Large-scale Knowledge Graphs (KGs) are increasingly relevant for humanities research, yet querying them via SPARQL poses challenges for non-technical users. While Text-to-SPARQL studies predominantly target popular KGs such as Wikidata or DBpedia, domain-specific KGs remain underexplored. This paper introduces a bilingual (English-Spanish) dataset designed for evaluating automatic text-to-SPARQL translation on GOLEM, a humanities KG containing metadata and extracted features from fanfiction stories hosted on Archive of Our Own (AO3). The dataset includes 477 manually crafted natural language questions paired with gold SPARQL queries, augmented to 1,895 questions through automatic paraphrasing. We benchmark several Large Language Models (LLMs) with prompt-based approaches, particularly examining in-context learning methods that select prompt examples based on semantic similarity, which yield the best results. Error analysis highlights entity linking as essential for improving query generation. This work provides practical insights and opens pathways for future research on natural language interfaces for querying domain-specific KGs in Digital Humanities. The dataset and output of our experiments are available at: https://github.com/GOLEM-lab/GOLEM_Text-to-SPARQL.
• A novel bilingual (English and Spanish) dataset containing 477 NL questions, each paired with manually curated SPARQL queries, that can be used to interrogate the GOLEM KG SPARQL endpoint. The dataset is augmented into 1,895 NL questions by automatically generated paraphrases.
We find that few-shot learning methods improve performance by selecting prompt examples based on semantic similarity to the input question. We provide a granular error analysis to identify current method limitations, laying the basis for future improvements.
The paper is structured as follows. In Section 2, we review existing text-to-SPARQL datasets, summarise recent approaches leveraging LLM-based prompting methods for NL-to-SPARQL translation, and position our work relative to these studies. Section 3 introduces GOLEM KG, describes the methodology for dataset construction, and provides a qualitative and quantitative characterisation of the resulting bilingual dataset. We then detail our experimental settings, including prompting strategies, LLMs tested, computational infrastructure, and evaluation metrics. In Section 4, we present experimental results along with an in-depth error analysis. Finally, we discuss conclusions and directions for future work. Our dataset and the output of our experiments are available at: https://github.com/GOLEM-lab/GOLEM_Text-to-SPARQL.
The transformation of NL questions into SPARQL queries (Text-to-SPARQL) is widely studied in
Knowledge Graph Question Answering (KGQA). However, existing Text-to-SPARQL methods do not
generalise well, as most research eforts rely on training or fine-tuning models [
With the emergence of LLMs, Text-to-SPARQL became intensively explored as an NL interface
for KGs [16]. Prompt-based techniques fostered the development of Text-to-SPARQL resources and
experiments beyond mainstream KGs toward domain-specific use cases in healthcare, scholarly data, and
cultural heritage. Sivasubramaniam et al. [17] experimented in the medical sector with electronic health
records (EHRs) and showed that SPARQL is underrepresented in LLM pre-training data, highlighting its
complexity and demonstrating SPARQL as the most complex language for LLM-based query generation
across diferent settings (zero-shot, few-shot, presence or absence of KG schema information in prompts).
