Pith. sign in

REVIEW 5 major objections 6 minor 25 references

The paper claims that a category-guided hybrid of dense retrieval and knowledge-graph reasoning—CatRAG—raises a regulation-answering chatbot to 98.89% F1, beating plain RAG.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 12:48 UTC pith:KEBOP3QD

load-bearing objection CatRAG is a sensible hybrid design and the new Vietnamese regulation dataset is a real asset, but the headline 98.89% F1 is likely measuring in/out-of-domain scope, not answer quality. the 5 major comments →

arxiv 2510.01800 v3 pith:KEBOP3QD submitted 2025-10-02 cs.AI

REBot: From RAG to CatRAG with Semantic Enrichment and Graph Routing

Thanh Ma , Tri-Tam La , Lam-Thu Le Huu , Minh-Nghi Nguyen , Khanh-Van Pham Luu This is my paper
classification cs.AI
keywords retrieval-augmented generationknowledge graphquestion answeringchatbotintent classificationnamed entity recognitionacademic regulationsgraph routing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that a university regulation-advising chatbot can answer more accurately by combining two retrieval paths: ordinary dense retrieval over document chunks, and symbolic retrieval over a knowledge graph whose nodes and edges are organised under categories predicted by a lightweight classifier. This hybrid, CatRAG, routes each query to the matching category subgraph, retrieves entities and relations from it, and merges that evidence with top vector matches before an LLM synthesises the answer. On a curated set of 1,313 regulation question–answer pairs, the paper reports that CatRAG reaches 98.89% F1 with one compact language model at a similarity threshold of 0.6, compared with 98.67% for plain RAG, and stays above 90% F1 at stricter thresholds. The intended contribution is a practical recipe for 'category-guided GraphRAG': keep RAG's factual grounding, add graph structure for context and explainability, and use classification to stop irrelevant domains from leaking in. The headline number should be read with caution, because the evaluation labels only whether a query is inside or outside the regulation domain, and the paper does not describe how precision and recall are computed from generated answers.

Core claim

The central claim is that adding a category-guided graph branch to standard RAG improves answer quality for domain-specific questions. The knowledge graph is built in two passes: chunks are embedded and stored in a vector index, NER extracts entities that are linked to chunks, an LLM extracts typed relations, and a classifier attaches each chunk to one of five regulation subgraphs. At query time the same classifier labels the question; the system then retrieves top-k chunks from the global vector store and from the matching subgraph, expands the graph chunks to entities and relations, and passes the combined context to an LLM. The paper reports that this construction yields 98.89% F1 with a

What carries the argument

The load-bearing component is the CatRAG-Query algorithm, supported by the Regulation-Enriched Graph (REG), a labelled graph G=(E,R) whose entities come from NER and whose typed relations come from LLM-based extraction, partitioned into category subgraphs by the classifier. The query algorithm embeds and classifies the question, retrieves top-k chunks from both the vector store and the predicted subgraph, expands those chunks to entities and relations, and feeds the unified context to the LLM. Graph retrieval is the differentiator: it supplies surrounding structured evidence that dense retrieval alone does not provide, while the category partition keeps that evidence focused on the right dom

Load-bearing premise

The load-bearing assumption is that the F1 score of 98.89% actually measures answer quality; the evaluation dataset only marks each query as belonging to the regulation domain or not, and the paper gives no protocol for scoring generated answers, so if F1 simply reflects in-domain versus out-of-domain discrimination, the 'accurate answers' claim is unsupported.

