Rhetorical-Role-Aware Retrieval-Augmented Generation for Legal Question Answering over Indian Supreme Court Judgments

Anand Kumar Madasamy; Navya Binu; Sangeetha Sivanesan; Sayed Ayaan Ahmed Sha

REVIEW 4 major objections 4 minor 21 references

Rhetorical-Role-Aware Retrieval-Augmented Generation for Legal Question Answering over Indian Supreme Court Judgments

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a legal question-answering system over Indian Supreme Court judgments is more faithful when retrieval follows the rhetorical structure of judgments—fact, precedent, ratio, ruling—than when it treats the text as a…

desk verdict Sensible pipeline, but no baseline, no ablation, and low contextual relevancy scores mean the effectiveness claim is unsupported; desk reject, though the idea is worth a resubmission. read the letter →

arxiv 2608.06828 v1 pith:6DI3NZJ3 submitted 2026-08-07 cs.IR

Sayed Ayaan Ahmed Sha , Sangeetha Sivanesan , Anand Kumar Madasamy , Navya Binu This is my paper · ORCID

classification cs.IR

keywords retrieval-augmentedgenerationlegalquestionansweringrhetoricalrolesIndianSupremeCourtjudgmentshybridretrievalcross-encoderrerankingqueryintentclassification

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

The paper claims that a retrieval-augmented question-answering system for Indian Supreme Court judgments works better when it respects the rhetorical structure of a judgment rather than treating the text as a flat collection. Each sentence is labelled with a role such as Fact, Precedent, Ratio of the Decision, or Ruling by Present Court, and chunks are built from consecutive sentences sharing a role. A query is assigned to one of five intents, and that intent decides which roles are allowed into the candidate set; dense and BM25 retrieval run in parallel, and a cross-encoder reranks the survivors before an LLM writes the answer. On 30 judgments across civil, corporate, and criminal law, the authors report faithfulness above 0.95 in every domain and answer relevancy near 0.9, which they read as evidence that role-aware retrieval supplies the context needed for grounded legal answers.

What carries the argument

The load-bearing mechanism is the rhetorical-role-aware chunking and filtering cycle. A sentence-level labeling model assigns one of seven roles—Argument, Statute, Fact, Ratio of the Decision, Precedent, Ruling by Lower Court, Ruling by Present Court—to every sentence; consecutive same-role sentences are merged into chunks, micro-chunks are folded into neighbors, and long groups are split into overlapping fifteen-sentence windows. A query is classified into one of five intents, and a hand-defined table maps each intent to the roles most likely to hold the answer (for example, inferential queries target Ratio of the Decision and Argument; precedent queries target Precedent and Statute). Retrieval then runs over a dense passage index and a BM25 sparse index in parallel, results are restricted to the allowed roles, and a cross-encoder scores each remaining query-passage pair before the top passages are handed to the generator.

What would settle it

Run the 25 queries on the same 30 judgments twice: once with the full role-filtering pipeline and once with filtering turned off, retrieving from all roles before reranking and generation. If unfiltered retrieval finds the answer-bearing passage at least as often, or if the generator's faithfulness does not drop when filtering is removed, then the intent-to-role mapping is not the cause of the reported performance; the same test could use an oracle role labeler to separate labeling error from mapping error.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the functional role a sentence plays in a judgment—fact, statute, precedent, ratio, ruling—is a useful retrieval signal for legal question answering. By grouping sentences by role into chunks and then filtering candidate passages by the roles mapped to the predicted query intent, the pipeline removes content that may be lexically or semantically similar but legally peripheral. The authors report that this role-aware pipeline, combined with hybrid dense/sparse retrieval and cross-encoder reranking, produces consistently high faithfulness (0.951–0.991 across the three domains) and answer relevancy (0.860–0.923), while contextual relevancy sits much lower, around 0.38–0.44. They interpret these numbers as showing that the retrieved context, even when it carries irrelevant material, still contains enough legally appropriate passages for the generator to stay close to the source.

Load-bearing premise

The load-bearing premise is that the model labelling each sentence with its role is accurate on these judgments and that the hand-built mapping from query intent to allowed roles never excludes the role that actually holds the answer; if either fails, the correct passage is removed before reranking no matter how well the retrievers perform.

