Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

Text-to-SPARQL Goes Beyond English: Multilingual Question Answering Over Knowledge Graphs through Human-Inspired Reasoning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An agent that plans, links entities, and refines queries beats prior multilingual Text-to-SPARQL systems without fine-tuning.

desk verdict A useful agent recipe for multilingual Text-to-SPARQL; the SOTA claim rests on unshown challenge results and a possible leakage path. read the letter →

arxiv 2507.16971 v1 pith:Y7GNLV6Z submitted 2025-07-22 cs.CL cs.AIcs.IR

Aleksandr Perevalov , Andreas Both This is my paper
classification cs.CLcs.AIcs.IR
keywords LLMagentsText-to-SPARQLKnowledgegraphquestionansweringMultilingualsemanticparsingIn-contextlearningExperiencepoolMachinetranslation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that multilingual question answering over knowledge graphs improves when a large language model does not write the SPARQL query in one pass but instead acts as an agent that plans, links entities, drafts, executes, and revises. The authors' mKGQAgent decomposes the task into a plan step, an action step with a named-entity linking tool, and a feedback step that runs the candidate query on a triplestore once, and it augments these steps with an experience pool of past successful and failed attempts retrieved by embedding similarity. On the QALD-9-plus benchmark across ten languages, the GPT-4o configuration reaches 54.83% F1 on English, ahead of every baseline, and remains the best system on German, Spanish, Belarusian, and Bashkir; the paper reports that the full system took first place in the Text2SPARQL 2025 challenge. The significance is that the gains come without supervised fine-tuning, pointing to task decomposition and non-parametric memory as an alternative route to multilingual semantic parsing.

What carries the argument

The carrying mechanism is the experience-pool-augmented agent loop. The plan step produces a step-by-step list of subtasks in the question's language; the action step executes each subtask with the same LLM, binding in a named-entity/relation-linking tool that maps candidate names to knowledge-graph URIs the model could not know; the feedback step takes the draft query, executes it on the triplestore, and feeds the response back for exactly one revision. The experience pool is built offline by running a simpler agent over the training subset and storing each question's embedding together with its plan, chat history, generated query, gold query, and F1 score; retrieval injects the top-N similar plans and top-N similar query examples into the plan and action prompts. This design carries the argument because it converts past successes and failures into prompt-level memory, and the ablation traces the largest single gain to action-step experience retrieval combined with GPT-4o.

What would settle it

Take each QALD-9-plus test question, compute the embedding of the question and of every sentence in the experience pool, and check whether any retrieved top-N example matches the test question or its machine-translated or paraphrased version. If such matches exist, rerun the evaluation with those examples excluded; a substantial F1 drop would show part of the reported gain is retrieval leakage rather than agent reasoning.

Watch

Extended reading notes

Core claim

mKGQAgent's central claim is that the bottleneck in multilingual Text-to-SPARQL is task organisation rather than the base model's query-language competence. Splitting generation into planning, tool-augmented action, and a single triplestore-feedback revision produces large gains over monolithic prompting: every component in the ablation adds positive F1, and the combined agent with GPT-4o reaches 54.83% on English, beats all baselines including 50.00% for HQA (GPT-4), and scores 43.08% on German, 38.28% on Spanish, 31.56% on Belarusian, and 40.48% on Bashkir. The experience pool, a vector store of training-set questions paired with plans, chat histories, generated and gold SPARQL, and F1 scores, lets the agent condition its prompts on similar past attempts, and the feedback step corrects the draft from the actual triplestore response. The paper presents this as a route to strong multilingual KGQA that avoids fine-tuning and preserves the base model's generalisation.

Load-bearing premise

The result rests on the assumption that embedding retrieval from the experience pool never returns the test question itself or a near-duplicate translation of it; the paper does not check similarity between retrieved examples and test questions, so a leak would inflate the reported F1 scores.