What would settle it

Regenerate answers for a sample of the 1,313 evaluation pairs, have human judges or gold answers grade correctness independent of query scope, and recompute precision/recall/F1; if the score drops to at or below the plain-RAG variant, the category-guided graph claim is falsified. A more refined check: take in-domain queries and deliberately corrupt the graph subgraph or classifier label; if answer quality does not degrade, the graph branch is not carrying the reported benefit.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • For one of the two tested language models, CatRAG outperforms plain RAG at every threshold from 0.6 to 0.8, with the largest margin at 0.8 (F1 91.05 vs 89.19) and the highest absolute F1 of 98.89 at 0.6.
  • Adding graph-based retrieval costs roughly two seconds per query (about 7.3 s vs 5.1 s for one model), which the paper still calls acceptable for real-time advising.
  • A lightweight classifier suffices for routing: the paper's best classifier reaches about 95.7% F1, and that accuracy is enough to yield the reported CatRAG gains.
  • The framework relies on compact language models rather than large ones, so the same category-guided recipe could be reproduced in other low-resource domains.
  • The paper's own qualitative review says answers are 'largely accurate and contextually enriched,' positioning CatRAG as a practical alternative to pure RAG for institutions with fragmented policy documents.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the reported advantage of CatRAG over RAG is model-dependent on the paper's own numbers—for the second model at threshold 0.6, plain RAG actually scores higher (98.95 F1 vs 98.73), so the headline 'CatRAG consistently outperforms RAG' would not survive across every model/threshold combination.
  • Editorial inference: because the evaluation dataset labels queries only as in-domain or out-of-domain, the 98.89% F1 may be measuring retrieval routing ability rather than answer correctness; a direct test would regrade outputs against gold answers or a human rubric before claiming factual superiority.
  • Editorial inference: the category-routing idea transfers naturally to other fragmented, policy-heavy domains (health, legal, financial), where a small taxonomy is known and documents are unstructured; a testable extension is replacing the fixed five-category schema with a learned hierarchy.
  • Editorial inference: the paper's own listed limitations—excessive context causing incoherent responses, hard long-context classification, and ~7-second latency—suggest the next bottleneck is context control, so adding reranking or selective entity expansion is the natural next experiment.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes REBot, an academic-regulation chatbot for Can Tho University, powered by CatRAG, a hybrid retrieval-augmented generation and knowledge-graph framework. CatRAG combines dense vector retrieval with a category-guided knowledge graph: documents are chunked, embedded, classified into five regulation domains, and linked to entities and relations in a Neo4j graph; at query time a fastText classifier routes the query to the corresponding subgraph, whose top chunks and entities are merged with vector retrieval results and passed to an LLM. The authors construct two Vietnamese datasets (a Q&A evaluation set and a classification training set), report classification results for fastText versus traditional baselines, compare RAG and CatRAG across thresholds and two LLMs, and report a headline F1-score of 98.89% for CatRAG with gpt-4o-mini at threshold 0.6. They also present a web application and qualitative examples.

Significance. If the evaluation were sound, CatRAG would be a useful and practical contribution: it provides a reusable recipe for combining category-guided graph retrieval with dense retrieval in a low-resource language (Vietnamese), includes a concrete knowledge-graph construction algorithm and query algorithm, offers complexity analysis for both, and releases data and code publicly. The paper also honestly discusses latency and context-bloat limitations. The architecture is not circular; the concern is whether the reported quantitative results measure what they claim. The headline number, however, rests on an evaluation protocol that is not described, and the similarity threshold is tuned on the same evaluation set. These issues are load-bearing for the central claim of state-of-the-art performance with an F1-score of 98.89%.