Editorial extensions

If this is right

Role-aware chunking produces semantically coherent retrieval units, so questions about reasoning are answered from the ratio and argument sections rather than from peripheral text.
Intent classification followed by role filtering narrows the candidate set, letting the cross-encoder rerank within the parts of a judgment that match the query's information need.
Hybrid dense and BM25 retrieval covers both paraphrase-level matches and exact legal identifiers such as section numbers and case names.
The reported faithfulness scores imply that answers generated over role-filtered context are largely supported by the retrieved passages across all three legal domains tested.

Reading between the lines

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

A direct ablation—running the same queries with role filtering disabled—would show what share of the faithfulness gain comes from the intent-to-role maps; the paper does not report such a comparison.
The contextual relevancy scores are markedly lower than faithfulness, suggesting the generator tolerates noisy context; tightening the retriever or the role mapping might raise both without hurting groundedness.
The hand-defined intent-to-role mapping is a natural place to automate, and learned mappings per domain or per court could transfer the approach to lower courts, other jurisdictions, or non-English judgments.
Because the paper's own limitations section notes that scores are automated and not reviewed by legal experts, a human annotation study of the generated answers would be the clearest next validation.

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.

Desk Editor's Note

read the letter

Quick take: the paper has a reasonable idea, but the evaluation doesn't support the headline claim, and one design choice (role filtering) could systematically remove the correct answer before reranking. I'd desk reject, though the idea is worth a resubmission.

What's actually new: combining rhetorical-role chunking with query-intent-to-role filtering for QA over Indian Supreme Court judgments is a real, modest integration. The paper correctly identifies that legal judgments have internal structure and that treating them as homogeneous text is suboptimal. It cites closely related systems (TaxFlow, CEFS) and inherits a role labeler from Malik et al. That's honest positioning.

The soft spots are serious. There is no baseline, no ablation, no error bars, no human evaluation. The reported Contextual Relevancy numbers (0.376–0.443) mean that most retrieved passages are judged irrelevant even after filtering and reranking. That directly contradicts the abstract's claim of 'proven very effective.' The Section 6 limitation note admits all scores come from automated LLM judges, which is honest, but it doesn't rescue the effectiveness claim.

The stress-test concern lands: Section 3.3.2 filters candidates by the intent-to-role mapping before reranking. If the labeler makes a mistake or the mapping omits the role containing the answer, the correct passage is gone and no reranker can recover it. The paper gives no labeler accuracy on these judgments and no ablation with filtering disabled. That is load-bearing for the claimed benefit of the approach.

Minor but genuine: the role taxonomy is sloppy—Section 3.1 says 'seven rhetorical categories' but lists six, and Table 2 uses a different set including RulingByPresentCourt. This matters because the intended role filtering depends on the taxonomy. Also, the same paper is cited twice as [18] and [21], which is sloppy but easy to fix.

Who this is for: people working on Indian legal RAG might find the pipeline description useful as a starting point, but the evidence doesn't justify deploying it. I'd recommend desk rejection, and I'd tell the authors to resubmit only after adding a no-filter baseline, ablations of each component, human evaluation on a small sample, and validation of the role labels.

Referee Report

4 major / 4 minor

Summary. This paper describes a retrieval-augmented generation system for legal question answering over Indian Supreme Court judgments. The pipeline performs sentence-level rhetorical role labeling, role-aware chunking, hybrid BM25/dense retrieval, intent-based role filtering, cross-encoder reranking, and LLM answer generation. The evaluation uses 30 judgments and 25 queries across civil, corporate, and criminal domains, with automated LLM-based metrics (Faithfulness, Answer Relevancy, Contextual Relevancy) reported in Tables 5–7. The authors conclude that the framework is “very effective” for context-heavy legal QA.