Sequeda et al. [
Having discussed available datasets, we now turn to recent prompt-based Text-to-SPARQL approaches
leveraging LLMs. Zahera et al. [
In this work, we test LLMs’ capability on the Text-to-SPARQL task. In particular, we evaluate
the flexibility of in-context learning with semantic search, avoiding costly fine-tuning and extensive
collection of NL-SPARQL examples. With this objective, we adapt Dynamic Few-Shot Learning [
The dataset construction has been curated manually by a Language and Communication Technologies
master student (author of this paper) as part of a curricular research internship. The annotator is
proficient in English and is a native Spanish speaker. This section describes our approach, introducing
GOLEM KG [
The GOLEM KG contains metadata and extracted features of fanfiction stories in various languages
from the popular online platform Archive of Our Own (AO3) [
Since the data in the GOLEM KG is multilingual, and potential end-users may want to query the KG in their native language, we aim to develop a multilingual dataset. We constructed NL questions in English and Spanish and paired them with their corresponding SPARQL query. We formulated the questions based on typical queries that end users interested in fanfiction might ask of the KG. Our methodology can be extended to additional languages. Dataset statistics are reported in Table 1. The following paragraphs describe how the dataset was constructed based on the included question types. 1https://cidoc-crm.org//extensions/lrmoo/html/LRMoo_v1.0.html 2https://www.dublincore.org/specifications/dublin-core/dcmi-terms/ 3http://search.golemlab.eu:3006/en/ 4http://graph.golemlab.eu:8890/sparql
The first steps towards creating a multilingual dataset of NL question-SPARQL query pairs in English and Spanish involved producing 39 generic template questions per language. Spanish questions are translations of English questions. These templates use placeholders instead of specific KG entities. There are several types of placeholders, each referring to one GOLEM predicate: • [[fandom]], corresponding to https://golemlab.eu/graph/fandom; • [[story]], corresponding to https://golemlab.eu/graph/title; • [[character]], corresponding to https://golemlab.eu/graph/character; • [[keyword]], corresponding to https://golemlab.eu/graph/keyword.
Below is one example of a template question per language: • How many [[fandom]] stories are there? • ¿CuÃąntas historias de [[fandom]] hay publicadas?
Among the template questions, some do not contain placeholders because they do not require the inclusion of any KG entities. Below is one example per language: • How many stories are tagged as explicit? • ¿CuÃąntas historias estÃąn marcadas como explÃŋcitas?
Before execution on the KG’s SPARQL endpoint, these generic template questions require an instantiation step, described in the following paragraph.
We instantiate the template questions by replacing placeholders with actual entities from the KG when required. These entities are obtained through a series of simple queries on the KG, performing one query per placeholder type. Below is an example of a query built for this purpose: prefix golem: <https://golemlab.eu/graph/> SELECT DISTINCT ?fandom WHERE {
?story golem:fandom ?fandom .
}
The resulting lists of instances are then used at random to replace the placeholders in the dataset, both in the NL question and its corresponding SPARQL query.
We refer to the resulting questions as instantiated questions, which can directly query the KG via the SPARQL endpoint. We created 10 instantiated questions for each template question. Each instantiated question has been paired with a manually crafted SPARQL query. If the corresponding SPARQL query generated an empty answer, the sample was removed. Below, we report two examples: • How many 1984 - George Orwell stories are there? prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT (COUNT(DISTINCT ?story) as ?uploads) WHERE {
?story golem:fandom "1984 - George Orwell" . • ¿CuÃąntas historias de Valley of Tears (TV) hay publicadas? prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT (COUNT(DISTINCT ?story) as ?uploads) WHERE {
?story golem:fandom "Valley of Tears (TV)" . prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT COUNT(?story) as ?explicit_stories WHERE {
?story golem:rating "Explicit" .
}
The template questions that do not contain any placeholders do not need to undergo any instantiation process and are directly associated with the corresponding SPARQL queries, as in the example below:
The two questions in English and Spanish reported in the above example are semantically equivalent (the Spanish one is the translation of the English one) and are therefore paired with the same SPARQL query.
The SPARQL queries corresponding to the instantiated questions are executed against the GOLEM KG SPARQL endpoint, and their responses are collected and evaluated.
Instantiated questions are paraphrased via data augmentation using deepseek-r1-7b. However, Spanish questions are frequently misparaphrased into English. We address this by verifying the correctness of the paraphrased questions’ language using a cross-check that implements the language detection library lingua5. Paraphrases detected by lingua in languages difering from the manually annotated source are discarded.6 Below, we list some examples of paraphrased questions: • What is the number of George Orwell 1984 works available? • How many George Orwell’s 1984 pieces exist? • ¿CuÃąntas narrativas de *Valley of Tears* (TV) existen ?
The prompt used in the paraphrasing step can be seen in A.1.1.
