Tatarstan Toponyms: A Bilingual Dataset and Hybrid RAG System for Geospatial Question Answering

# Tatarstan Toponyms: A Bilingual Dataset and Hybrid RAG System for Geospatial Question Answering
Source: [https://arxiv.org/html/2605.05962](https://arxiv.org/html/2605.05962)
[![[Uncaptioned image]](https://arxiv.org/html/2605.05962v1/x1.png)M\. K\. Arabov](https://orcid.org/0000-0003-2525-1183) Department of Data Analysis and Programming Technologies Institute of Computational Mathematics and Information Technologies Kazan \(Volga Region\) Federal University Kazan, Russia MKArabov@kpfu\.ru

###### Abstract

This paper addresses the end\-to\-end task of automatic geospatial question answering over multilingual toponymic data\. An original bilingual dataset of toponyms of the Republic of Tatarstan is introduced, comprising 9,688 structured records with detailed linguistic, etymological, administrative, and coordinate information \(93\.1% of objects are georeferenced\)\. Based on this dataset, a specialized question\-answering corpus of approximately 39,000 “question–context–extractable answer” triples is constructed, with guaranteed answer localization within the text\. To solve the task, an architecture combining two key components is proposed: a hybrid retriever that integrates dense semantic indexing using multilingual\-e5\-large with a geospatial filter and ranking \(KD\-trees, haversine distance\), and an extractive reader based on fine\-tuned transformer models\. On 500 test queries, the hybrid search achieves Recall@1 = 0\.988, Recall@5 = 1\.000, and MRR = 0\.994, statistically significantly outperforming both BM25 and the purely spatial method\. Among the tested reader architectures \(RuBERT, XLM\-RoBERTa\-large, T5\-RUS\), the best answer extraction quality is attained by the multilingual XLM\-RoBERTa\-large model: EM = 0\.992, F1 = 0\.994\. A contrasting effect is observed: on raw outputs, RuBERT\-based models fail to answer coordinate\-related questions \(F1 = 0\), whereas XLM\-RoBERTa\-large achieves F1 = 0\.984; however, simple post\-processing completely eliminates the gaps in numerical values and restores RuBERT accuracy to 100%\. This discrepancy is attributed to tokenization specifics and the composition of pre\-training corpora\. The created resources \(dataset, QA corpus, trained model weights, and web demonstrator\) are openly published on the Hugging Face platform\. The obtained results can be directly applied in the development of geospatial question\-answering services, geocoding systems, and digital humanities projects for multilingual regions\.

*Keywords*Tatarstan toponyms⋅\\cdotquestion answering⋅\\cdotRAG⋅\\cdotextractive reading⋅\\cdothybrid search⋅\\cdotsemantic embeddings⋅\\cdotgeospatial analysis⋅\\cdotbilingual dataset⋅\\cdotTatar language⋅\\cdotRussian language⋅\\cdotXLM\-RoBERTa⋅\\cdotmultilingual E5⋅\\cdotnatural language processing

## 1Introduction

Geographical names constitute a fundamental component of spatial data infrastructure and play a key role in cartography, navigation, regional governance, and digital humanities research\(Suleymanovet al\.,[2021](https://arxiv.org/html/2605.05962#bib.bib1); Galimovet al\.,[2023](https://arxiv.org/html/2605.05962#bib.bib2)\)\. In multilingual regions such as the Republic of Tatarstan, where Russian and Tatar are both official state languages, toponyms function simultaneously in two linguistic codes and accumulate rich etymological, historical, and dialectological information\. This characteristic transforms Tatarstan toponymy into an exceptionally valuable yet difficult\-to\-process information resource\. Traditional geospatial resources \(GeoNames, OpenStreetMap\) provide basic coordinate and classification data; however, they lack deep linguistic and etymological details and are not designed for semantic search that accommodates synonymy of geographical terms \(“selo” – “derevnya” \[village\], “reka” – “pritok” \[river – tributary\]\) and multilingual spelling variants of the same object\(Rehurek and Sojka,[2006](https://arxiv.org/html/2605.05962#bib.bib3)\)\.

In parallel, the past decade has witnessed notable progress in the computational processing of the Tatar language: representative text corpora have been created\([24](https://arxiv.org/html/2605.05962#bib.bib4);[28](https://arxiv.org/html/2605.05962#bib.bib5);[16](https://arxiv.org/html/2605.05962#bib.bib6)\), morphological analysis systems\(Gilmullin and Gataullin,[2017](https://arxiv.org/html/2605.05962#bib.bib7)\)and tokenization tools\(Arabov,[2026b](https://arxiv.org/html/2605.05962#bib.bib8)\)have been developed, and methods for semantic annotation\(Mukhamedshinet al\.,[2025](https://arxiv.org/html/2605.05962#bib.bib9); Mindubaev and Gatiatullin,[2024](https://arxiv.org/html/2605.05962#bib.bib10)\)and the construction of distributional word and document representations\(Arabov,[2026a](https://arxiv.org/html/2605.05962#bib.bib11); Gafarovet al\.,[2025](https://arxiv.org/html/2605.05962#bib.bib12)\)have been proposed\. Nevertheless, specialized question\-answering datasets that account for toponymic specificity have been absent until now\. Existing QA collections \(SQuAD, SberQuAD, and their Russian\-language counterparts\) are oriented toward news and encyclopedic texts and do not cover structured geographic information that includes coordinates, object types, and etymological descriptions\. At the same time, queries such as “Where is the Mesha River located?”, “What are the coordinates of the village of Rantamak?”, and “Why is Lake Kaban so named?” are natural for users of geoservices and educational platforms\.

Modern generative models underlying Retrieval\-Augmented Generation \(RAG\) require reliable retrieval components capable of supplying relevant context\. In the domain of geographic search, the combination of dense semantic embeddings obtained from multilingual transformers \(multilingual E5\(Lianget al\.,[2022](https://arxiv.org/html/2605.05962#bib.bib13)\), XLM\-RoBERTa\(Conneauet al\.,[2020](https://arxiv.org/html/2605.05962#bib.bib14)\)\) with geospatial ranking and filtering is of particular interest\. However, the integration of such approaches for multilingual toponymic data, especially coupled with subsequent extractive reading, remains underexplored\. The identified gap—the simultaneous absence of both a dataset and an end\-to\-end question\-answering architecture for geographic queries in a bilingual setting—determines the relevance of the present work\.

## 2Related Work

Toponymic Data and Resources for the Tatar Language\.International projects such as GeoNames and OpenStreetMap provide open coordinate and attribute information; however, they do not contain the detailed etymological and linguistic data necessary for scholarly onomastics\([14](https://arxiv.org/html/2605.05962#bib.bib15)\)\. Academic works on the toponymy of Tatarstan \(dictionaries by Garipova, Sattarov, Äkhmätÿanov\) have been partially digitized on the “Toponyms of Tatarstan” portal but lack a machine\-readable application programming interface\. Within the framework of the present study, the authors collected and openly published a structured dataset comprising 9,688 records with names in Russian and Tatar, object types, etymology, and coordinate referencing for 93% of entries\([TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib16)\); it is this dataset that serves as the empirical foundation of the present work\.

Tools for Automatic Processing of the Tatar Language\.For the Tatar language, general\-purpose corpora\([24](https://arxiv.org/html/2605.05962#bib.bib4);[28](https://arxiv.org/html/2605.05962#bib.bib5);[16](https://arxiv.org/html/2605.05962#bib.bib6)\), morphological analysis systems\(Gilmullin and Gataullin,[2017](https://arxiv.org/html/2605.05962#bib.bib7)\), tokenizers\(Arabov,[2026b](https://arxiv.org/html/2605.05962#bib.bib8)\), as well as methods for semantic annotation based on knowledge graphs\(Mukhamedshinet al\.,[2025](https://arxiv.org/html/2605.05962#bib.bib9)\)and machine learning\(Mindubaev and Gatiatullin,[2024](https://arxiv.org/html/2605.05962#bib.bib10)\), have been developed\. Software solutions for constructing vector representations of words and documents have been created\(Arabov,[2026a](https://arxiv.org/html/2605.05962#bib.bib11); Gafarovet al\.,[2025](https://arxiv.org/html/2605.05962#bib.bib12)\)\. All of the listed resources are consolidated on the Hugging Face platform within the TatarNLPWorld organization, ensuring reproducibility of experiments\([TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib16)\)\.