Significance. If validated, the idea of exploiting rhetorical roles in legal judgments for chunking and filtering is potentially useful for domain-specific RAG. The paper makes a concrete pipeline choice and compiles a small evaluation corpus, which is a useful starting point. However, the current evidence does not establish the central effectiveness claim: there is no comparison against a standard RAG baseline, no ablation of the role-aware components, no validation of the external role labeler, and the reported Contextual Relevancy scores are low. The strengths are the clearly described architecture and the acknowledgment of limitations in Section 6, but these do not compensate for the missing experimental support.

major comments (4)

[Section 4 (Tables 5–7); Section 3.3.2] The central effectiveness claim is not supported because no baseline or ablation isolates the contribution of the proposed components. The reported Contextual Relevancy scores of 0.376–0.443 indicate that most retrieved nodes are judged irrelevant, and without a comparison to a vanilla RAG pipeline, to retrieval without role filtering, or to chunking without rhetorical roles, the high Faithfulness and Answer Relevancy values cannot be attributed to the proposed architecture. A no-filter baseline is especially important because Section 3.3.2 removes passages before reranking.
[Section 3.3.2 and Table 3] The intent-to-role mapping is hand-defined and filters candidates before reranking, so an omission or error is unrecoverable. The paper never validates the Malik et al. role labeler on these 30 judgments, reports no labeler accuracy or human agreement, and does not test whether the mapping excludes passages that contain the answer. Without such validation or a filtering-disabled ablation, the retrieval results are consistent with a system that works only when the answer happens to fall in the mapped roles.
[Section 4.1 and Section 6] All three metrics are computed by an LLM acting as an evaluation judge, and Section 6 admits there is no human evaluation. Because the same class of LLM both generates answers and judges relevance, the scores are self-referential; no evidence is given that the LLM judge agrees with human judgments or with gold labels. The low Contextual Relevancy values in Tables 5–7 further suggest that the judge is not simply lenient, and the lack of a human-validated subset makes it impossible to interpret the metric values as evidence of effectiveness.
[Section 3.1 and Table 2] The rhetorical-role taxonomy is inconsistently defined. Section 3.1 says “one of the six rhetorical roles” and then lists “seven rhetorical categories” with only six items (Argument, Statute, Fact, Ratio of the Decision, Precedent, Ruling by Lower Court), while Table 2 reports a seventh role, RulingByPresentCourt. This inconsistency matters because the role labels are the input to chunking and filtering, and the paper does not clarify which roles the labeler actually produces or how the intent-to-role mapping was derived.

minor comments (4)

[Section 3.1 and Section 3.3.2] Typos: “labled” should be “labeled”; “correspondoing” should be “corresponding”; “alloted” should be “allotted”; Table 6: “Qwen3-32B A” appears to have a stray “A”.
[References] The reference list contains the same Malik et al. work as [18] and [21]; one entry should be removed or cross-referenced.
[Section 3.2 and Section 3.3.3] The prompt design and retrieval hyperparameters (top-k, rerank size, small-group merging threshold, sub-splitting window and overlap) are not specified; they should be reported for reproducibility.
[Abstract and Conclusion] The abstract claims the framework “demonstrated strong performance” and “proves” effectiveness, but Section 6 acknowledges limitations; the wording should be modest and match the evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the framework is an empirical RAG system and its reported scores are not forced by construction.

full rationale

The paper makes no analytic claim that a predicted quantity reduces to a fitted input. Each pipeline component—rhetorical-role chunking via the external Malik et al. labeler [21], dual dense/BM25 indexing, the hand-defined intent-to-role mapping in Table 3, cross-encoder reranking, and LLM generation—is implemented and then evaluated with separate metrics (Equations 1–3). The only self-referential element is that automated LLM judges compute Faithfulness, Answer Relevancy, and Contextual Relevancy while LLMs also generate answers; Section 6 explicitly limits the claim: "the scores reported come entirely from automated LLM-based metrics rather than review by legal experts." This is a methodological risk (judge bias), not a reduction: Contextual Relevancy scores of 0.376–0.443 (Tables 5–7) show the pipeline is not successful by construction. No parameter is fitted and renamed a prediction, and no load-bearing premise relies on a self-citation; Malik et al. [21] is external prior work, not the authors' own. The role-filtering step (Section 3.3.2) is a hard pre-retrieval constraint, and its unvalidated labeler and hand-crafted mapping could in principle filter away the answer, but the paper reports empirical scores rather than deriving effectiveness from the mapping. Those are correctness and robustness concerns, not circularity in the derivation. The inconsistent 'six vs seven rhetorical roles' description in Section 3.1 is likewise a clarity issue, not a circular step. Therefore no significant circularity is found.