Editorial extensions

If this is right

  • With GPT-4o, the full mKGQAgent raises F1 on English QALD-9-plus from a 34.37% no-memory baseline to 54.83%, a 59.53% relative improvement.
  • Each component contributes: the plan-step experience pool adds up to +35.23%, the action-step experience pool up to +53.27%, and the feedback step up to +17.74% for GPT-4o.
  • The full agent needs 13.03 LLM calls per question and costs about USD 3.06 per 100 questions with GPT-4o, showing a concrete quality-versus-cost trade-off.
  • Machine-translating non-English questions into English usually improves F1, with particularly large gains for Russian and Spanish, though GPT-4o degrades on German under translation.
  • The approach outperforms prior multilingual systems on English and on several non-English languages, and the authors report first place in the Text2SPARQL 2025 challenge.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: Because the experience pool is populated from the training subset and retrieval is by embedding similarity, the reported F1 could be inflated if a retrieved example is the same question as the test item or a near-duplicate translation; a leakage-controlled rerun would settle how much of the gain is memory versus reasoning.
  • Editorial inference: The intermediate plan, tool calls, and triplestore response are human-readable by design, so the same decomposition could serve as an explanation or debugging layer for deployed KGQA systems.
  • Editorial inference: Since no weights are fine-tuned, the architecture may transfer to other knowledge graphs or domains by rebuilding the experience pool; the paper does not test this, but it is a direct consequence of the design.
  • Editorial inference: The translation results suggest a per-language routing policy, using native processing for languages like German with strong models and translation for lower-resource or weaker models, could combine the best of both settings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes mKGQAgent, a multilingual Text-to-SPARQL system built as an LLM agent with a plan step, an action step featuring a named-entity-linking tool, a feedback step that executes intermediate SPARQL on a triplestore, and a non-parametric experience pool used for in-context learning of plans and queries. The system is evaluated on QALD-9-plus with ten languages using four backbone LLMs (GPT-3.5, GPT-4o, Qwen2.5-72B, Llama3.1-70B), and the reported results include comparisons with previously published KGQA baselines, an ablation study over components, a cost analysis, and experiments on machine-translating non-English questions into English. The abstract and conclusion claim first place in the Text2SPARQL challenge 2025 and state-of-the-art performance, with an English F1 score of 54.83%.

Significance. If the reported results hold, the paper offers a practically useful demonstration that an agentic decomposition with an experience pool can improve multilingual KGQA without supervised fine-tuning. The systematic ablations, the inclusion of open-source and proprietary models, and the token/cost analysis are valuable contributions that go beyond a single benchmark score. The paper also makes its code publicly available. However, the headline claims of first place in the Text2SPARQL challenge and state-of-the-art performance are currently not supported by evidence inside the manuscript, and a retrieval-memorization confound threatens the causal interpretation of the reported gains.