In this section, we first introduce the prompting approaches tested (Zero-shot, Naive Few-shot, and
an adaptation of Dynamic Few-shot Learning [
Zero-shot In the Zero-shot (ZS) prompting approach, we include in the prompt only the task instructions and the list of predicates from the KG. We report an example of ZS prompt per language in Appendix A.1.2.
Naive few-shots In the Naive Few-shot (NFS) prompting approach, we include in the prompt the task instructions and the list of predicates from the KG. Plus, we include three random examples in the prompt, taken from the same language as the question being processed. We report an example of an NFS prompt per language in Appendix A.1.3.
Dynamic Few-shot Learning (adapted) In the adapted Dynamic Few-shot Learning (a-DFSL)
prompting approach, we include in the prompt the task instructions and the list of predicates from the
KG. Additionally, adapting what is described in [
In this section, we present the results of our experiments, and we perform a detailed analysis of the errors.
Precision consistently exceeds recall across all models and prompting approaches, indicating more frequent query failures or empty results rather than incorrect answers. For all the models tested across all prompting approaches (except the zero-shot approach, which is less relevant due to its extremely low performance), performance on the English-language subset is better than on the Spanish-language subset.
Thus, we select deepseek-coder-v2 (a-DFSL) for further experiments on the augmented dataset
(Table 3). Performance notably declines on augmented data due to entity alterations introduced by
paraphrasing. Entity mismatches between paraphrases and KG canonical forms cause the SPARQL
query to obtain erroneous results. Such results closely align with the ablation studies performed in the
original DFSL study [
We conducted a systematic error analysis on all 70 errors produced by our best-performing model (deepseek-coder-v2), which optimises the accuracy-runtime tradeof. Notably, our analysis could not spot the "triple-flip" errorâĂŤa well-documented phenomenon in text-to-SPARQL conversion where subject and object positions are reversed in generated triples, resulting in empty or incorrect query results. The identified errors were instead categorised into two primary classes: (i) Failed execution, covering queries unable to execute successfully against the SPARQL endpoint (e.g., QueryBadFormed, timeouts), and (ii) Incorrect results, referring to queries executed successfully but returning answers difering from the gold annotations. The second macro-category was further divided into specific error types (see Table 4), namely Failed KG Entity Recognition, where the model incorrectly recognized entities from the NL question; Wrong Predicate, where the model selected an incorrect predicate among those provided in the prompt; Incomplete Query, where generated queries lacked suficient complexity to retrieve all expected results; and SPARQL Syntax Error, involving syntactic mistakes in the generated query that led to incorrect results. Additionally, we identified errors attributable to the evaluation approach itself (Evaluation Method Errors), where queries returned correct results but were mistakenly lfagged as wrong due to minor variations in variable naming or result grouping. Future work will address these evaluation inaccuracies.
Table 5 illustrates each error type among queries executed successfully but returning incorrect results. In the first example, comparing the gold SPARQL query with the generated one reveals that both queries are structurally identical, difering only in the KG entity used as the object of predicate golem:title. The generated query incorrectly uses the entity "Wolfstar prompts" instead of the correct entity "(fanart) Wolfstar prompts", causing an empty (incorrect) result.
Another common issue is represented by queries of the error type Incomplete Query, exemplified by subjective NL questions such as "What is Forbidden Like The Forest about?", reported as the second example in the table. The annotator associated this question with a highly articulated SPARQL query, which retrieves multiple detailed elements (e.g., keywords, romantic categories, content warnings, collections, series, and summaries). In contrast, the model produced a minimal query, interpreting the ambiguous phrase "about" simply as retrieving the story’s summary. This discrepancy highlights a modelling issue where subjective interpretations by annotators may lead to overly complex gold standard queries compared to the minimalistic outputs generated by the model. In future work, annotation guidelines will be refined to address this kind of subjectivity.
A final error type illustrated in the table is Wrong Predicate, exemplified by the fourth and last query pair. The gold SPARQL query correctly uses the predicate golem:author, whereas the modelgenerated query mistakenly selects the predicate dc:creator. In such cases, the model erroneously selects a predicate from among those provided in the prompt or hallucinates predicates that do not exist in the KG, leading to incorrect or incomplete results.