major comments (5)
  1. [Section 5.1, Table 1 and Section 5.2, Table 2] The F1 metric is not interpretable as an answer-quality measure. The Q&A dataset is labeled only as 'Truth - Knowledge related to CTU academic regulation' (909 samples) versus 'Other - Knowledge outside of CTU academic regulation' (404 samples). No gold answers, human rubric, LLM-judge procedure, exact-match criterion, or any other grading protocol is given for computing Precision, Recall, and F1 in Table 2. As written, the reported 98.89% F1 could simply reflect in-domain/out-of-domain discrimination of the response (i.e., whether the model's answer stays within the regulation scope), not whether the answer is factually correct or complete. This invalidates the headline claim as stated. The authors must specify the exact evaluation protocol, provide the grading rubric or judge prompts, and ideally report per-question accuracy on the 909 in-domain items separately.
  2. [Section 5.2, grid search over threshold] The similarity threshold is fitted on the same evaluation set used to report the headline result. The text says 'we conduct a gridsearch from 0.1 to 1.0. Then, [0.6,0.8] obtained the high accuracy,' and the abstract and Table 2 report the best threshold (0.6) as the system's performance. This is a selection artifact: with 10 thresholds explored, the maximum F1 is expected to be optimistic relative to out-of-sample performance. The paper should use a validation split for threshold selection or report the full threshold sweep with a clear train/validation/test separation. The statement in Section 5.2 that 'a similarity threshold of 0.7 is applied' also conflicts with the headline 0.6 result; please clarify which threshold is actually used in the deployed system.
  3. [Table 1 and Section 5.1] The dataset sizes are internally inconsistent: the text reports 1,319 Q&A pairs and 3,256 classification questions, while Table 1 totals 1,313 and 3,252 respectively. The arithmetic in Table 1 is correct for its rows (474+611+655+477+485+550 = 3,252; 909+404 = 1,313), so the discrepancy is in the prose or in the table labels. This matters because the classification and evaluation sets are the only empirical basis for the paper's claims; the authors should correct the numbers and report the exact split used for training and testing the classifier.
  4. [Section 5.2, Table 2 discussion] The narrative claim that CatRAG 'consistently outperforms its RAG counterpart across all threshold settings' is contradicted by the Mistral row at threshold 0.6, where RAG achieves F1 98.95 vs CatRAG 98.73, and the Mistral row at threshold 0.7 has nearly equal recall. Please rephrase the claim to describe the actual pattern, and report whether the differences are statistically significant under any reasonable test (e.g., paired bootstrap or McNemar). Without significance testing, the 0.22-point F1 gain for gpt-4o-mini at 0.6 may be noise.
  5. [Section 5.2 / Conclusion] The paper calls the result 'state-of-the-art performance' but provides no external baselines from the literature, no comparison with prior academic-advising chatbots, and no baseline beyond the paper's own RAG variant. Given the evaluation metric is undefined (see first comment), the SOTA claim is unsupported. At minimum, the authors should compare against a simple fastText-routing plus vector-retrieval baseline, and against a standard GraphRAG baseline, with the same evaluation protocol.
minor comments (6)
  1. [Section 4, Algorithms 1 and 2] The algorithm listings contain LaTeX artifacts: 'v_q leftarrow f_emb(q)' and similar arrows are typeset with a leading minus sign, making the code difficult to read. Please fix the arrow commands (e.g., leftarrow without the hyphen).
  2. [Section 3 / Section 5.3] The framework description says NER and relation extraction use Underthesea, but the paper does not specify which NER model or relation-extraction prompt is used, nor the accuracy of the NER step. A short description or reference would help reproducibility.
  3. [Table 2 / Table 3] The caption for Table 2 lists 'Eval. Time' with only two cells filled; other rows show '–'. Please either report the time for all rows or remove the column and describe the timing measurement separately. Table 3 reports 'average of 10 runs' but does not state the standard deviation or the number of queries used; please add this.
  4. [Section 5.3] The text says 'we distilled four key questions' but only two examples are shown. Either show all four or rephrase to 'two illustrative examples'.
  5. [Abstract and Section 2.5] The abstract states that CatRAG 'integrates RAG with GraphRAG' but elsewhere GraphRAG is described as a technique that 'often underperforms standard RAG' [7]. The framing is mildly inconsistent; please clarify whether the paper's contribution is a new variant of GraphRAG or a union of RAG and GraphRAG.
  6. [Throughout] There are several typographical issues: 'T ogether' (Section 4), inconsistent comma use around 'e.g.,', and the repeated footnote markers 2/5 and 3/8. A careful proofreading pass would improve readability.

Circularity Check

0 steps flagged

No construction-level circularity; the reported F1 is an evaluation result with validity caveats, not a derivation from its own inputs.

full rationale

Walking the claimed derivation chain: CatRAG-Construct (Algorithm 1) and CatRAG-Query (Algorithm 2) are constructive algorithms whose outputs—a vector store/graph and a generated response—depend on inputs such as chunks, an embedding function, a classifier, a knowledge graph, and an LLM. No output quantity is defined in terms of the headline F1, and the headline F1 is not fed back into the algorithms. The F1=98.89 in Table 2 is an evaluation result of one configuration (threshold 0.6) and is not a fitted prediction derived from the model equations. The threshold grid search reported in Section 5.2 ('we conduct a gridsearch from 0.1 to 1.0. Then, [0.6,0.8] obtained the high accuracy') is a model-selection/validity concern, not a circularity: it does not make the architecture's derivation equivalent to its inputs. Similarly, the absence of a grading protocol for the Table 1 labels ('Truth 909 — Knowledge related to CTU academic regulation; Other 404 — Knowledge outside of CTU academic regulation') is a benchmark-validity issue, not a circular reduction. The only self-citation, [15], appears in the introduction as background context ('an LLM-powered chatbot that offers CTU students [15] fast, accurate, and up-to-date answers') and is not load-bearing: no uniqueness theorem or ansatz is imported from it. The paper also reports honest limitations in Section 5.4, including excessive graph context and latency, and compares CatRAG against its own plain-RAG baseline. Because the central architecture and its reported comparison are independently grounded in the algorithms and measurements, no specific circular step can be exhibited, and the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central evaluation rests on a proprietary expert-labeled dataset, a hand-built KG, and several hand-chosen retrieval constants (τ=0.6, k_vec=5, chunk size). The threshold τ was selected by grid search on the same data used to report the headline F1, which is the main free-parameter concern. No new physical or theoretical entities are postulated; the REG graph and category labels are engineering schema, not invented entities with independent falsifiable handles.