major comments (4)
  1. [Abstract, Section 1, Section 5.1] The paper repeatedly claims that mKGQAgent 'took first place among the other participants' in the Text2SPARQL challenge 2025, but no results table from that challenge appears anywhere in the manuscript. Section 5.1 only compares against previously published baselines quoted from an external leaderboard. Since the first-place claim is one of the central selling points, the official challenge results (or a clear statement that the claim is based on the challenge organizers' report) must be included, or else the claim should be tempered to 'competitive with published baselines.'
  2. [Section 3.2.1, Section 3.2.2, Algorithms 5 and 6] The experience pool is built from the training subset and stores each training question, its vector embedding, and its ground-truth SPARQL query (Algorithm 4). In the evaluation phase, Algorithms 5 and 6 retrieve the top-N most similar plans and queries by embedding similarity to the test question and inject them into the prompts. QALD-9-plus is a multilingual parallel benchmark in which the same underlying factoid question appears in multiple languages; if the train/test split is performed per language rather than per underlying question, or if near-duplicate surface forms cross the split, a retrieved example can directly supply the gold SPARQL query for the test question. The paper does not report the split granularity, an overlap analysis, the retrieval similarity distribution, or any exclusion of self/near-self matches. Without such an audit, the 54.83% result cannot be attributed to the agent's reasoning rather than to retrieval memorization. Please conduct and report an overlap analysis and, if needed, filter out retrieved examples that are translations or near-duplicates of the test question.
  3. [Section 4.3, Figure 4, Table 1] The comparison against baselines is based on numbers reused from the KGQA leaderboard rather than rerun under the same evaluation conditions, and no significance tests or confidence intervals are reported. The headline margin over HQA (GPT-4) is 4.83 percentage points (54.83 vs 50.00); without variance estimates or multiple runs, this difference may not be statistically reliable. The paper should either rerun the baselines in a controlled setting, report per-question variance and significance tests, or explicitly state the comparison is indicative only.
  4. [Table 2] In Table 2, the machine-translation rows contain identical F1 scores for different source languages: mKGQAgent (GPT-4o) reports 35.66 for both German MT and Russian MT, and mKGQAgent (Llama 3.1 70B) reports 17.29 for both. Given that the native-language scores for German and Russian differ substantially (e.g., 43.08 vs 31.67 for GPT-4o), identical MT results are implausible and suggest a copy/paste or spreadsheet error. These entries must be verified and corrected, as they directly support the claim in Section 5.3 that machine translation generally improves quality.
minor comments (5)
  1. [Section 5.2] The sentence about Lithuanian inverts the presentation order used for Russian and Ukrainian: it reads 'Lithuanian (25.54 vs 31.15)' while the previous clauses quote the outperforming system first. The numbers themselves match Table 1, but the order should be made consistent for readability.
  2. [Section 5.1, Table 4] The ablation study reports only single F1 values per configuration, with no indication of run-to-run variance or statistical significance. The claim of 'synergistic effects' from combining all components would be stronger with error bars or repeated evaluations.
  3. [Section 4.2] The implementation section mentions that SPARQL queries are executed on the 'official Wikidata SPARQL endpoint,' while the benchmarks in the challenge include DBpedia-based questions. Please clarify which endpoint(s) are used for each evaluation set.
  4. [Section 1 and Section 4] The abstract and introduction say the evaluation covers 'ten languages, including two classified as endangered,' but the two endangered languages are never explicitly identified. Please state which languages these are.
  5. [Figure 4] The bar chart is visually dense and the ordering of bars on the y-axis is nontrivial; adding numeric labels to each bar would improve legibility and match the F1 values reported in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the experience pool is populated only from the training subset with ground-truth supervision, and the reported F1 scores are measured against an external benchmark rather than equated to any input by construction.

full rationale

The paper's derivation chain is self-contained. The only learned or stored component, the experience pool, is built exclusively from the training subset with external supervision: Algorithm 4 unpacks a training example (q_i, phi_i), computes F1_i against the ground-truth SPARQL query phi_i, and stores {q_i, v_qi, phi_i, phi_hat_i, p_i, H, F1_i}. During evaluation, Algorithms 5 and 6 embed the test question and retrieve top-N examples from that pool, injecting them into prompts; but the retrieved ground-truth queries are inputs supplied by the training set, not predictions derived from the test set. No equation in the paper equates the reported test F1 values (e.g., 54.83 for GPT-4o) with any fitted parameter or with the F1 scores stored in the experience pool. The ablations measure quality against the same external QALD-9-plus ground truth rather than imposing the result by construction. The authors do cite their own QALD-9-plus dataset and KGQA leaderboard for benchmark and baseline numbers, but these citations supply externally defined test questions and independently reported comparison scores; they are not the load-bearing argument for the method's effectiveness. The residual concern that near-duplicate multilingual questions could cross the train/test split and leak gold queries through embedding retrieval is a data-integrity or leakage risk, not a circularity reduction: nothing in the paper's stated construction forces the evaluation result to equal a training input.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No fitted parameters in the mathematical sense; the ledger lists manual hyperparameters and assumptions about external benchmarks and services. These are typical for a systems paper, but they mean the reported numbers are conditional on many unreported choices.

