbdata.2019.2921572</mixed-citation><mixed-citation xml:lang="en">Johnson J., Douze M., Jégou H. Billion-scale similarity search with GPUs // IEEE Transactions on Big Data. 2021. V. 7. N 3. P. 535–547. https://doi.org/10.1109/tbdata.2019.2921572</mixed-citation></citation-alternatives></ref><ref id="cit8"><label>8</label><citation-alternatives><mixed-citation xml:lang="ru">Jégou H., Douze M., Schmid C. Product quantization for nearest neighbor search // IEEE Transactions on Pattern Analysis and Machine Intelligence. 2011. V. 33. N 1. P. 117–128. https://doi.org/10.1109/tpami.2010.57</mixed-citation><mixed-citation xml:lang="en">Jégou H., Douze M., Schmid C. Product quantization for nearest neighbor search // IEEE Transactions on Pattern Analysis and Machine Intelligence. 2011. V. 33. N 1. P. 117–128. https://doi.org/10.1109/tpami.2010.57</mixed-citation></citation-alternatives></ref><ref id="cit9"><label>9</label><citation-alternatives><mixed-citation xml:lang="ru">Kong W., Dudek J.M., Li C., Zhang M., Bendersky M. SparseEmbed: Learning sparse lexical representations with contextual embeddings for retrieval // Proc. of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2023. P. 2399–2403. https://doi.org/10.1145/3539618.3592065</mixed-citation><mixed-citation xml:lang="en">Kong W., Dudek J.M., Li C., Zhang M., Bendersky M. SparseEmbed: Learning sparse lexical representations with contextual embeddings for retrieval // Proc. of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2023. P. 2399–2403. https://doi.org/10.1145/3539618.3592065</mixed-citation></citation-alternatives></ref><ref id="cit10"><label>10</label><citation-alternatives><mixed-citation xml:lang="ru">Tiancheng Z., Lu X., Lee K. SPARTA: Efficient Open-Domain Question Answering via Sparse Transformer Matching Retrieval // Proc. of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021. P. 565–575. https://doi.org/10.18653/v1/2021.naacl-main.47</mixed-citation><mixed-citation xml:lang="en">Tiancheng Z., Lu X., Lee K. SPARTA: Efficient Open-Domain Question Answering via Sparse Transformer Matching Retrieval // Proc. of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021. P. 565–575. https://doi.org/10.18653/v1/2021.naacl-main.47</mixed-citation></citation-alternatives></ref><ref id="cit11"><label>11</label><citation-alternatives><mixed-citation xml:lang="ru">Dobrynin V., Abramovich R., Platonov A. Building a full-text search index using “Transformer” neural network // Proc. of the 2023 IEEE 17th International Conference on Application of Information and Communication Technologies (AICT). 2023. P. 1–5. https://doi.org/10.1109/aict59525.2023.10313200</mixed-citation><mixed-citation xml:lang="en">Dobrynin V., Abramovich R., Platonov A. Building a full-text search index using “Transformer” neural network // Proc. of the 2023 IEEE 17th International Conference on Application of Information and Communication Technologies (AICT). 2023. P. 1–5. https://doi.org/10.1109/aict59525.2023.10313200</mixed-citation></citation-alternatives></ref><ref id="cit12"><label>12</label><citation-alternatives><mixed-citation xml:lang="ru">Malkov Y.A., Yashunin D.A. Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs // IEEE Transactions on Pattern Analysis and Machine Intelligence. 2020. V. 42. N 4. P. 824–836. https://doi.org/10.1109/tpami.2018.2889473</mixed-citation><mixed-citation xml:lang="en">Malkov Y.A., Yashunin D.A. Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs // IEEE Transactions on Pattern Analysis and Machine Intelligence. 2020. V. 42. N 4. P. 824–836. https://doi.org/10.1109/tpami.2018.2889473</mixed-citation></citation-alternatives></ref><ref id="cit13"><label>13</label><citation-alternatives><mixed-citation xml:lang="ru">Arthur D., Vassilvitskii S. k-means++: the advantages of careful seeding // Proc. of the 18th ACM-SIAM Symposium on Discrete Algorithms Mathematics. 2007. P. 1027–1035.</mixed-citation><mixed-citation xml:lang="en">Arthur D., Vassilvitskii S. k-means++: the advantages of careful seeding // Proc. of the 18th ACM-SIAM Symposium on Discrete Algorithms Mathematics. 2007. P. 1027–1035.</mixed-citation></citation-alternatives></ref><ref id="cit14"><label>14</label><citation-alternatives><mixed-citation xml:lang="ru">Bajaj P., Campos D., Craswell N., Deng L., Gao J., Liu X., Majumder R., McNamara A., Mitra B., Nguyen T., Rosenberg M., Song X., Stoica A, Tiwary S., Wang T. MS MARCO: a human generated MAchine Reading COmprehension dataset // arXiv. 2016. arXiv:1611.09268. https://doi.org/10.48550/arXiv.1611.09268</mixed-citation><mixed-citation xml:lang="en">Bajaj P., Campos D., Craswell N., Deng L., Gao J., Liu X., Majumder R., McNamara A., Mitra B., Nguyen T., Rosenberg M., Song X., Stoica A, Tiwary S., Wang T. MS MARCO: a human generated MAchine Reading COmprehension dataset // arXiv. 2016. arXiv:1611.09268. https://doi.org/10.48550/arXiv.1611.09268</mixed-citation></citation-alternatives></ref><ref id="cit15"><label>15</label><citation-alternatives><mixed-citation xml:lang="ru">Thakur N., Nils R., Rücklé A., Srivastava A., Gurevych I. BEIR: A heterogenous benchmark for zero-shot evaluation of information retrieval models // arXiv. 2021. arXiv:2104.08663. https://doi.org/10.48550/arXiv.2104.08663</mixed-citation><mixed-citation xml:lang="en">Thakur N., Nils R., Rücklé A., Srivastava A., Gurevych I. BEIR: A heterogenous benchmark for zero-shot evaluation of information retrieval models // arXiv. 2021. arXiv:2104.08663. https://doi.org/10.48550/arXiv.2104.08663</mixed-citation></citation-alternatives></ref></ref-list><fn-group><fn fn-type="conflict"><p>The authors declare that there are no conflicts of interest present.</p></fn></fn-group></back></article>