To analyse the errors of the run on the paraphrased question dataset, reported in Table 3, we selected a sample of 30 erroneous cases. In this sample, we do not consider those errors due to the failed execution of the query. Questions were chosen to maximise semantic diversity. Semantic embeddings of questions were computed using the same model used for selecting a-DFSL examples. A greedy selection method then iteratively picked the most semantically distinct questions, measured by cosine distance, ensuring coverage of the broadest possible range of error cases. This approach allowed us to eficiently maximise the variety of errors considered while minimising redundancy in our analysis.
As we report in Table 4.2, unlike the analysis of instantiated questions, no instances of Incomplete Query errors emerged, possibly due to the smaller sample size. However, we identified a new error category, Misleading Paraphrase, grouping queries arising from flawed or unclear paraphrased questions. Typical examples include nonsensical, mixed-language, or malformed paraphrases such as “It never entered my mind’s key points or topics”, “¿CÃşmo many kudos se le otorgan a If only?”, or “¿QuÃľ pÃąjaros aparecen en If You Could Be Anywhere?”. Future analyses will focus on addressing this error type through improved paraphrase quality control.
One limitation of our dataset construction is that the questions were not directly informed by domain experts or end-users from computational literature or fanfiction communities. Instead, they were primarily inspired by information already present in the KG, which could potentially introduce bias and limit method efectiveness. Future work should involve gathering explicit requirements from target endusers to create more representative queries. Additionally, the dataset augmentation process produced paraphrases with varying linguistic quality, including language inconsistencies such as mixed-language outputs (e.g., English instead of expected Spanish). To address this, future eforts should integrate automatic quality control mechanisms, such as leveraging LLMs as evaluators, or conduct manual assessments to enhance dataset quality. Some collected NL questions exhibit subjective interpretations, for example, queries about the content of stories, which lead to subjective gold SPARQL annotations. Future work will refine the annotation guidelines to minimise ambiguity.
Another limitation concerns the systemâĂŹs generalizability beyond the semi-templated scope of the constructed dataset. While the tested methodology demonstrates efectiveness within the instantiated questions, its performance degrades on paraphrased variants, indicating limited robustness to linguistic variation. Furthermore, its behaviour in more open-ended or less structured query scenarios remains underexplored. Future work will involve evaluating system performance on more naturalistic and user-authored queries.
We acknowledge the explicit predicate listing as a simplification. In deployed applications, these would be stored as pre-defined variables hidden from end-users. Future work should examine how prompt complexity afects performance and explore more compact schema representations.
The best-performing prompting approach (a-DFSL) is a simplified version of the original DFSL [
The current evaluation strategy compares answers from generated and gold SPARQL queries, but has limitations. Queries returning identical results with diferent groupings or slight variable diferences are incorrectly marked as errors. Future work will refine the evaluation method for greater flexibility and accuracy.
In this work, we introduced a bilingual dataset of NL in English and Spanish questions paired with
corresponding SPARQL queries. The dataset targets the GOLEM KG, containing metadata and features
extracted from fanfiction hosted on the Archive of Our Own (AO3) platform. The dataset comprises
instantiated questions, manually crafted and automatically populated with KG entities, and an augmented
version generated via automatic paraphrasing. We used the dataset to benchmark various LLMs on
the text-to-SPARQL task, exploring several prompting strategies. An adapted, simplified version of
DFSL [
The author(s) have not employed any Generative AI tools.