free parameters (5)
  • top-N similar plans/queries retrieved from experience pool = not reported
    The number of examples injected into the planning and action prompts is a manually chosen hyperparameter; the paper does not specify its value or sensitivity.
  • experience pool success threshold = F1 = 1.0
    Examples are stored as successful only if the generated query achieves a Macro F1 of exactly 1.0 against the gold query; this threshold is chosen arbitrarily.
  • text embedding model for retrieval = multilingual-e5-large
    The embedding model determines which training examples are retrieved; selected based on MTEB rank rather than task-specific tuning.
  • sampling temperature = 0
    Set to zero for all LLM calls to make outputs deterministic; a common but hand-picked setting.
  • context window size = 16384 tokens
    Set for open-source models to avoid out-of-memory errors; may truncate long prompts and affect results.
assumptions (4)
  • domain assumption Macro F1 computed on the answer sets of the generated and gold SPARQL queries is a valid measure of query correctness.
    This is the benchmark standard (Usbeck et al. [29]), but it ignores query structure and can reward queries that happen to return the same answers for the wrong reasons.
  • domain assumption Wikidata SPARQL endpoint, Wikidata entity lookup, and Falcon 2.0 are reliable and available during evaluation.
    The feedback step and NEL tool depend on these external services; any change in them changes the reported results.
  • domain assumption QALD-9-plus training and test splits are disjoint at the question-ID level and no test question appears in the experience pool in any language.
    The paper does not explicitly confirm this, but the experience pool is built from the training subset; if violated, retrieval would leak answers.
  • domain assumption Machine translation with OPUS-MT preserves the information need of the original question.
    The MT comparison in Section 4.4 assumes that translated English questions are faithful to the native-language intent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Text-to-SPARQL Goes Beyond English: Multilingual Question Answering Over Knowledge Graphs through Human-Inspired Reasoning." pith.science (2026). https://pith.science/paper/Y7GNLV6Z

@misc{pith2026250716971,
  author       = {Pith},
  title        = {Pith review of: Text-to-SPARQL Goes Beyond English: Multilingual Question Answering Over Knowledge Graphs through Human-Inspired Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y7GNLV6Z}},
  note         = {Machine review of arXiv:2507.16971}
}
read the original abstract

Accessing knowledge via multilingual natural-language interfaces is one of the emerging challenges in the field of information retrieval and related ones. Structured knowledge stored in knowledge graphs can be queried via a specific query language (e.g., SPARQL). Therefore, one needs to transform natural-language input into a query to fulfill an information need. Prior approaches mostly focused on combining components (e.g., rule-based or neural-based) that solve downstream tasks and come up with an answer at the end. We introduce mKGQAgent, a human-inspired framework that breaks down the task of converting natural language questions into SPARQL queries into modular, interpretable subtasks. By leveraging a coordinated LLM agent workflow for planning, entity linking, and query refinement - guided by an experience pool for in-context learning - mKGQAgent efficiently handles multilingual KGQA. Evaluated on the DBpedia- and Corporate-based KGQA benchmarks within the Text2SPARQL challenge 2025, our approach took first place among the other participants. This work opens new avenues for developing human-like reasoning systems in multilingual semantic parsing.

Figures

Figures reproduced from arXiv: 2507.16971 by the authors.

Figure 1
Figure 1. mKGQAgent workflow demonstration (online phase). In the evaluation phase, the mKGQAgent is using the experience pool examples to improve planning, the in-context learning training examples to improve SPARQL query generation awareness and the feedback to correct possible errors. The offline phase, which is required for gathering experience pool. The evaluation or online phase – the routing of the mKGQAgent’s componen… view at source ↗
Figure 2
Figure 2. 𝒮𝒜𝑔𝑒𝑛𝑡 workflow demonstration (offline phase). In the offline phase, the 𝒮𝒜𝑔𝑒𝑛𝑡 is used to gather the experience pool of positive and negative answers over the train subsets. 𝒮𝒜𝑔𝑒𝑛𝑡 uses the plan step (cf. Section 3.1.2) to generate a structured step-by-step plan and the action step, that either calls the LLM or the named entity linking (NEL) tool (cf. Section 3.1.1) ultimately leading to the SPARQL query generation… view at source ↗
Figure 3
Figure 3. The English versions of the prompts used within the mKGQAgent. Placeholders are color-coded blue. Comments are color-coded red. Important: We use language-specific prompts for every considered language. In case of the evaluation within Text2SPARQL challenge, we used English prompts only despite the input language. 4.3. Baselines To compare the performance of the mKGQAgent we select both “pre-LLM era” KGQA systems an… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison between our mKGQAgent approach (teal) and the baselines (grey) on English questions of QALD-9-plus [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. SAGA: Schema-Aware Grounding for Agentic Text-to-SPARQL Generation

    cs.AI 2026-07 conditional novelty 6.0 of 10

    Schema-aware property filtering during interactive KBQA grounding improves answer F1 on nine benchmarks and reduces empty results.