Semantic and Geospatial Search\.Classical lexical methods such as BM25\(Rehurek and Sojka,[2006](https://arxiv.org/html/2605.05962#bib.bib3)\)struggle with cross\-lingual and synonymic queries\. The development of Transformer architectures has led to the emergence of dense retrieval models\. The E5 family\(Lianget al\.,[2022](https://arxiv.org/html/2605.05962#bib.bib13)\)is trained with a contrastive loss function on multilingual collections and is capable of encoding documents and queries into a unified semantic space\. The multilingual\-e5\-large model demonstrates high effectiveness for low\-resource and multilingual scenarios\(Conneauet al\.,[2020](https://arxiv.org/html/2605.05962#bib.bib14); Schweter and Tekir,[2020](https://arxiv.org/html/2605.05962#bib.bib17)\)\. In parallel, geoinformatics employs spatial indices \(KD\-trees\(Bentley,[1975](https://arxiv.org/html/2605.05962#bib.bib18)\), R\-trees\) and metrics that account for the Earth’s curvature \(haversine distance\(Sinnott,[1984](https://arxiv.org/html/2605.05962#bib.bib19)\)\)\. Works\(Galimovet al\.,[2023](https://arxiv.org/html/2605.05962#bib.bib2); Burnashevet al\.,[2025](https://arxiv.org/html/2605.05962#bib.bib20)\)proposed geolinguistic systems combining fuzzy logic with spatial data for dialect analysis; however, they did not utilize dense embeddings\. Hybrid approaches that combine textual and geographic relevance in a weighted manner have been successfully applied in recommendation services and point\-of\-interest search, but such solutions have not been previously investigated for multilingual toponyms with detailed etymological attribution\.

Question\-Answering Systems and Extractive Reading\.For the extractive QA task, SQuAD\(Rajpurkaret al\.,[2016](https://arxiv.org/html/2605.05962#bib.bib21)\)and its multilingual versions have become the de facto standard datasets\. Russian\-language counterparts, such as SberQuAD\(Efimov and others,[2020](https://arxiv.org/html/2605.05962#bib.bib22)\), are oriented primarily toward news and encyclopedic texts and do not contain structured geographic data with coordinates\. Broader diagnostic benchmarks for the Russian language, in particular Russian SuperGLUE\(Fenogenovaet al\.,[2021](https://arxiv.org/html/2605.05962#bib.bib28)\), enable the evaluation of multiple aspects of language understanding but also do not include tasks requiring the extraction of numerical coordinates and etymological information\. Among models, BERT\-based architectures lead the field: RuBERT\(Devlinet al\.,[2019](https://arxiv.org/html/2605.05962#bib.bib23)\)\(further pre\-trained on Russian texts\) and the multilingual XLM\-RoBERTa\(Conneauet al\.,[2020](https://arxiv.org/html/2605.05962#bib.bib14)\), trained on one hundred languages\. The latter has demonstrated excellent results in recognizing nested named entities, including geographical ones, as confirmed, for example, within the Russian\-language competition RuNNE\-2022\(Artemovaet al\.,[2022](https://arxiv.org/html/2605.05962#bib.bib29)\)\. Generative T5 models\(Raffelet al\.,[2020](https://arxiv.org/html/2605.05962#bib.bib24)\)are also applied, but typically in the answer generation setting\. A significant limitation of monolingual models, revealed in recent experiments\(Choureet al\.,[2022](https://arxiv.org/html/2605.05962#bib.bib25)\), is their inability to process numerical coordinates due to WordPiece tokenization specifics and a shortage of relevant examples in training sets\. Multilingual models employing SentencePiece, by contrast, successfully extract numerical sequences\. To date, no specialized QA benchmarks combining structured geographic information with etymological and coordinate components have been proposed\.

Thus, the literature review attests to the existence of disparate components \(language resources for Tatar, dense retrieval models, geospatial filtering methods, extractive reading models\); however, their integration into a unified system capable of finding the required toponym and extracting the precise answer for an arbitrary geographic query in Russian or Tatar is lacking\. The present work is aimed at filling this gap\.

## 3Methodology

### 3\.1Hybrid Retrieval: Architecture and Method

The purpose of the retrieval component is to take a textual queryqq, optionally accompanied by a geographic point\(latq,lonq\)\(\\textit\{lat\}\_\{q\},\\textit\{lon\}\_\{q\}\)and a radiusRR, and produce a ranked list of thekkmost relevant toponyms from a set of documentsDD, each of which is equipped with coordinates\. The architecture comprises two parallel indexing stages—semantic and spatial—which interact during query processing in a hybrid ranking procedure\.

Semantic Indexing\.To represent the content portion of a document, the multilingual encoder modelintfloat/multilingual\-e5\-large\(Lianget al\.,[2022](https://arxiv.org/html/2605.05962#bib.bib13)\)is employed, which transforms the contextual fieldcontextinto a 1024\-dimensional dense vector\. The model was chosen as it demonstrates state\-of\-the\-art results in cross\-lingual search and supports both Russian and Tatar languages\. All document embeddings are normalized to unitL2L\_\{2\}\-norm, allowing cosine proximity to be assessed via the dot product\. Computation is performed in batches of 32 documents using a GPU\. For efficient nearest\-neighbor search, a flat inner product index \(IndexFlatIP\) is constructed using the FAISS library\(Johnsonet al\.,[2019](https://arxiv.org/html/2605.05962#bib.bib26)\); when a GPU accelerator is available, the index is placed in GPU memory\. Query processing reduces to encoding its text with the same model,L2L\_\{2\}\-normalizing the query vector, and extracting the top\-kkdocuments with the maximum dot product\. The resulting quantity

sem\_score​\(i\)=⟨eq,ei⟩\\textit\{sem\\\_score\}\(i\)=\\langle e\_\{q\},e\_\{i\}\\rangleis interpreted as the semantic relevance of documentii\.

Geospatial Filter and Ranking\.Spatial selection is implemented in two passes\. In the first stage, a bounding box is constructed around the query point:

Δ​lat=R/111320,Δ​lon=R/\(111320⋅cos⁡\(ϕ\)\),\\Delta\\textit\{lat\}=R/111320,\\quad\\Delta\\textit\{lon\}=R/\(111320\\cdot\\cos\(\\phi\)\),whereϕ\\phiis the latitude in radians andRRis the search radius\. Objects whose coordinates fall within this region are declared candidates\. Then, for each candidate, the exact haversine distance\(Sinnott,[1984](https://arxiv.org/html/2605.05962#bib.bib19)\)is computed, and those objects whose distance exceedsRRare filtered out\. The spatial score of the remaining candidates is determined by an exponentially decaying function of distance:

geo\_score​\(i\)=exp⁡\(−diR\)\.\\textit\{geo\\\_score\}\(i\)=\\exp\\left\(\-\\frac\{d\_\{i\}\}\{R\}\\right\)\.To accelerate the filtering stage, a KD\-tree\(Bentley,[1975](https://arxiv.org/html/2605.05962#bib.bib18)\)is built over the coordinates of all documents during corpus preparation, enabling logarithmic\-time retrieval of the subset of objects within a given neighborhood\.

Combined Relevance\.The final relevance is formed as a weighted sum of normalized semantic and spatial scores\. Semantic scores are normalized using min\-max scaling among the candidates that passed the geofilter:

sem\_norm​\(i\)=sem\_score​\(i\)−minsemmaxsem−minsem,\\textit\{sem\\\_norm\}\(i\)=\\frac\{\\textit\{sem\\\_score\}\(i\)\-\\min\_\{\\textit\{sem\}\}\}\{\\max\_\{\\textit\{sem\}\}\-\\min\_\{\\textit\{sem\}\}\},and in degenerate cases \(when the difference is close to zero\), all values are set to 1\. Spatial scores are normalized by dividing by the maximum value among the candidates:

geo\_norm​\(i\)=geo\_score​\(i\)maxgeo\.\\textit\{geo\\\_norm\}\(i\)=\\frac\{\\textit\{geo\\\_score\}\(i\)\}\{\\max\_\{\\textit\{geo\}\}\}\.The combined score is computed as

score​\(i\)=α⋅sem\_norm​\(i\)\+\(1−α\)⋅geo\_norm​\(i\),\\textit\{score\}\(i\)=\\alpha\\cdot\\textit\{sem\\\_norm\}\(i\)\+\(1\-\\alpha\)\\cdot\\textit\{geo\\\_norm\}\(i\),whereα∈\[0,1\]\\alpha\\in\[0,1\]is a hyperparameter governing the contribution of the semantic component\. Objects are ranked by descendingscore​\(i\)\\textit\{score\}\(i\), and the top\-kkare returned\. The value ofα\\alphais tuned empirically on a validation set \(200 queries\) via grid search over\{0\.1,0\.3,0\.5,0\.7,0\.9\}\\\{0\.1,0\.3,0\.5,0\.7,0\.9\\\}by maximizing the average Recall@5\. In the experiments,α=0\.1\\alpha=0\.1was found to be optimal, with a fixed radius ofR=50R=50km, chosen based on the typical scale of local search within Tatarstan\.