free parameters (5)
  • KG similarity threshold τ = 0.6 (best of grid search over [0.1, 1.0])
    Chosen to maximize F1 on the evaluation set; the headline 98.89% figure depends on this post hoc choice. No validation split is described (Section 5.2).
  • k_vec (top vector chunks) = 5
    Stated in Section 5.2: 'retrieves the top five most relevant chunks'; chosen by hand, no sensitivity analysis.
  • k_graph (top graph chunks) = not specified
    TopKChunks depth in Algorithm 2 is never given a value; it affects the CatRAG context that produces the reported F1.
  • fastText hyperparameters = epochs=100, ngrams=3
    Training choices reported in Table 4; standard but tuned, and the router's errors propagate into the QA pipeline.
  • chunk size = not specified
    Chunking granularity from Docling output is never stated; retrieval quality and the construct cost O(M·d) depend on it (Section 4.1).
axioms (5)
  • domain assumption The expert-authored Q&A dataset (labels 'Truth 909' / 'Other 404') and its gold answers are correct and complete.
    Section 5.1 states data was 'manually verified' and answers authored by CTU academic support experts; the entire F1 evaluation rests on this unverifiable ground truth.
  • domain assumption The 250-node knowledge graph extracted from CTU PDFs via Docling, Underthesea NER, and LLM relation extraction faithfully represents the regulations.
    Section 3 and Figure 1; no accuracy audit of the KG is reported, yet CatRAG's graph route relies on it.
  • domain assumption PhoBERTv2 embeddings with cosine similarity at threshold 0.6 yield semantically valid matches for Vietnamese regulation text.
    Section 5.2; no ablation over embedding models or similarity measures.
  • domain assumption The fastText classifier (95.67% F1) routes queries to the correct subgraph often enough that routing errors do not materially degrade the QA result.
    Section 5.2 Table 4; misrouted queries receive context from the wrong subgraph, and this failure coupling is not analyzed.
  • domain assumption gpt-4o-mini / mistral-small-2506 generate faithful answers from the retrieved context.
    Section 5.2; the paper itself notes (Section 5.4) that graph context can yield 'incoherent or inaccurate responses', so faithfulness is not guaranteed.

pith-pipeline@v1.3.0-alltime-deepseek · 8574 in / 20138 out tokens · 137925 ms · 2026-08-04T12:48:31.416772+00:00 · methodology

0 comments
read the original abstract

Academic regulation advising is essential for helping students interpret and comply with institutional policies, yet building effective systems requires domain specific regulatory resources. To address this challenge, we propose REBot, an LLM enhanced advisory chatbot powered by CatRAG, a hybrid retrieval reasoning framework that integrates retrieval augmented generation with graph based reasoning. CatRAG unifies dense retrieval and graph reasoning, supported by a hierarchical, category labeled knowledge graph enriched with semantic features for domain alignment. A lightweight intent classifier routes queries to the appropriate retrieval modules, ensuring both factual accuracy and contextual depth. We construct a regulation specific dataset and evaluate REBot on classification and question answering tasks, achieving state of the art performance with an F1 score of 98.89%. Finally, we implement a web application that demonstrates the practical value of REBot in real world academic advising scenarios.

Figures

Figures reproduced from arXiv: 2510.01800 by Khanh-Van Pham Luu, Lam-Thu Le Huu, Minh-Nghi Nguyen, Thanh Ma, Tri-Tam La.