  2. MARS: Multi-hop Adaptive Retrieval and SPARQL Generation for KGQA

    cs.CL 2026-07 conditional novelty 5.0 of 10

    MARS answers multi-hop knowledge-graph questions by iteratively retrieving ranked triple patterns and letting an LLM decide when to emit a SPARQL query, beating agentic baselines on QALD-10 without fine-tuning.

  3. LLMs+Graphs: Toward Graph-Native, Synergistic AI Systems

    cs.DB 2026-06 unverdicted novelty 2.0 of 10

    The paper synthesizes three synergies between LLMs and graphs—augmented retrieval/reasoning, bidirectional KG integration, and graph-enhanced agents—plus LLM uses in graph data management and ML.

Reference graph

Works this paper leans on

36 extracted references · 23 canonical work pages · cited by 3 Pith papers

  1. [1]

    Diefenbach, P

    D. Diefenbach, P. H. Migliatti, O. Qawasmeh, V. Lully, K. Singh, P. Maret, QAnswer: A question answering prototype bridging the gap between a considerable part of the lod cloud and end-users, in: The World Wide Web Conference, WWW ’19, Association for Computing Machinery, New York, NY, USA, 2019, p. 3507–3510. doi:10.1145/3308558.3314124

    arXiv 2019

  2. [2]

    Turganbay, V

    R. Turganbay, V. Surkov, D. Evseev, M. Drobyshevskiy, Generative question answering sys- tems over knowledge graphs and text, volume 22, 2023, pp. 1112–1126. doi: 10.28995/ 2075-7182-2023-22-1112-1126

    work page 2023

  3. [3]

    MST5 -- Multilingual Question Answering over Knowledge Graphs

    N. Srivastava, M. Ma, D. Vollmers, H. Zahera, D. Moussallem, A.-C. N. Ngomo, MST5–multilingual question answering over knowledge graphs, arXiv preprint arXiv:2407.06041 (2024)

    work page Pith review arXiv 2024

  4. [4]

    Jiang, K

    J. Jiang, K. Zhou, W. X. Zhao, Y. Song, C. Zhu, H. Zhu, J.-R. Wen, KG-Agent: An efficient autonomous agent framework for complex reasoning over knowledge graph, arXiv preprint arXiv:2402.11163 (2024). 15https://www.itzbund.de/

    arXiv 2024

  5. [5]

    QueryAgent: A Reliable and Efficient Reasoning Framework with Environmental Feedback-based Self-Correction

    X. Huang, S. Cheng, S. Huang, J. Shen, Y. Xu, C. Zhang, Y. Qu, QueryAgent: A reliable and efficient reasoning framework with environmental feedback based self-correction, arXiv preprint arXiv:2403.11886 (2024)

    work page Pith review arXiv 2024

  6. [6]

    Y. Li, Y. Zhang, L. Sun, MetaAgents: Simulating interactions of human behaviors for LLM-based task-oriented coordination via collaborative generative agents, CoRR abs/2310.06500 (2023). doi:10.48550/ARXIV.2310.06500. arXiv:2310.06500

    arXiv doi:10.48550/arxiv.2310.06500 2023

  7. [7]

    Diefenbach, K

    D. Diefenbach, K. Singh, A. Both, D. Cherix, C. Lange, S. Auer, The Qanary ecosystem: Getting new insights by composing question answering pipelines, in: J. Cabot, R. De Virgilio, R. Torlone (Eds.), Web Engineering, Springer International Publishing, Cham, 2017, pp. 171–189

    work page 2017

  8. [8]

    C. G. Correa, M. K. Ho, F. Callaway, T. L. Griffiths, Resource-rational task decomposition to minimize planning costs, in: Proceedings of the 42th Annual Meeting of the Cognitive Science Society - Developing a Mind: Learning in Humans, Animals, and Machines, CogSci 2020, virtual, July 29 - August 1, 2020, cognitivesciencesociety.org, 2020. URL: https://co...

    work page 2020

Show all 36 references
  1. [9]

    Perevalov, D

    A. Perevalov, D. Diefenbach, R. Usbeck, A. Both, QALD-9-plus: A multilingual dataset for ques- tion answering over DBpedia and Wikidata translated by native speakers, in: 2022 IEEE 16th International Conference on Semantic Computing (ICSC), IEEE, 2022, pp. 229–234

    2022

  2. [10]

    Perevalov, X

    A. Perevalov, X. Yan, L. Kovriguina, L. Jiang, A. Both, R. Usbeck, Knowledge graph question answering leaderboard: A community resource to prevent a replication crisis, in: Proceedings of the Language Resources and Evaluation Conference, European Language Resources Association...