### 3\.2Extractive Reading: QA Corpus Generation and Model Training

For the extractive reading component, the original toponym dataset is transformed into a set of question\-answer pairs conforming to the SQuAD format\(Rajpurkaret al\.,[2016](https://arxiv.org/html/2605.05962#bib.bib21)\)\. The procedure pursues two objectives: to ensure precise answer localization within the context for training extractive models, and to cover all major information categories of the source records\.

Context and Question Generation\.Each significant field of a source record \(name, object type, location, etymology, coordinates, region, physiographic characteristics, sources\) is prefixed with a Russian\-language label \(Название \(рус\): \[Name \(Rus\):\], Объект: \[Object:\], Этимология: \[Etymology:\], Расположение: \[Location:\], Координаты: \[Coordinates:\], etc\.\) and concatenated with other fields via the “ \| ” separator\. If the resulting string exceeds 2048 characters, proportional truncation of each field is applied while preserving the prefix\. For seven information categories, Russian\-language question templates with a\{name\}placeholder have been developed, into which the Russian \(or, if unavailable, Tatar\) name of the object is inserted\. Examples of templates: Что такое \{name\}? \[What is \{name\}?\] \(object type\), Где находится \{name\}? \[Where is \{name\} located?\] \(location\), Какие координаты у \{name\}? \[What are the coordinates of \{name\}?\] \(coordinates\), Почему \{name\} так называется? \[Why is \{name\} so named?\] \(etymology\), and others\. The full list of templates is provided in Section 4\.2\.

Answer Construction and Annotation\.For each question, the answer is the value of the corresponding field \(for coordinates, the string “latitude, longitude”\)\. The start position of the answer within the context is computed algorithmically: as the sum of the lengths of all fields preceding the target prefix, plus the length of the prefix itself\. This ensures precise character\-level annotation required for training extractive models\. The maximum number of QA pairs per record is limited to ten to ensure uniform coverage\. The resulting corpus comprises 38,696 pairs and is split into training \(90%\) and validation \(10%\) sets, stratified by question type\.

Models and Training\.For the extractive reading task, two groups of transformer architectures were fine\-tuned: monolingual Russian\-language RuBERT models \(base and large versions, pre\-trained on SberQuAD\(Efimov and others,[2020](https://arxiv.org/html/2605.05962#bib.bib22); Devlinet al\.,[2019](https://arxiv.org/html/2605.05962#bib.bib23)\)\) and the multilingual XLM\-RoBERTa\-large model\(Conneauet al\.,[2020](https://arxiv.org/html/2605.05962#bib.bib14)\), pre\-trained on SQuAD 2\.0\. The generative model T5\-RUS\(Raffelet al\.,[2020](https://arxiv.org/html/2605.05962#bib.bib24)\)was not fine\-tuned due to technical reasons and was used in its base variant as an additional reference point\. All models were trained on the generated QA corpus for three epochs using the AdamW optimizer, a learning rate of3×10−53\\times 10^\{\-5\}, a batch size of 4, and linear warm\-up over 500 steps\. For extractive models, the standard procedure of predicting start and end positions of the answer was applied; for T5\-RUS, the format “question: … context: …→\\toanswer” was employed\. As a simple baseline, a heuristic rule was implemented that extracts the answer based on question keywords and prefixes in the context\.

The trained models, as well as the generated QA corpus, are published on Hugging Face\([TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib31)\)and are available for reproducibility of the results\.

## 4Data

### 4\.1Source Dataset of Tatarstan Toponyms

The empirical foundation of the study is a specialized bilingual dataset of Tatarstan toponyms, collected, structured, and published by the authors in open access on the Hugging Face platform\([TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib30)\)as part of the present work\. The necessity of creating this resource is dictated by the absence of machine\-readable datasets that simultaneously contain coordinate, linguistic, and etymological information about the region’s geographical objects\. The dataset aggregates data from fundamental academic works on the toponymy of Tatarstan: the hydronym dictionaries of F\. G\. Garipova, the studies of G\. F\. Sattarov on Tatar toponymy, the etymological dictionaries of R\. G\. хмтьянов, the multi\-volume Tatar Encyclopedia, dialectological dictionaries, as well as materials from the digital portal Топонимы Татарстана \(toponym\.antat\.ru\)\. Reliance on peer\-reviewed academic sources ensures high reliability and completeness of the included information\.

The total volume of the dataset amounts to 9,688 records, each corresponding to a single geographical object within the territory of the Republic of Tatarstan and adjacent regions of compact Tatar settlement\. The record structure includes the following fields: a unique identifier, source URL, toponym type \(toponym or microtoponym\), toponym subtype \(oikonym, hydronym, oronym, or no type\), geographical object \(76 unique categories: деревня \[village\], село \[selo\], река \[river\], озеро \[lake\], поле \[field\], гора \[mountain\], луг \[meadow\], приток реки \[river tributary\], поляна \[glade\], etc\.\), name in Russian, name in Tatar, federal subject, physiographic details, geographical location, name etymology, bibliographic sources, latitude and longitude coordinates \(where available\), and a map availability flag\. Etymological information, which is of particular value for scholarly research in onomastics and historical geography, is present in approximately 63% of records\.

A key characteristic of the dataset is the presence of coordinate referencing: 9,023 records \(93\.1%\) are furnished with latitude and longitude values, enabling their use in geospatial analysis and distance\-based filtering procedures\. The remaining 665 records \(6\.9%\) lacking coordinates are excluded from the hybrid search experiments but were employed in generating the question\-answering corpus for categories not requiring spatial information \(etymology, sources\)\.

The distribution of records by toponym subtype is presented in Table[1](https://arxiv.org/html/2605.05962#S4.T1)\. The predominant share is constituted by oikonyms—names of populated places \(деревни \[villages\], сёла \[selos\], посёлки \[settlements\]\), which reflects the settlement structure in the region and the thematic coverage of the sources used\. Hydronyms \(names of water bodies: rivers, lakes, streams\) and oronyms \(names of landforms: mountains, hills, ravines\) together form approximately one quarter of the corpus\. The “no type” category unites records for which the toponym subtype was not explicitly indicated in the sources; however, they retain information about the geographical object and etymology\.

Table 1:Distribution of dataset records by toponym subtypeGranularity by geographical object type \(Table[2](https://arxiv.org/html/2605.05962#S4.T2)\) reveals the most frequent categories\. Populated places dominate: деревни \[villages\] \(36\.1%\) and сёла \[selos\] \(20\.6%\), which corresponds to the share of oikonyms in Table[1](https://arxiv.org/html/2605.05962#S4.T1)\. Among natural objects, the most represented are луга \[meadows\] \(8\.3%\), реки \[rivers\] \(7\.2%\), поля \[fields\] \(5\.2%\), and горы \[mountains\] \(4\.1%\)\. About 10\.7% of records belong to other, less frequent categories, such as lakes, river tributaries, glades, springs, natural tracts, swamps, ravines, and other objects\. The presence of 76 unique types of geographical objects ensures high entity diversity and allows testing search algorithms under conditions of heterogeneous taxonomy\.

Table 2:Top\-10 most frequent geographical objects in the datasetThe regional distribution lawfully reflects the geographic focus of the dataset: approximately 93% of records pertain to objects within the territory of the Republic of Tatarstan, about 4% to Tyumen Oblast \(areas of compact settlement of Siberian Tatars\), and the remaining 3% are distributed among adjacent subjects of the Russian Federation \(Башкортостан \[Bashkortostan\], Ulyanovsk, Samara, Orenburg Oblasts, and others\)\.