Figure 1
Figure 1. Figure 1: Knowledge graph of REBot with 250 nodes In REBot, this design allows the system to provide domain-specific, context￾aware, and verifiable responses. The subgraph retrieval ensures that answers are grounded not only in relevant documents but also in the correct entity rela￾tionships, which is essential for regulation-focused dialogue. The next section explains details about our REBot framework. 3 REBot Fram… view at source ↗
Figure 2
Figure 2. Figure 2: The REBot framework cleaned, segmented, and enriched via NER and relation extraction with Under￾thesea5 , then embedded using PhoBERTv2 [19] for reliable Vietnamese represen￾tations; (2) Student Affairs Classification (Φ), in which queries are categorized into five regulation domains using fastText [25], chosen for its efficiency and ac￾curacy on short Vietnamese texts; (3) REBot Knowledge Graph, a Neo4j v… view at source ↗
Figure 3
Figure 3. Figure 3: The Interface of our REBot System 6 Conclusion This study presents the effectiveness of a retrieval-augmented chatbot system using standard Retrieval-Augmented Generation (RAG) and knowledge-graph￾enhanced GraphRAG to provide accurate and detailed responses to student ques￾tions at Can Tho University. Quantitative tests indicate that CatRAG, combined with a compact generative model, achieves top performanc… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 7 linked inside Pith

  1. [1]

    High-Confidence Computing4(2), 100211 (2024)

    A survey on large language model (llm) security and privacy: The good, the bad, and the ugly. High-Confidence Computing4(2), 100211 (2024)

    2024

  2. [2]

    Opdahl, A., Veres, C.: Named entity ex- traction for knowledge graphs: A literature overview

    Al-Moslmi, T., Gallofré Oca˜ na, M., L. Opdahl, A., Veres, C.: Named entity ex- traction for knowledge graphs: A literature overview. IEEE Access8, 32862–32881 (2020)

    2020

  3. [3]

    Journal of Big Data9(1), 10 (2022)

    Chiche, A., Yitagesu, B.: Part of speech tagging: a systematic review of deep learn- ing and machine learning approaches. Journal of Big Data9(1), 10 (2022)

    2022

  4. [4]

    Hue University Journal of Science: Techniques and Technology 131, 39–52 (06 2023) 12 Thanh Ma et al

    Doan, H., Le, V., van, K.: Xây dựng khung ứng dụng ai chatbot trong lĩnh vực quy chế đào tạo. Hue University Journal of Science: Techniques and Technology 131, 39–52 (06 2023) 12 Thanh Ma et al

    2023

  5. [5]

    arXiv preprint arXiv:2310.04560 (2023)

    Fatemi, B., Halcrow, J., Perozzi, B.: Talk like a graph: Encoding graphs for large language models. arXiv preprint arXiv:2310.04560 (2023)

    Pith/arXiv arXiv 2023

  6. [6]

    Computer Science Review29, 21–43 (2018)

    Goyal, A., Gupta, V., Kumar, M.: Recent named entity recognition and classifica- tion techniques: a systematic review. Computer Science Review29, 21–43 (2018)

    2018

  7. [7]

    graphrag: A systematic evaluation and key insights

    Han, H., Shomer, H., Wang, Y., Lei, Y., Guo, K., Hua, Z., Long, B., Liu, H., Tang, J.: Rag vs. graphrag: A systematic evaluation and key insights. arXiv preprint arXiv:2502.11371 (2025)

    arXiv 2025

  8. [8]

    arXiv preprint arXiv:2501.00309 (2024)

    Han, H., Wang, Y., Shomer, H., Guo, K., Ding, J., Lei, Y., Halappanavar, M., Rossi, R.A., Mukherjee, S., Tang, X., et al.: Retrieval-augmented generation with graphs (graphrag). arXiv preprint arXiv:2501.00309 (2024)

    Pith/arXiv arXiv 2024

  9. [9]

    ACM Com- puting Surveys54(4), 1–37 (Jul 2021)

    Hogan, A., Blomqvist, E., Cochez, M., et al, D.: Knowledge graphs. ACM Com- puting Surveys54(4), 1–37 (Jul 2021)

    2021

  10. [10]

    arXiv preprint arXiv:2405.16506 (2024)

    Hu, Y., Lei, Z., Zhang, Z., Pan, B., Ling, C., Zhao, L.: Grag: Graph retrieval- augmented generation. arXiv preprint arXiv:2405.16506 (2024)