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

The method introduces no new theoretical entities; its contributions are design choices (chunking thresholds, intent-to-role mapping) and an unvalidated reliance on an external annotation model. The free parameters are hand-set thresholds that directly determine what text enters the retrieval index.

free parameters (5)

Small-group merging threshold
Groups with fewer than two sentences are merged with the preceding group (Section 3.1, Step 2). This hand-chosen threshold shapes chunk granularity and retrieval boundaries.
Long-group sub-splitting window size = 15 sentences
Groups with more than 15 sentences are split into overlapping windows of 15 sentences (Section 3.1, Step 3).
Sub-splitting overlap = 2 sentences
Adjacent windows overlap by two sentences to preserve contextual continuity (Section 3.1, Step 3).
Intent-to-role mapping = Five intents mapped to role subsets (Table 3)
Hand-defined mapping between query intent and target rhetorical roles; the role filter excludes any passage outside the mapped roles.
Retrieval top-k and rerank size
The number of candidates retrieved and reranked is not reported in the paper, so the system cannot be exactly reproduced.

assumptions (4)

domain assumption The external rhetorical role labeler (Malik et al., 2022) assigns correct labels to sentences in Indian Supreme Court judgments.
Chunking and role filtering depend entirely on these labels (Section 3.1); no accuracy or validation of the labeler on this corpus is reported.
domain assumption The five query intent categories and their target-role mappings are correct and complete for legal QA over these judgments.
Section 3.3.1 and Table 3 define the mapping by hand; if a query needs text from a non-mapped role, the correct passage is discarded.
domain assumption LLM-as-a-judge metrics are valid proxies for legal QA quality.
Section 4.1; no human evaluation or correlation with expert judgment is provided, and the Limitations section concedes this.
domain assumption The 30 selected judgments and 25 queries are representative of Indian Supreme Court legal QA.
Selection from Indian Kanoon, 10 per domain, with no stated sampling procedure or comparison to a broader benchmark.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Rhetorical-Role-Aware Retrieval-Augmented Generation for Legal Question Answering over Indian Supreme Court Judgments." pith.science (2026). https://pith.science/paper/6DI3NZJ3

@misc{pith2026260806828,
  author       = {Pith},
  title        = {Pith review of: Rhetorical-Role-Aware Retrieval-Augmented Generation for Legal Question Answering over Indian Supreme Court Judgments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6DI3NZJ3}},
  note         = {Machine review of arXiv:2608.06828}
}

read the original abstract

This research paper proposes a Retrieval Augmented Generation (RAG) framework that is specific to the legal field in order to assist interactive retrieval and reason about judgments from the Supreme Court of India. The solution uses an enhanced version of RAG framework which consists of rhetorically based chunking, fusion-based retrieval, and cross encoder reranking methods to increase the relevancy of the information retrieved. In order to improve conversations, the proposed framework uses chat history along with query classification and rewriting in order to understand user intention from successive queries. Additionally, there are features that take into account structural aspects of legal documents, such as isolated names of judges that could have an impact on retrieval quality. The evaluation was done using the DeepEval framework and demonstrated strong performance on metrics including contextual recall and answer relevancy, which proves that the framework is very effective in dealing with legal question-answering tasks that require a lot of context. The results emphasize the importance of domain specific enhancements in developing legal AI systems that are both reliable and explainable.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 13 canonical work pages

[18]

In: Proceedings of the Natural Legal Language Processing Workshop 2022, pp

Malik, V., Sanjay, R., Guha, S.K., Hazarika, A., Nigam, S.K., Bhattacharya, A., Modi, A.: Semantic segmentation of legal documents via rhetorical roles. In: Proceedings of the Natural Legal Language Processing Workshop 2022, pp. 153– 171 (2022)

work page 2022
[21]

In: Aletras, N., Chalkidis, I., Barrett, L., Goant ,˘ a, C., Preot,iuc-Pietro, D

Malik, V., Sanjay, R., Guha, S.K., Hazarika, A., Nigam, S.K., Bhattacharya, A., Modi, A.: Semantic segmentation of legal documents via rhetorical roles. In: Aletras, N., Chalkidis, I., Barrett, L., Goant ,˘ a, C., Preot,iuc-Pietro, D. (eds.) Proceedings of the Natural Legal Language Processing Workshop 2022, pp. 153–171. Association for Computational Ling...