To facilitate semantic search, all textual fields of each record, excluding coordinates, were merged into a single contextual fieldcontextusing English\-language key prefixes\. The context format is as follows:

> Name \(rus\): <name\_rus\> \| Name \(tat\): <name\_tat\> \| Type: <toponym\_type\> \| Subtype: <toponym\_subtype\> \| Object: <geographical\_object\> \| Etymology: <etymology\> \| Details: <physiographic\_details\> \| Location: <geographical\_location\> \| Sources: <bibliographic\_sources\>

Fields containing no information were excluded from the context\. The choice of English\-language prefixes is dictated by the use of the multilingual modelmultilingual\-e5\-large, for which English\-language keys provide more robust matching of semantically similar fields in cross\-lingual space\. The described procedure enabled the formation of 9,023 documents with non\-empty contextual fields and coordinate referencing, which constituted the corpus for indexing and testing the retrieval component\.

### 4\.2Generation of the Question\-Answering Corpus

Based on the structured toponym dataset, automatic generation of a synthetic question\-answering corpus was performed, intended for training and evaluating extractive QA models\. The key requirement for the generation procedure was guaranteed answer localization within the context text, which is necessary for the correct computation of answer start and end positions during extractive model training\.

For each information type in a source record \(name, object type, location, etymology, coordinates, region, physiographic characteristics, sources\), the corresponding field was transformed into a string with a Russian\-language prefix unambiguously identifying the category: Название \(рус\): \[Name \(Rus\):\], Название \(тат\): \[Name \(Tat\):\], Объект: \[Object:\], Этимология: \[Etymology:\], Расположение: \[Location:\], Координаты: \[Coordinates:\], Регион: \[Region:\], Физико\-географические сведения: \[Physiographic Details:\], Источники: \[Sources:\]\. All prefixed strings were concatenated via the “ \| ” separator into a single context\. When the maximum permissible length \(2048 characters\) was exceeded, proportional truncation of each field was applied, with mandatory preservation of the prefix\. This approach guarantees the presence of any potential answer within the context together with the prefix marking it, which subsequently enables unambiguous determination of the answer start position as the sum of the lengths of preceding fields and separators\.

For seven information categories, Russian\-language question templates were developed\. Each template contains a\{name\}placeholder, into which the object name is inserted\. Examples of templates by category:

- •Object type: Что такое \{name\}? \[What is \{name\}?\], Какой тип у \{name\}? \[What type is \{name\}?\], К какому типу относится \{name\}? \[To which type does \{name\} belong?\];
- •Location: Где находится \{name\}? \[Where is \{name\} located?\], В каком месте расположен \{name\}? \[In what place is \{name\} situated?\], Где именно расположен \{name\}? \[Where exactly is \{name\} situated?\];
- •Etymology: Что означает название \{name\}? \[What does the name \{name\} mean?\], Почему \{name\} так называется? \[Why is \{name\} so named?\], Каково происхождение названия \{name\}? \[What is the origin of the name \{name\}?\];
- •Coordinates: Какие координаты у \{name\}? \[What are the coordinates of \{name\}?\], Где на карте находится \{name\}? \[Where on the map is \{name\}?\];
- •Region: В каком регионе находится \{name\}? \[In which region is \{name\} located?\], Какой федеральный субъект у \{name\}? \[What federal subject does \{name\} belong to?\];
- •Sources: Какие источники упоминают \{name\}? \[Which sources mention \{name\}?\], Где можно прочитать о \{name\}? \[Where can one read about \{name\}?\];
- •Physiographic characteristics: Какие физико\-географические сведения о \{name\}? \[What physiographic details are there about \{name\}?\], Что известно о географических особенностях \{name\}? \[What is known about the geographical features of \{name\}?\]\.

During generation, for each record, one of the templates was randomly selected, into which the Russian name of the object \(or, if absent, the Tatar name\) was inserted\. The answer was formed as the value of the corresponding field \(for coordinates, a string with latitude and longitude separated by a comma\)\. The answer start position within the context was computed algorithmically as the total length of all fields preceding the target prefix, plus the length of the prefix itself\. Thus, each QA pair contains precise positional labels conforming to the SQuAD format\(Rajpurkaret al\.,[2016](https://arxiv.org/html/2605.05962#bib.bib21)\)\.

The maximum number of question\-answer pairs per record was limited to ten to prevent the dominance of objects with a large number of populated fields and to ensure uniform coverage\. The total volume of the generated corpus amounted to 38,696 QA pairs\. The distribution by question type is presented in Table[3](https://arxiv.org/html/2605.05962#S4.T3)\.

Table 3:Distribution of the synthetic QA corpus by question typeThe largest share \(31\.9%\) is comprised of coordinate questions, which reflects the primary importance of spatial information for geographical objects\. Questions about object type \(20\.8%\) and location \(14\.8%\) together form more than one third of the corpus\. Etymological questions \(13\.0%\) exploit a unique feature of the dataset—the availability of data on name origins\. The smallest share \(1\.6%\) is held by questions about physiographic characteristics, which is attributable to the fragmentary population of this field in the source records\.

The average context length in the QA corpus is 1250 characters; answers vary from 2 to 150 characters\. Coordinate answers are the shortest \(on average 20 characters\), while the longest are answers containing a list of bibliographic sources \(up to 500 characters in individual cases\)\. An example of a single question\-answer pair in SQuAD format is provided below:

> \{ "id":"1530\_coordinates\_0", "context":"Название \(рус\): Рантамак \| Название \(тат\): Рантамак \| Объект: Село \| Этимология: Топоним произошел от ойконима ‶Рангазар\-Тамак˘\. \| Расположение: Расположено на р\. Мелля, в 21 км к востоку от с\. Сарманово\. \| Источники: хмтьянов Р\.Г\. Татар телене этимологик сзлеге\.\.\. \| Координаты: 55\.205461, 52\.881862", "question":"Какие координаты у Рантамак?", "answers": \[\{"text":"55\.205461, 52\.881862","answer\_start":312\}\] \}

The generated corpus was randomly partitioned into training \(90%, 34,826 pairs\) and validation \(10%, 3870 pairs\) sets while preserving stratification by question type\. Both parts of the corpus, as well as the source toponym dataset, are openly published on the Hugging Face platform\([TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib30);[TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib31)\)under the CC BY\-SA 4\.0 license and are available for use in scholarly and applied purposes\.

## 5Experimental Study

The experimental validation of the proposed question\-answering system was conducted in two stages, corresponding to its architectural components\. In the first stage, the effectiveness of the hybrid retrieval module, combining semantic indexing with geospatial filtering, was evaluated; in the second, the ability of the extractive reading component to accurately localize the answer within the context provided by the retrieval mechanism was assessed\. Such two\-part evaluation protocol made it possible, on the one hand, to characterize each subsystem in isolation, and on the other, to demonstrate their compatibility within a unified pipeline\. All experiments were carried out on the corpus of 9,023 coordinate\-referenced documents described above and on the synthetic question\-answering set of 38,696 examples generated according to the procedure in Section 4\.2\.

To evaluate the retrieval component, an independent test set was prepared by randomly selecting 500 records from the source dataset and generating natural\-language queries for them using five templates \(Что такое \{name\}? \[What is \{name\}?\], Где находится \{name\}? \[Where is \{name\}?\], Расскажи о \{name\} \[Tell about \{name\}\], etc\.\), in which the object name was inserted either in Russian or in Tatar with probabilities of 0\.7 and 0\.3, respectively\. This approach guaranteed that for each query there exists exactly one relevant document—the source record from which the query was generated\. Additionally, a validation set of 200 queries, disjoint from the test set, was allocated; it was used exclusively for tuning the hybrid search hyperparameters—the weighting coefficientα\\alphaand the spatial filter radiusRR\.

Retrieval quality was measured by two main metrics: Recall@kk\(k=1,3,5k=1,3,5\) and Mean Reciprocal Rank \(MRR\)\. The former indicates the proportion of queries for which the relevant document appears among the topkkretrieval results; the latter averages the reciprocal of the rank of the first relevant document and is thus sensitive to early hits\. To obtain statistically sound conclusions, all metrics were accompanied by 95% confidence intervals constructed via bootstrap with 1,000 resamples with replacement\. This enabled not only comparison of average values but also assessment of the significance of differences between methods\.