This work is part of the Graphs and Ontologies for Literary Evolution Models (GOLEM) project funded by the European Commission. Jose Maldonado-RodrÃŋguez is supported by the Erasmus Mundus Masters Program in Language and Communication Technologies (LCT), EU grant no. 2019-1508. Arianna Graciotti is supported by the European UnionâĂŹs Horizon 2020 research and innovation programme under grant agreement No 101004746. [6] A. Perevalov, D. Diefenbach, R. Usbeck, A. Both, Qald-9-plus: A multilingual dataset for question answering over dbpedia and wikidata translated by native speakers, in: 2022 IEEE 16th International Conference on Semantic Computing (ICSC), 2022, pp. 229–234. doi:10.1109/ICSC52841. 2022.00045. [7] L.-A. Kafee, S. Razniewski, P. Vougiouklis, R. Usbeck, X. Yan, A. Perevalov, L. Jiang, J. Schulz, A. Kraft, C. MÃűller, J. Huang, J. Reineke, A.-C. N. Ngomo, M. Saleem, A. Both, Qald-10 âĂŞ the 10th challenge on question answering over linked data: Shifting from dbpedia to wikidata as a kg for kgqa, Semantic Web 15 (2024) 2193–2207. URL: https://doi.org/10.3233/SW-233471. doi:10.3233/SW-233471. arXiv:https://doi.org/10.3233/SW-233471. [8] S. Cao, J. Shi, L. Pan, L. Nie, Y. Xiang, L. Hou, J. Li, B. He, H. Zhang, KQA pro: A dataset with explicit compositional programs for complex question answering over knowledge base, in: S. Muresan, P. Nakov, A. Villavicencio (Eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, Dublin, Ireland, 2022, pp. 6101–6119. URL: https://aclanthology.org/2022.acl-long.422/. doi:10.18653/v1/2022.acl-long.422. [9] M. Dubey, D. Banerjee, A. Abdelkawi, J. Lehmann, Lc-quad 2.0: A large dataset for complex question answering over wikidata and dbpedia, in: C. Ghidini, O. Hartig, M. Maleshkova, V. Svátek, I. Cruz, A. Hogan, J. Song, M. Lefrançois, F. Gandon (Eds.), The Semantic Web – ISWC 2019, Springer International Publishing, Cham, 2019, pp. 69–78. [10] P. Trivedi, G. Maheshwari, M. Dubey, J. Lehmann, Lc-quad: A corpus for complex question answering over knowledge graphs, in: C. d’Amato, M. Fernandez, V. Tamma, F. Lecue, P. Cudré-Mauroux, J. Sequeda, C. Lange, J. Heflin (Eds.), The Semantic Web – ISWC 2017, Springer International Publishing, Cham, 2017, pp. 210–218. [11] J. Berant, A. Chou, R. Frostig, P. Liang, Semantic parsing on Freebase from question-answer pairs, in: D. Yarowsky, T. Baldwin, A. Korhonen, K. Livescu, S. Bethard (Eds.), Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Seattle, Washington, USA, 2013, pp. 1533–1544. URL: https:// aclanthology.org/D13-1160/. [12] Y. Su, H. Sun, B. Sadler, M. Srivatsa, I. Gür, Z. Yan, X. Yan, On generating characteristic-rich question sets for QA evaluation, in: J. Su, K. Duh, X. Carreras (Eds.), Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Austin, Texas, 2016, pp. 562–572. URL: https://aclanthology.org/D16-1054/. doi:10. 18653/v1/D16-1054. [13] Y. Gu, S. Kase, M. Vanni, B. Sadler, P. Liang, X. Yan, Y. Su, Beyond i.i.d.: Three levels of generalization for question answering on knowledge bases, in: Proceedings of the Web Conference 2021, WWW ’21, Association for Computing Machinery, New York, NY, USA, 2021, p. 3477âĂŞ3488. URL: https://doi.org/10.1145/3442381.3449992. doi:10.1145/3442381.3449992. [14] D. Keysers, N. SchÃďrli, N. Scales, H. Buisman, D. Furrer, S. Kashubin, N. Momchev, D. Sinopalnikov, L. Stafiniak, T. Tihon, D. Tsarkov, X. Wang, M. van Zee, O. Bousquet, Measuring compositional generalization: A comprehensive method on realistic data, 2020. URL: https://arxiv.org/abs/1912. 09713. arXiv:1912.09713. [15] D. Banerjee, S. Awale, R. Usbeck, C. Biemann, DBLP-QuAD: A question answering dataset over the DBLP scholarly knowledge graph, in: BIR 2023: 13th International Workshop on Bibliometricenhanced Information Retrieval at ECIR 2023, 2023. [16] H. Khorashadizadeh, F. Z. Amara, M. Ezzabady, F. Ieng, S. Tiwari, N. Mihindukulasooriya, J. Groppe, S. Sahri, F. Benamara, S. Groppe, Research trends for the interplay between large language models and knowledge graphs, 2024. URL: https://arxiv.org/abs/2406.08223. arXiv:2406.08223. [17] S. Sivasubramaniam, C. E. Osei-Akoto, Y. Zhang, K. Stockinger, J. Fuerst, Sm3-textto-query: Synthetic multi-model medical text-to-query benchmark, in: A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, C. Zhang (Eds.), Advances in Neural Information Processing Systems, volume 37, Curran Associates, Inc., 2024, pp. 88627–88663. URL: https://proceedings.neurips.cc/paper_files/paper/2024/file/
In the sections below, we report the prompts used in the diferent experimental settings tested in this paper.