    2022

  3. [11]

    Perevalov, A

    A. Perevalov, A. Both, A.-C. Ngonga Ngomo, Multilingual question answering systems for knowledge graphs—a survey, Semantic Web 15 (2024) 2089–2124

    2024

  4. [12]

    Punjani, K

    D. Punjani, K. Singh, A. Both, M. Koubarakis, I. Angelidis, K. Bereta, T. Beris, D. Bilidas, T. Ioannidis, N. Karalis, C. Lange, D. Pantazi, C. Papaloukas, G. Stamoulis, Template-based question answering over linked geospatial data, in: Proceedings of the 12th Workshop on Geog...

    2018

  5. [13]

    Pellissier Tanon, M

    T. Pellissier Tanon, M. D. de Assunção, E. Caron, F. M. Suchanek, Demoing Platypus – a multilingual question answering platform for Wikidata, in: A. Gangemi, A. L. Gentile, A. G. Nuzzolese, S. Rudolph, M. Maleshkova, H. Paulheim, J. Z. Pan, M. Alam (Eds.), The Semantic Web: ES...

    2018

  6. [14]

    R. Omar, I. Dhall, P. Kalnis, E. Mansour, A universal question-answering platform for knowledge graphs, Proceedings of the ACM on Management of Data 1 (2023) 1–25

    2023

  7. [15]

    Y. Zhou, X. Geng, T. Shen, W. Zhang, D. Jiang, Improving zero-shot cross-lingual transfer for multilingual question answering over knowledge graph, in: K. Toutanova, A. Rumshisky, L. Zettlemoyer, D. Hakkani-Tur, I. Beltagy, S. Bethard, R. Cotterell, T. Chakraborty, Y. Zhou (Ed...

    2021

  8. [16]

    Zhang, J

    Y. Zhang, J. Wang, Z. Wang, R. Zhang, XSemPLR: Cross-lingual semantic parsing in multiple natural languages and meaning representations, in: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computa...

    2023

  9. [17]

    Y. Tan, X. Zhang, Y. Chen, Z. Ali, Y. Hua, G. Qi, CLRN: A reasoning network for multi-relation question answering over cross-lingual knowledge graphs, Expert Systems with Applications 231 (2023) 120721. URL: https://www.sciencedirect.com/science/article/pii/S095741742301223X. ...

    2023

  10. [18]

    C. Zong, Y. Yan, W. Lu, J. Shao, Y. Huang, H. Chang, Y. Zhuang, Triad: A framework leveraging a multi-role LLM-based agent to solve knowledge base question answering, in: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024, pp. 1698–1710

    2024

  11. [19]

    Lehmann, D

    J. Lehmann, D. Bhandiwad, P. Gattogi, S. Vahdati, Beyond boundaries: A human-like approach for question answering over structured and unstructured information sources, Transactions of the Association for Computational Linguistics 12 (2024) 786–802

    2024

  12. [20]

    Xiong, J

    G. Xiong, J. Bao, W. Zhao, Interactive-KBQA: Multi-turn interactions for knowledge base question answering with large language models, CoRR abs/2402.15131 (2024). doi:10.48550/ARXIV.2402. 15131

    2024 doi arXiv

  13. [21]

    Mialon, R

    G. Mialon, R. Dessi, M. Lomeli, C. Nalmpantis, R. Pasunuru, R. Raileanu, B. Roziere, T. Schick, J. Dwivedi-Yu, A. Celikyilmaz, et al., Augmented language models: a survey, Transactions on Machine Learning Research (2023)

    2023

  14. [22]

    L. Wang, C. Ma, X. Feng, Z. Zhang, H. Yang, J. Zhang, Z. Chen, J. Tang, X. Chen, Y. Lin, et al., A survey on large language model based autonomous agents, Frontiers of Computer Science 18 (2024) 186345. doi:10.1007/s11704-024-40231-1