Four ranking strategies were compared: classical lexical BM25 search, purely spatial search \(top\-kknearest by haversine distance\), semantic search based onmultilingual\-e5\-largewith a FAISS index, and the proposed hybrid method\. The hybrid search was performed with a fixed radius ofR=50R=50km andα=0\.1\\alpha=0\.1, selected on the validation set as yielding the maximumRecall@5=1\.0\\text\{Recall@5\}=1\.0\. The comparison results are summarized in Table[4](https://arxiv.org/html/2605.05962#S5.T4), and Figure[1](https://arxiv.org/html/2605.05962#S5.F1)visualizes them as bar charts with interval estimates\.

Table 4:Comparison of search methods \(95% bootstrap confidence intervals\)The figures in Table[4](https://arxiv.org/html/2605.05962#S5.T4)indicate a substantial superiority of the hybrid scheme\. First of all, the valueRecall@5=1\.000\\text\{Recall@5\}=1\.000is striking, meaning that for each of the 500 test queries, the relevant object is guaranteed to be present among the first five results\. Such high recall practically eliminates the risk of losing the necessary information before the reading stage, which is a critical requirement for RAG systems\. TheRecall@1=0\.988\\text\{Recall@1\}=0\.988score indicates that only in six cases out of five hundred did the target document fail to occupy the first position; moreover, the lower bound of the confidence interval \(0\.978\) lies substantially above the upper bounds ofRecall@1for all competing methods\. This arrangement of intervals confirms the statistical significance of the hybrid’s superiority over the alternatives\.

Semantic search by itself demonstrates fairly high quality \(Recall@1=0\.774\\text\{Recall@1\}=0\.774\), confirming the ability of themultilingual\-e5\-largemodel to adequately encode multilingual toponymic contexts and capture semantic proximity between different name variants\. However, the absence of spatial filtering allows objects with similar names but located far apart to appear in top positions, which reduces accuracy by approximately 0\.2 compared to the hybrid\. Spatial search, which ignores the query text, showsRecall@1=0\.536\\text\{Recall@1\}=0\.536\. This is expected: within a 50 km radius, several objects are often found, and the geographically nearest one does not always coincide with the sought one\. Nevertheless, even this simple strategy outperforms BM25 \(Recall@1=0\.438\\text\{Recall@1\}=0\.438\), underscoring the fundamental role of geographic proximity for toponymic queries\. BM25’s most modest results are explained not only by the lexical gap between Tatar and Russian names but also by the synonymy of geographical terms: a classical inverted index is unable to equate село \[selo\] and деревня \[village\], река \[river\] and приток \[tributary\]\.

A detailed picture is presented in Figure[1](https://arxiv.org/html/2605.05962#S5.F1), where the Recall@1 and Recall@5 bars for each method are supplemented with confidence intervals\. The hybrid method stands out not only for the height of its bars but also for their minimal variability, indicating high stability of the algorithm\.

![Refer to caption](https://arxiv.org/html/2605.05962v1/ris1.png)Figure 1:Comparison of Recall@1 and Recall@5 with 95% confidence intervals for BM25, purely spatial, semantic, and hybrid methods\.As a next step, we analyzed the performance of the hybrid approach broken down by toponym type\. The source dataset is divided into two main categories: Топоним \[Toponym\] \(large objects: populated places, significant natural features, etc\.\) and Микротопоним \[Microtoponym\] \(local landscape elements: meadows, fields, natural tracts, springs\)\. In the test set, toponyms accounted for 349 queries and microtoponyms for 151\. The results of the separate analysis are presented in Table[5](https://arxiv.org/html/2605.05962#S5.T5)\.

Table 5:Recall@1 of the hybrid method by toponym typeBoth categories demonstrate near\-ceiling values; however, microtoponyms show a slightly higher result\. This difference is explained by their local nature: for small objects, the density of toponyms within a 50 km radius is generally lower, meaning the spatial filter leaves fewer candidates, and the semantic component finds it easier to identify the sole correct one\. For comparison, Figure[2](https://arxiv.org/html/2605.05962#S5.F2)presents a bar chart that clearly illustrates this distinction\.

![Refer to caption](https://arxiv.org/html/2605.05962v1/ris2.png)Figure 2:Recall@1 of the hybrid method for the “Toponym” and “Microtoponym” categories\.Despite the impressive aggregate metrics, the hybrid retrieval made six errors in which the relevant document did not appear in the first position\. Manual analysis of these cases, undertaken to identify root causes, showed that all of them are related not to shortcomings of the combination algorithm but to the quality of the data themselves\. In two cases, complete namesakes were present within a 50 km radius—two different objects with identical names \(e\.g\., two natural tracts named Красная Горка \[Krasnaya Gorka\]\)\. Their contexts proved practically identical, and the semantic component could not give preference to one over the other\. In three cases, the coordinates of the target object were offset by 10–15 kilometers from the true position, due to which it did not fall within the search zone and was eliminated at the geofiltering stage\. In one further case, thecontextfield contained a minimum of information \(only name and type\), which yielded a predictably low semantic score and allowed another, more informative candidate to take the lead\. Consequently, further improvement of the retrieval component lies primarily in enhancing data quality: verifying coordinates through cross\-validation with OpenStreetMap or satellite imagery, enriching contextual fields, and developing a component for resolving toponymic homonymy\.

Turning to the second stage, extractive reading, we note that the task here was to precisely locate the answer within the context retrieved by the retrieval module\. Three architectures, representing both monolingual and multilingual approaches, were selected for training and evaluation: RuBERT base, RuBERT large, and XLM\-RoBERTa large, each fine\-tuned on the synthetic QA corpus of 34,826 pairs\. The generative model T5\-RUS, unfortunately, could not be trained due to tokenizer version incompatibility and was used only in its base pre\-trained variant as an additional reference point\. Additionally, as a simple baseline, a heuristic rule was implemented that extracts the answer based on question keywords and the corresponding prefixes in the context\. The training hyperparameters, common to all models, are presented in Table[6](https://arxiv.org/html/2605.05962#S5.T6)\.

Table 6:Hyperparameters for QA model fine\-tuningThe results obtained on the validation set of 3,870 QA pairs uncovered an important nuance that substantially affected interpretation\. It turned out that the RuBERT models, operating with the WordPiece tokenizer, tend to insert extraneous spaces inside numerical coordinates \(“55\. 175195” instead of “55\.175195”\) and hyphenated constructions \(“северо \- западу” \[“north \- west”\]\)\. This is a purely tokenization artifact unrelated to the semantic capability of the model\. By applying elementary post\-processing \(removal of breaks within floating\-point numbers and unification of hyphens and brackets\), we ensured that RuBERT answers became identical to the reference ones\. Table[7](https://arxiv.org/html/2605.05962#S5.T7)presents both raw Exact Match and F1 metrics and the values obtained after normalization\.

Table 7:QA model results on the validation set \(raw and normalized metrics\)The multilingual XLM\-RoBERTa large initially delivers virtually perfect answers \(EM=0\.992\\text\{EM\}=0\.992\) and requires no post\-processing\. Both RuBERT models, after minimal normalization, achieve exact match with the reference \(EM=1\.000\\text\{EM\}=1\.000\), while operating 3\.5 times faster—6\.5 ms versus 22\.4 ms per query\. This speed advantage makes RuBERT the preferred choice for production systems, provided a thin post\-processing layer is added\. Figure[3](https://arxiv.org/html/2605.05962#S5.F3)presents a comparison of models in terms of EM and F1 with 95% bootstrap confidence intervals\.

![Refer to caption](https://arxiv.org/html/2605.05962v1/ris3.png)Figure 3:Comparison of models by Exact Match and F1 score after normalization \(error bars represent 95% confidence intervals\)\.For a deeper understanding of capabilities and limitations, we conducted a per\-category quality analysis, presented in Table[8](https://arxiv.org/html/2605.05962#S5.T8)\. After normalization, all three fine\-tuned models achieve 100% F1 on questions about etymology, location, region, and object type\. The only noticeable deviation is observed for XLM\-RoBERTa large on coordinates \(F1=0\.984\\text\{F1\}=0\.984\), which is associated with rare cases of inaccurate copying of the last digit of a decimal fraction\.