1 Generate 10 paraphrases for the following text, making the responses short and querylike. Please do not say anything after </think>, only the paraphrases. If there are any words within double square brackets (such as [[story]] or [[character]] ), please do not modify them.: 2 3 How many 1984 - George Orwell stories are there?
Listing 1: English question ata augmentation prompt example
SPANISH 1 Generate 10 paraphrases for the following text, making the responses short and querylike. Please do not say anything after </think>, only the paraphrases. If there are any words within double square brackets (such as [[story]] or [[character]] ), please do not modify them.: 2 3 £CuÃąntas historias de Valley of Tears (TV) hay publicadas?
Listing 2: English question ata augmentation prompt example
Listing 3: English question ZS Prompt Example
Listing 4: Spanish question NFS Prompt Example
Question: How many chapters does Guardian of Hogwarts have? Query: <SPARQL> prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT DISTINCT ?chapters WHERE { ? story golem:title "Guardian of Hogwarts" . ?story golem:numberOfChapters ? chapters . } </SPARQL> Question: What is the average number of comments for stories from Glory of the
Special Forces (TV)? Query: <SPARQL> prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT AVG(?comments) AS ? avg_number_of_comments WHERE { ?story golem:fandom "Glory of the Special Forces (TV)" . ?story golem:numberOfComments ?comments . } </SPARQL> Question: Is the story Luna Lovegood and the Chamber of Innocence completed? Query: <SPARQL> prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT DISTINCT ?status WHERE { ? story golem:title "Luna Lovegood and the Chamber of Innocence" . ?story golem: publicationStatus ?status . } </SPARQL> 77 78 79 ### 80 81 82 Question: How many 1984 - George Orwell stories are there? 83 84 Query:
Listing 5: English question NFS Prompt Example
1 Your task is to translate a question in natural language into a SPARQL query for the
GOLEM knowledge graph. 2 The query must follow these guidelines: 3 1. SPARQL queries must include the following prefix: 4 prefix golem: <https://golemlab.eu/graph/> 5 2. Enclose SPARQL queries within <SPARQL> </SPARQL> tags.
Question: £CuÃąntos capÃŋtulos tiene this title is a wip of which i hate, like this fic and my life? 53 54 55 56
Query: <SPARQL> prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT DISTINCT ?chapters WHERE { ? story golem:title "this title is a wip of which i hate, like this fic and my life" . ?story golem:numberOfChapters ?chapters . } </SPARQL> 60 61 62 Question: £EstÃą completada la historia A Small Steep Valley? Query: <SPARQL> prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT DISTINCT ?status WHERE { ? story golem:title "A Small Steep Valley" . ?story golem:publicationStatus ? status . } </SPARQL> 77 78 79 ### 80 81 82 Question: £CuÃąntas historias de Valley of Tears (TV) hay publicadas? 83 84 Query:
Listing 6: Spanish question NFS Prompt Example
Below are two examples of prompts for transforming natural language questions into SPARQL queries, one in English and one in Spanish.