    2024 doi

  15. [23]

    Y. Luo, Z. Yang, F. Meng, Y. Li, J. Zhou, Y. Zhang, An empirical study of catastrophic forgetting in large language models during continual fine-tuning, arXiv preprint arXiv:2308.08747 (2023)

    2023 arXiv

  16. [25]

    Q. J. Huys, N. Lally, P. Faulkner, N. Eshel, E. Seifritz, S. J. Gershman, P. Dayan, J. P. Roiser, Interplay of approximate planning strategies, Proceedings of the National Academy of Sciences 112 (2015) 3098–3103. doi:10.1073/pnas.1414219112

    2015 doi

  17. [26]

    S. Auer, C. Bizer, G. Kobilarov, J. Lehmann, R. Cyganiak, Z. Ives, DBpedia: A nucleus for a web of open data, in: The semantic web, Springer, 2007, pp. 722–735

    2007

  18. [27]

    Vrandečić, M

    D. Vrandečić, M. Krötzsch, Wikidata: A free collaborative knowledgebase, Commun. ACM 57 (2014) 78–85. doi:10.1145/2629489

    2014 doi

  19. [28]

    Soruco, D

    J. Soruco, D. Collarana, A. Both, R. Usbeck, QALD-9-ES: A Spanish dataset for question answering systems, in: Knowledge Graphs: Semantics, Machine Learning, and Languages, IOS Press, 2023, pp. 38–52

    2023

  20. [29]

    Usbeck, M

    R. Usbeck, M. Röder, M. Hoffmann, F. Conrads, J. Huthmann, A.-C. Ngonga-Ngomo, C. Demmler, C. Unger, Benchmarking question answering systems, Semantic Web 10 (2019) 293–304

    2019

  21. [30]

    J. Lin, J. Tang, H. Tang, S. Yang, W.-M. Chen, W.-C. Wang, G. Xiao, X. Dang, C. Gan, S. Han, AWQ: Activation-aware weight quantization for on-device LLM compression and acceleration, in: P. Gibbons, G. Pekhimenko, C. D. Sa (Eds.), Proceedings of Machine Learning and Sys- tems,...

    2024

  22. [31]

    W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, I. Stoica, Efficient memory management for large language model serving with pagedattention, in: Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023

    2023

  23. [32]

    L. Wang, N. Yang, X. Huang, L. Yang, R. Majumder, F. Wei, Multilingual E5 text embeddings: A technical report, arXiv preprint arXiv:2402.05672 (2024)

    2024 arXiv

  24. [33]

    Muennighoff, N

    N. Muennighoff, N. Tazi, L. Magne, N. Reimers, MTEB: Massive text embedding benchmark, arXiv preprint arXiv:2210.07316 (2022)

    2022 arXiv

  25. [34]

    Sakor, K

    A. Sakor, K. Singh, A. Patel, M.-E. Vidal, Falcon 2.0: An entity and relation linking tool over Wikidata, in: Proceedings of the 29th ACM International Conference on Information & Knowledge Management, CIKM ’20, Association for Computing Machinery, New York, NY, USA, 2020, p. ...

    2020

  26. [35]

    Perevalov, A

    A. Perevalov, A. Both, D. Diefenbach, A.-C. Ngonga Ngomo, Can machine translation be a reasonable alternative for multilingual question answering systems over knowledge graphs?, in: Proceedings of the ACM Web Conference 2022, WWW ’22, Association for Computing Machinery, New Y...

    2022

  27. [36]

    Srivastava, A

    N. Srivastava, A. Perevalov, D. Kuchelev, D. Moussallem, A.-C. Ngonga Ngomo, A. Both, Lingua franca – entity-aware machine translation approach for question answering over knowledge graphs, in: Proceedings of the 12th Knowledge Capture Conference 2023, K-CAP ’23, Association f...

    2023

  28. [37]

    Tiedemann, S

    J. Tiedemann, S. Thottingal, OPUS-MT — Building open translation services for the World, in: Proceedings of the 22nd Annual Conference of the European Association for Machine Translation (EAMT), Lisbon, Portugal, 2020

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.