Table 8:F1 score by question type after normalizationOn the heatmap \(Figure[4](https://arxiv.org/html/2605.05962#S5.F4)\), this result is vividly reflected: all cells, with the exception of one, are colored in the warmest tones, corresponding to an F1 of unity\.

![Refer to caption](https://arxiv.org/html/2605.05962v1/ris4.png)Figure 4:Heatmap of F1 score by question type for fine\-tuned models \(after normalization\)\. Black cells are absent, demonstrating the complete resolution of the coordinate problem for RuBERT\.The experimental study is concluded by the analysis of answer length and inference time\. The distribution of predicted answer lengths for the XLM\-RoBERTa large model is practically identical to the reference \(Figure[5](https://arxiv.org/html/2605.05962#S5.F5)\)\. The bulk of answers is concentrated in the 20–40 character range, corresponding to coordinates and short names; there is also a small peak around 200 characters, corresponding to bibliographic sources\. For RuBERT after normalization, the distribution shape coincides with the reference without significant deviations\.

![Refer to caption](https://arxiv.org/html/2605.05962v1/ris5.png)Figure 5:Distribution of answer length \(in characters\) for reference values and predictions of the xlm\_roberta\_large model\.Speed measurements showed \(Figure[6](https://arxiv.org/html/2605.05962#S5.F6)\) that RuBERT spends only about 6\.5 ms per query, XLM\-RoBERTa large spends 22\.4 ms, and the generative T5 spends 27\.2 ms\. Thus, for tasks where latency is critical, RuBERT with post\-processing appears to be the optimal alternative, whereas XLM\-RoBERTa large may be preferable when maximum quality of numerical coordinate extraction is required without additional programming\.

![Refer to caption](https://arxiv.org/html/2605.05962v1/ris6.png)Figure 6:Average processing time per example \(ms\) for the considered QA models\.Thus, the totality of the experimental data convincingly demonstrates that the proposed two\-component architecture achieves near\-ideal performance at all stages—from retrieval of the relevant document to precise extraction of the answer—and the identified tokenization peculiarities of RuBERT not only explain previous observations but also provide a simple recipe for their complete elimination\.

## 6Discussion

The conducted experimental study covered two key components of the proposed question\-answering architecture—hybrid retrieval and extractive reading—and yielded a series of interrelated results that are expedient to discuss in aggregate\. The most important outcome of the retrieval part is the demonstration that the combination of semantic search based on dense embeddings frommultilingual\-e5\-largeand geospatial filtering is capable of delivering virtually perfect recall for toponymic queries at the scale of a single region\. The valuesRecall@5=1\.000\\text\{Recall@5\}=1\.000andMRR=0\.994\\text\{MRR\}=0\.994, achieved on an independent test set of 500 queries, not only significantly surpass traditional BM25 and purely spatial search but also leave minimal room for further improvement\. The lower bound of the hybridRecall@1confidence interval \(0\.978\) lies above the upper bounds of the analogous intervals for all alternatives, confirming the statistical significance of the superiority and attesting to the stability of the method\.

Analysis of the optimal value of the weighting coefficientα=0\.1\\alpha=0\.1, obtained via grid search on the validation set, shows that in the task of toponym search, geographic proximity is the dominant relevance feature\. The spatial filter with a 50 km radius reduces the candidate set from several thousand to several dozen, after which even a small contribution from the semantic component suffices for accurate identification of the target object\. This conclusion accords well with intuition but had not previously been confirmed quantitatively on multilingual toponymic material enriched with etymological data\. It is noteworthy that microtoponyms demonstrate a marginally higherRecall@1compared to larger toponyms \(0\.993 vs\. 0\.986\)\. The reason, apparently, lies in their local character: within a 50 km radius around small natural tracts or springs, fewer candidate objects are found, which further eases the task for the semantic component\.

Manual analysis of the six retrieval errors in which the relevant document did not occupy the first position revealed that all of them are attributable to the quality of the source data rather than algorithmic shortcomings\. Three out of six cases are associated with coordinate inaccuracy—objects were offset by 10–15 km from their true position and fell outside the search radius\. Two further cases involved complete namesakes within the same radius, and one involved insufficient contextual description\. This observation has a direct practical implication: further improvement of retrieval quality requires, first and foremost, not model complication but coordinate verification \(e\.g\., through cross\-validation with OpenStreetMap or satellite imagery data\), enrichment of contextual fields, and the development of a module for resolving toponymic homonymy\.

The results obtained for the extractive reading component are no less indicative, and in some aspects, unexpected\. First of all, the critical importance of domain\-specific fine\-tuning was confirmed: the base versions of both RuBERT and XLM\-RoBERTa large, without adaptation to the synthetic QA corpus, demonstrate unacceptably low quality \(F1does not exceed 0\.574\), whereas after three epochs of fine\-tuning, the metrics reach near\-ceiling values\. This contrast underscores the value of the created question\-answering resource, which contains patterns absent from general QA collections such as SQuAD and SberQuAD\.

A noteworthy result was the elucidation of the nature of RuBERT’s “failure” on geographic coordinate questions\. The initially recorded zeroF1on this category was misleading and could have created the impression of a fundamental inability of monolingual WordPiece models to process numerical information\. Detailed analysis showed that the problem is of a purely tokenization nature: RuBERT splits coordinates into tokens with spaces \(“55\. 175195”\), which hinders accurate extraction\. Elementary post\-processing—removal of extraneous spaces within floating\-point numbers and unification of hyphens—completely eliminates the artifact, after which both versions of RuBERT achieve 100% accuracy on all question categories, including coordinates\. Moreover, RuBERT’s inference speed \(approximately 6\.5 ms per query\) is 3\.5 times that of XLM\-RoBERTa large \(22\.4 ms\)\. Thus, the widespread notion that multilingual SentencePiece models are necessary for working with numerical data requires refinement: when light post\-processing is permissible, monolingual models may be preferable due to gains in speed and lower memory consumption\.

The synergy of the two components—retrieval and reader—merits a separate comment\. The achieved metrics ofRecall@5=1\.000\\text\{Recall@5\}=1\.000at the search stage andEM≈1\.000\\text\{EM\}\\approx 1\.000at the reading stage \(with post\-processing\) mean that, within the model setting where the query is generated from one of the dataset records, the system is capable of errorlessly finding the relevant document and extracting the exact answer\. In the context of RAG systems, this property is critically important: the generative model receives a correct context with a probability tending to unity, which radically reduces the risk of hallucinations\. Although an end\-to\-end retrieval\-plus\-reader experiment was not conducted in this work, the separate metrics provide an upper\-bound estimate of the full pipeline’s quality and allow the developed components to be regarded as ready for integration\.

It is necessary to note a number of limitations inherent to the present study\. First, the test queries for both components were generated from templates and do not reflect the full diversity of formulations characteristic of real users\. This is standard practice for the initial stages of creating specialized datasets, but it leaves open the question of the system’s behavior on live queries\. Second, the geospatial filtering radius is fixed at 50 km; for objects of different scales \(a spring versus a mountain range\), such a value may be suboptimal\. Third, the generative model T5\-RUS could not be trained due to technical reasons, so the comparison of extraction and generation paradigms remained incomplete\. In addition, questions about physiographic characteristics constituted only 1\.6% of the QA corpus, which is insufficient for reliable conclusions, and this category deserves further supplementation\.

The practical significance of the performed work is determined not only by the achieved metrics but also by the complete cycle of open publication of artifacts\. The source toponym dataset\([TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib30)\), the question\-answering corpus\([TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib31)\), the weights of the trained models \(including both versions of RuBERT and XLM\-RoBERTa large\), and the interactive web demo\([TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib32)\)are hosted on Hugging Face and are ready for use in geoinformation services, educational platforms, and digital humanities projects\. The openness of the resources ensures reproducibility and enables other researchers to advance the direction, in particular, to adapt the proposed approach to other regions and language pairs\.

Promising directions for further research flow directly from the discussed limitations\. These include: the development of adaptive strategies for selecting the geofiltering radius depending on the object type or local toponym density; cross\-validation of coordinates against external sources to minimize positioning errors; conducting an end\-to\-end retrieval\-plus\-reader experiment in a full\-fledged RAG setting with measurement of final answer quality; fine\-tuning generative models \(such as T5\) on the created QA corpus for direct comparison with the extractive approach; as well as the application of parameter\-efficient fine\-tuning methods \(LoRA/QLoRA\), which have already demonstrated promise for the closely related Bashkir language\(Arabov and Khaybullina,[2026](https://arxiv.org/html/2605.05962#bib.bib27)\)\. Addressing the enumerated tasks will enable the transition from a demonstration prototype to a production\-grade question\-answering system for multilingual geographic data\.