    Pith/arXiv arXiv 2024

  11. [11]

    Proceedings of Vietnamese Speech and Language Processing (VLSP) (2016)

    Huyen, N.T.M., Luong, V.X.: Vlsp 2016 shared task: Named entity recognition. Proceedings of Vietnamese Speech and Language Processing (VLSP) (2016)

    2016

  12. [12]

    Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., K¨ uttler, H., Lewis, M., tau Yih, W., Rockt¨ aschel, T., Riedel, S., Kiela, D.: Retrieval-augmented generation for knowledge-intensive nlp tasks (2021)

    2021

  13. [13]

    ACM Transactions on Intelligent Systems and Technology (TIST)13(2), 1–41 (2022)

    Li, Q., Peng, H., Li, J., Xia, C., Yang, R., Sun, L., Yu, P.S., He, L.: A survey on text classification: From traditional to deep learning. ACM Transactions on Intelligent Systems and Technology (TIST)13(2), 1–41 (2022)

    2022

  14. [14]

    Luong, H., Luong, K.: A chatbot-based academic advising model for student in informationtechnology:Acasestudy.SaudiJournalofEngineeringandTechnology 10(3), 93–100 (2025)

    2025

  15. [15]

    In: International Conference on Intelligent Systems and Data Science

    Ma, T., Chau, T.K., Thai, P.A., Tram, T.M., Huynh, K., Tran-Nguyen, M.T.: Racos: Ai-routed chat-voice admission consulting support system. In: International Conference on Intelligent Systems and Data Science. pp. 295–310. Springer (2024)

    2024

  16. [16]

    Minh, P.Q.N.: A feature-rich vietnamese named-entity recognition model (2018), https://arxiv.org/abs/1803.04375

    Pith/arXiv arXiv 2018

  17. [17]

    Expert Systems with Applications106, 36–54 (2018)

    Miro´ nczuk, M.M., Protasiewicz, J.: A recent overview of the state-of-the-art ele- ments of text classification. Expert Systems with Applications106, 36–54 (2018)

    2018

  18. [18]

    Meta-Radiology1(2), 100022 (2023)

    Nazir, A., Wang, Z.: A comprehensive survey of chatgpt: Advancements, applica- tions, prospects, and challenges. Meta-Radiology1(2), 100022 (2023)

    2023

  19. [19]

    arXiv preprint arXiv:2003.00744 (2020)

    Nguyen, D.Q., Nguyen, A.T.: Phobert: Pre-trained language models for viet- namese. arXiv preprint arXiv:2003.00744 (2020)

    Pith/arXiv arXiv 2003

  20. [20]

    arXiv preprint arXiv:1711.04951 (2017)

    Nguyen, D.Q., Vu, T., Nguyen, D.Q., Dras, M., Johnson, M.: From word segmen- tation to pos tagging for vietnamese. arXiv preprint arXiv:1711.04951 (2017)

    Pith/arXiv arXiv 2017

  21. [21]

    IEEE Transactions on Knowledge and Data Engineering36(7), 3580–3599 (2024)

    Pan, S., Luo, L., Wang, Y., Chen, C., Wang, J., Wu, X.: Unifying large language models and knowledge graphs: A roadmap. IEEE Transactions on Knowledge and Data Engineering36(7), 3580–3599 (2024)

    2024

  22. [22]

    Pande, A., Patil, R., Mukkemwar, R., Panchal, R., Bhoite, S.: Comprehensive study of google gemini and text generating models: Understanding capabilities and performance (11 2024)

    2024

  23. [23]

    Artificial Intelligence Review56(11), 13071–13102 (2023)

    Peng, C., Xia, F., Naseriparsa, M., Osborne, F.: Knowledge graphs: Opportunities and challenges. Artificial Intelligence Review56(11), 13071–13102 (2023)

    2023

  24. [24]

    arXiv preprint arXiv:2307.07697 (2023)

    Sun, J., Xu, C., Tang, L., Wang, S., Lin, C., Gong, Y., Ni, L.M., Shum, H.Y., Guo, J.: Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph. arXiv preprint arXiv:2307.07697 (2023)

    Pith/arXiv arXiv 2023

  25. [25]

    ICAIIS’20 pp

    Yao, T., Zhai, Z., Gao, B.: Text classification model based on fasttext. ICAIIS’20 pp. 154–157 (2020)

    2020