1 Your task is to translate a question in natural language into a SPARQL query for the
GOLEM knowledge graph. 2 The query must follow specific guidelines to ensure accuracy and correctness: 3 1. SPARQL queries must include the following prefix: 4 prefix golem: <https://golemlab.eu/graph/> 5 2. Enclose SPARQL queries within <SPARQL> </SPARQL> tags. 6 3. You must generate 1 query(ies). 7 4. It is very important that you use only the predicates provided below. 8 5. Examples are provided below for guidance. 9 10 ### 11 Predicates: 12 https://golemlab.eu/graph/numberOfComments 13 https://golemlab.eu/graph/numberOfKudos 14 https://golemlab.eu/graph/publicationStatus 15 https://golemlab.eu/graph/dateModified 16 https://golemlab.eu/graph/characters
Question: How many stories are there on Archive of Our Own? Query: <SPARQL> prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT COUNT(DISTINCT ?story) as ? stories WHERE { ?story golem:story_id ?id } </SPARQL> ### Question: How many åijĂåŸŐ stories are there? Query: <SPARQL> prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT (COUNT(DISTINCT ?story) as ? uploads) WHERE { ?story golem:fandom "åijĂåŸŐ" . } </SPARQL> ###
Question: How many Kudos do Mr.Vampire (1985) stories get on average? Query: <SPARQL> prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT AVG(?kudos) AS ? average_number_of_kudos WHERE { ?story golem:fandom "Mr.Vampire (1985)" . ?story golem:numberOfKudos ?kudos . } </SPARQL>
Listing 7: English question a-DFSL Prompt Example
1 Your task is to translate a question in natural language into a SPARQL query for the
GOLEM knowledge graph. 2 The query must follow specific guidelines to ensure accuracy and correctness: 3 1. SPARQL queries must include the following prefix: 4 prefix golem: <https://golemlab.eu/graph/> 5 2. Enclose SPARQL queries within <SPARQL> </SPARQL> tags. 6 3. You must generate 1 query(ies). 7 4. It is very important that you use only the predicates provided below. 8 5. Examples are provided below for guidance. 9 10 ### 11 Predicates: 12 https://golemlab.eu/graph/numberOfComments 13 https://golemlab.eu/graph/numberOfKudos 14 https://golemlab.eu/graph/publicationStatus 15 https://golemlab.eu/graph/dateModified 16 https://golemlab.eu/graph/characters 17 https://golemlab.eu/graph/collections 18 https://golemlab.eu/graph/fandom 19 https://golemlab.eu/graph/publisher 20 https://golemlab.eu/graph/rating 21 https://golemlab.eu/graph/series 22 https://golemlab.eu/graph/story_id 23 https://golemlab.eu/graph/summary 24 https://golemlab.eu/graph/numberOfChapters 25 https://golemlab.eu/graph/datePublished 26 https://golemlab.eu/graph/keyword 27 https://golemlab.eu/graph/contentWarning 28 https://golemlab.eu/graph/numberOfWords 29 https://golemlab.eu/graph/socialRelationships 30 https://golemlab.eu/graph/datePackaged 31 https://golemlab.eu/graph/romanticCategory 32 https://golemlab.eu/graph/noOfPairings 33 https://golemlab.eu/graph/topPartner 34 https://golemlab.eu/graph/topPartnerPairings 35 https://golemlab.eu/graph/averageWordLength 36 https://golemlab.eu/graph/MSTTR Query: <SPARQL> prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT (COUNT(DISTINCT ?story) as ? uploads) WHERE { ?story golem:fandom "[[fandom]]" . } </SPARQL> Question: £CuÃąntas historias se publican al aÃśo? Query: <SPARQL> prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT (COUNT(DISTINCT ?story) as ? uploads) WHERE { ?story golem:story_id ?id } </SPARQL> Question: £CuÃąntos autores han publicado alguna historia? Query: <SPARQL> prefix golem: <https://golemlab.eu/graph/> prefix dc: <http://purl.org/dc/terms/> prefix gc: <https://ontology.golemlab.eu/> SELECT (COUNT(DISTINCT ?author) as ? uploads) WHERE { ?story golem:author ?author . } </SPARQL>