## 7Conclusion

In this work, a comprehensive question\-answering system for multilingual toponymic data has been presented and experimentally substantiated, encompassing the stages of resource collection and structuring, hybrid retrieval of relevant documents, and extractive reading\. The conducted research has yielded the following main scientific and practical results\.

1. 1\.An original bilingual dataset of Tatarstan toponyms has been collected, verified, and openly published, comprising 9,688 records, of which 9,023 are furnished with geographic coordinates and contain extensive linguistic, etymological, and administrative information\. Based on this dataset, a specialized question\-answering corpus of 38,696 “question–context–extractable answer” triples was automatically generated with guaranteed answer localization and annotated in the SQuAD format\([TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib30);[TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib31)\)\.
2. 2\.A hybrid retrieval method has been proposed and implemented, combining semantic indexing with the multilingual modelmultilingual\-e5\-largeand geospatial ranking using KD\-trees and haversine distance\. It has been shown that, with the optimal weighting coefficientα=0\.1\\alpha=0\.1and a radius of 50 km, the method achievesRecall@1=0\.988\\text\{Recall@1\}=0\.988,Recall@5=1\.000\\text\{Recall@5\}=1\.000, andMRR=0\.994\\text\{MRR\}=0\.994on 500 test queries, statistically significantly outperforming BM25, purely semantic, and purely spatial search\.
3. 3\.A multi\-architecture benchmark of extractive reading models \(RuBERT\-base, RuBERT\-large, XLM\-RoBERTa\-large\) trained on the synthetic QA corpus has been conducted\. The multilingual XLM\-RoBERTa\-large demonstrates near\-perfect Exact Match \(0\.992\) and F1 \(0\.994\) without any additional processing\. It has been established that the initially recorded complete failure of the RuBERT models on coordinate questions \(F1=0\\text\{F1\}=0\) is attributable to WordPiece tokenization artifacts and is completely eliminated by trivial post\-processing, after which both versions of RuBERT achieve 100% accuracy on all question categories with an inference speed 3\.5 times higher than that of XLM\-RoBERTa\-large\.
4. 4\.All resources created in the course of this work—the source dataset, the QA corpus, the weights of the trained models, and the interactive web application—are hosted in open access on the Hugging Face platform\([TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib30);[TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib31);[TatarNLPWorld,](https://arxiv.org/html/2605.05962#bib.bib32)\), ensuring full reproducibility of the results and the possibility of their immediate practical use\.

The practical significance of the developed system lies in its readiness for integration into geoinformation services, digital archives, educational platforms, and projects for the preservation of the cultural heritage of multilingual regions\. The achieved search recall and answer extraction accuracy metrics guarantee that a generative model within a RAG pipeline will receive relevant context in virtually all cases, which critically reduces the risk of hallucinations\.

Further research will be directed toward the implementation of adaptive geofiltering radius selection, cross\-validation of coordinates against external sources, end\-to\-end evaluation of a complete RAG pipeline involving generative models, as well as the expansion of the dataset with open\-ended questions\. The proposed approach can be scaled to other regions and language pairs, provided that toponymic datasets of comparable structure with coordinate referencing are available\.

## References

- M\. K\. Arabov and S\. S\. Khaybullina \(2026\)Adapting large language models to a low\-resource agglutinative language: a comparative study of lora and qlora for bashkir\.Note:arXiv preprint arXiv:2605\.04948Cited by:[§6](https://arxiv.org/html/2605.05962#S6.p9.1)\.
- M\. K\. Arabov \(2026a\)Tatar2Vec\.Kazan \(Volga Region\) Federal University,Kazan\.Note:Certificate of State Registration of Computer Program No\. 2026610619 dated 14\.01\.2026\(In Russian\)Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p2.1),[§2](https://arxiv.org/html/2605.05962#S2.p2.1)\.
- M\. K\. Arabov \(2026b\)TatarTokenizers\.Kazan \(Volga Region\) Federal University,Kazan\.Note:Certificate of State Registration of Computer Program No\. 2026611049 dated 16\.01\.2026\(In Russian\)Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p2.1),[§2](https://arxiv.org/html/2605.05962#S2.p2.1)\.
- E\. Artemova, M\. Zmeev, N\. Loukachevitch, I\. Rozhkov, T\. Batura, V\. Ivanov, and E\. Tutubalina \(2022\)RuNNE\-2022 shared task: recognizing nested named entities\.InProceedings of the Dialogue 2022 Conference,Moscow\.Cited by:[§2](https://arxiv.org/html/2605.05962#S2.p4.1)\.
- J\. L\. Bentley \(1975\)Multidimensional binary search trees used for associative searching\.Communications of the ACM18\(9\),pp\. 509–517\.Cited by:[§2](https://arxiv.org/html/2605.05962#S2.p3.1),[§3\.1](https://arxiv.org/html/2605.05962#S3.SS1.p3.5)\.
- R\. Burnashev, A\. Gatiatullin, and M\. Khamidullin \(2025\)Geolinguistic system for dialect similarity analysis based on associative rules and fuzzy logic\.InProceedings of the 10th International Conference on Computer Science and Engineering \(UBMK\),Istanbul, Turkiye,pp\. 1782–1785\.Cited by:[§2](https://arxiv.org/html/2605.05962#S2.p3.1)\.
- A\. A\. Choure, R\. B\. Adhao, and V\. K\. Pachghare \(2022\)NER in hindi language using transformer model: xlm\-roberta\.In2022 IEEE International Conference on Blockchain and Distributed Systems Security \(ICBDS\),Pune, India,pp\. 1–5\.External Links:[Document](https://dx.doi.org/10.1109/ICBDS53701.2022.9935841)Cited by:[§2](https://arxiv.org/html/2605.05962#S2.p4.1)\.
- A\. Conneau, K\. Khandelwal, N\. Goyal, V\. Chaudhary, G\. Wenzek, F\. Guzmán, E\. Grave, M\. Ott, L\. Zettlemoyer, and V\. Stoyanov \(2020\)Unsupervised cross\-lingual representation learning at scale\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics \(ACL\),Online,pp\. 8440–8451\.Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p3.1),[§2](https://arxiv.org/html/2605.05962#S2.p3.1),[§2](https://arxiv.org/html/2605.05962#S2.p4.1),[§3\.2](https://arxiv.org/html/2605.05962#S3.SS2.p4.2)\.
- J\. Devlin, M\.\-W\. Chang, K\. Lee, and K\. Toutanova \(2019\)BERT: pre\-training of deep bidirectional transformers for language understanding\.InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies \(NAACL\-HLT\),Minneapolis, Minnesota,pp\. 4171–4186\.Cited by:[§2](https://arxiv.org/html/2605.05962#S2.p4.1),[§3\.2](https://arxiv.org/html/2605.05962#S3.SS2.p4.2)\.
- S\. Efimovet al\.\(2020\)SberQuAD – russian reading comprehension dataset\.InProceedings of the 22nd Conference on Artificial Intelligence \(DCAI\),Cited by:[§2](https://arxiv.org/html/2605.05962#S2.p4.1),[§3\.2](https://arxiv.org/html/2605.05962#S3.SS2.p4.2)\.
- A\. Fenogenova, M\. Tikhonova, V\. Mikhailov, T\. Shavrina, A\. Emelyanov, D\. Shevelev, A\. Kukushkin, V\. Malykh, and E\. Artemova \(2021\)Russian superglue 1\.1: revising the lessons not learned by russian nlp models\.InComputational Linguistics and Intellectual Technologies: Papers from the Annual International Conference "Dialogue",Cited by:[§2](https://arxiv.org/html/2605.05962#S2.p4.1)\.
- F\. M\. Gafarov, V\. R\. Gafarova, and M\. M\. Ayupov \(2025\)Explainable artificial intelligence methods in text classification of machine learning models for tatar language\.InProceedings of the 10th International Conference on Computer Science and Engineering \(UBMK\),Istanbul, Turkiye,pp\. 1796–1800\.Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p2.1),[§2](https://arxiv.org/html/2605.05962#S2.p2.1)\.
- M\. Galimov, R\. Burnashev, and A\. Gatiatullin \(2023\)Designing a prototype of a fuzzy expert system for a dialectologist using geographic information systems and technologies\.InProceedings of the 8th International Conference on Computer Science and Engineering \(UBMK\),Burdur, Turkiye,pp\. 382–386\.Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p1.1),[§2](https://arxiv.org/html/2605.05962#S2.p3.1)\.
- \[14\]GeoNames\.Note:\[Electronic resource\]Accessed: 28\.02\.2026External Links:[Link](https://www.geonames.org/)Cited by:[§2](https://arxiv.org/html/2605.05962#S2.p1.1)\.
- R\. A\. Gilmullin and R\. R\. Gataullin \(2017\)Morphological analysis system of the tatar language\.InProceedings of the 9th International Conference on Computational Collective Intelligence \(ICCCI\),pp\. 519–528\.Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p2.1),[§2](https://arxiv.org/html/2605.05962#S2.p2.1)\.
- \[16\]IPSANTat\_monocorpus\_v2\.Note:Hugging Face\. \[Electronic resource\]Accessed: 28\.02\.2026External Links:[Link](https://huggingface.co/datasets/IPSAN/tat_monocorpus_v2/)Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p2.1),[§2](https://arxiv.org/html/2605.05962#S2.p2.1)\.
- J\. Johnson, M\. Douze, and H\. Jégou \(2019\)Billion\-scale similarity search with gpus\.IEEE Transactions on Big Data7\(3\),pp\. 535–547\.Cited by:[§3\.1](https://arxiv.org/html/2605.05962#S3.SS1.p2.3)\.
- P\. Liang, M\. Zhang,et al\.\(2022\)EmbEddings from bidirectional encoder representations \(e5\): a universal embedding model for information retrieval\.Note:arXiv preprint arXiv:2212\.03533Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p3.1),[§2](https://arxiv.org/html/2605.05962#S2.p3.1),[§3\.1](https://arxiv.org/html/2605.05962#S3.SS1.p2.3)\.
- A\. Mindubaev and A\. R\. Gatiatullin \(2024\)Problems of semantic relation extraction from tatar text corpus ‘tugan tel’\.InProceedings of the 3rd International Conference on Problems of Informatics, Electronics and Radio Engineering \(PIERE\),Novosibirsk, Russian Federation,pp\. 1690–1694\.Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p2.1),[§2](https://arxiv.org/html/2605.05962#S2.p2.1)\.
- D\. R\. Mukhamedshin, A\. R\. Gatiatullin, N\. A\. Prokopyev, and R\. A\. Gilmullin \(2025\)Semantic annotation in electronic corpus of tatar language ‘tugan tel’ based on knowledge graph\.InProceedings of the 10th International Conference on Computer Science and Engineering \(UBMK\),Istanbul, Turkiye,pp\. 1792–1795\.Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p2.1),[§2](https://arxiv.org/html/2605.05962#S2.p2.1)\.
- C\. Raffel, N\. Shazeer, A\. Roberts, K\. Lee, S\. Narang, M\. Matena, Y\. Zhou, W\. Li, and P\. J\. Liu \(2020\)Exploring the limits of transfer learning with a unified text\-to\-text transformer\.Journal of Machine Learning Research21,pp\. 1–67\.Cited by:[§2](https://arxiv.org/html/2605.05962#S2.p4.1),[§3\.2](https://arxiv.org/html/2605.05962#S3.SS2.p4.2)\.
- P\. Rajpurkar, J\. Zhang, K\. Lopyrev, and P\. Liang \(2016\)SQuAD: 100,000\+ questions for machine comprehension of text\.InProceedings of the 2016 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),Austin, Texas,pp\. 2383–2392\.Cited by:[§2](https://arxiv.org/html/2605.05962#S2.p4.1),[§3\.2](https://arxiv.org/html/2605.05962#S3.SS2.p1.1),[§4\.2](https://arxiv.org/html/2605.05962#S4.SS2.p5.1)\.
- R\. Rehurek and P\. Sojka \(2006\)Gensim – python framework for vector space modelling\.InProceedings of the 9th International Conference on Text, Speech and Dialogue \(TSD\),Brno, Czech Republic,pp\. 45–50\.Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p1.1),[§2](https://arxiv.org/html/2605.05962#S2.p3.1)\.
- \[24\]M\. R\. Saykhunov, R\. R\. Khusainov, T\. I\. Ibragimov,et al\.Written corpus of the tatar language\.Note:\[Electronic resource\]Accessed: 03\.01\.2026External Links:[Link](https://www.corpus.tatar/en)Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p2.1),[§2](https://arxiv.org/html/2605.05962#S2.p2.1)\.
- S\. Schweter and S\. Tekir \(2020\)BERT for turkish: a comprehensive evaluation\.InProceedings of the 12th Language Resources and Evaluation Conference \(LREC\),Marseille, France,pp\. 4325–4332\.Cited by:[§2](https://arxiv.org/html/2605.05962#S2.p3.1)\.
- R\. W\. Sinnott \(1984\)Virtues of the haversine\.Sky and Telescope68\(2\),pp\. 159\.Cited by:[§2](https://arxiv.org/html/2605.05962#S2.p3.1),[§3\.1](https://arxiv.org/html/2605.05962#S3.SS1.p3.3)\.
- D\. Sh\. Suleymanov, A\. Ya\. Fridman, R\. A\. Gilmullin, and B\. A\. Kulik \(2021\)System analysis of the problem of natural language modeling\.Transactions of the Kola Science Centre\. Information Technologies12\(5\),pp\. 57–66\.Note:\(In Russian\)Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p1.1)\.
- \[28\]Tatar mixed corpus – tatar corpus from the web\.Note:Sketch Engine\. \[Electronic resource\]Accessed: 03\.01\.2026External Links:[Link](https://www.sketchengine.eu/tatar-corpus-from-the-web/)Cited by:[§1](https://arxiv.org/html/2605.05962#S1.p2.1),[§2](https://arxiv.org/html/2605.05962#S2.p2.1)\.
- \[29\]TatarNLPWorldTatarstan toponyms dataset\.Note:Hugging Face\. \[Electronic resource\]Accessed: 28\.02\.2026External Links:[Link](https://huggingface.co/datasets/TatarNLPWorld/tatarstan-toponyms)Cited by:[§4\.1](https://arxiv.org/html/2605.05962#S4.SS1.p1.1),[§4\.2](https://arxiv.org/html/2605.05962#S4.SS2.p10.1),[§6](https://arxiv.org/html/2605.05962#S6.p8.1),[item 1](https://arxiv.org/html/2605.05962#S7.I1.i1.p1.1),[item 4](https://arxiv.org/html/2605.05962#S7.I1.i4.p1.1)\.
- \[30\]TatarNLPWorldTatarstan toponyms qa dataset\.Note:Hugging Face\. \[Electronic resource\]Accessed: 28\.02\.2026External Links:[Link](https://huggingface.co/datasets/TatarNLPWorld/tatarstan-toponyms-qa)Cited by:[§3\.2](https://arxiv.org/html/2605.05962#S3.SS2.p5.1),[§4\.2](https://arxiv.org/html/2605.05962#S4.SS2.p10.1),[§6](https://arxiv.org/html/2605.05962#S6.p8.1),[item 1](https://arxiv.org/html/2605.05962#S7.I1.i1.p1.1),[item 4](https://arxiv.org/html/2605.05962#S7.I1.i4.p1.1)\.
- \[31\]TatarNLPWorldToponymic rag explorer\.Note:Hugging Face Spaces\. \[Electronic resource\]Accessed: 28\.02\.2026External Links:[Link](https://huggingface.co/spaces/TatarNLPWorld/tatar-toponym-rag-space)Cited by:[§6](https://arxiv.org/html/2605.05962#S6.p8.1),[item 4](https://arxiv.org/html/2605.05962#S7.I1.i4.p1.1)\.
- \[32\]TatarNLPWorldTurkic nlp & low resource languages research hub\.Note:Hugging Face\. \[Electronic resource\]Accessed: 28\.02\.2026External Links:[Link](https://huggingface.co/TatarNLPWorld)Cited by:[§2](https://arxiv.org/html/2605.05962#S2.p1.1),[§2](https://arxiv.org/html/2605.05962#S2.p2.1)\.