doi:10.18653/v1/2022.nllp-1.13 2022
[1]

Foundations and trends®in information retrieval4(1-2), 1–174 (2009)

Robertson, S., Zaragoza, H.: The probabilistic relevance framework: Bm25 and beyond. Foundations and trends®in information retrieval4(1-2), 1–174 (2009)

work page 2009
[2]

In: Burstein, J., Doran, C., Solorio, T

Devlin, J., Chang, M.-W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North Ameri- can Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers...

work page 2019
[3]

Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)

2018
[4]

Advances in neural information processing systems 33, 9459–9474 (2020)

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., K¨ uttler, H., Lewis, M., Yih, W.-t., Rockt¨ aschel, T.,et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems 33, 9459–9474 (2020)

2020
[5]

In: 11 International Conference on Smart Computing and Informatics, pp

Arora, Y., Sharma, G., Maurya, S., Bansal, N., Jain, P., Rehman, A., Shelke, N.: Building legal intelligence: Designing and developing a chatbot with llm. In: 11 International Conference on Smart Computing and Informatics, pp. 1–12 (2025). Springer

work page 2025
[6]

arXiv preprint arXiv:2302.13971 (2023)

Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozi` ere, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

arXiv 2023

Show all 21 references

[7]

Human-Intelligent Systems Integration7(1), 87–98 (2025)

Mubeen, F., Mehdi, A., Haque, M.A., Nomani, M., Uddin, N.S.: Redefining legal access: a rag-based ai system for indian law. Human-Intelligent Systems Integration7(1), 87–98 (2025)

2025
[8]

arXiv preprint arXiv:2309.14735 (2023)

Nigam, S.K., Mishra, S.K., Mishra, A.K., Shallum, N., Bhattacharya, A.: Legal question-answering in the indian context: Efficacy, challenges, and potential of modern ai models. arXiv preprint arXiv:2309.14735 (2023)

2023 arXiv
[9]

IEEE Access (2025)

Wahidur, R.S., Kim, S., Choi, H., Bhatti, D.S., Lee, H.-N.: Legal query rag. IEEE Access (2025)

2025
[10]

Nigam, S.K., Patnaik, B.D., Mishra, S., Thomas, A.V., Shallum, N., Ghosh, K., Bhattacharya, A.: Nyayarag: Realistic legal judgment prediction with rag under the indian common law system. In: Proceedings of the 14th International Joint Conference on Natural Language Processing ...

2025
[11]

Applied Artificial Intelligence40(1), 2626097 (2026)

Karna, V.R., M, R.R., Babu, B.S., S, N., M, M., V, H.: A hybrid rag-llama frame- work for scalable and accurate interpretation of legal texts. Applied Artificial Intelligence40(1), 2626097 (2026)

2026
[12]

arXiv preprint arXiv:2311.12719 (2023)

Devaraj, P.N., PV, R.T., Gangrade, A., et al.: Development of a legal document ai-chatbot. arXiv preprint arXiv:2311.12719 (2023)

2023 arXiv
[13]

In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

Louis, A., Van Dijck, G., Spanakis, G.: Interpretable long-form legal question answering with retrieval-augmented large language models. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pp. 22266–22275 (2024)

2024
[14]

In: European Conference on Information Retrieval, pp

Askari, A., Yang, Z., Ren, Z., Verberne, S.: Answer retrieval in legal community question answering. In: European Conference on Information Retrieval, pp. 477– 485 (2024). Springer

2024
[15]

In: Findings of the Association for Computational Linguistics: NAACL 2025, pp

Hou, A.B., Weller, O., Qin, G., Yang, E., Lawrie, D., Holzenberger, N., Blair- Stanek, A., Van Durme, B.: Clerc: A dataset for us legal case retrieval and retrieval-augmented analysis generation. In: Findings of the Association for Computational Linguistics: NAACL 2025, pp. 78...

2025
[16]

URL https://arxiv

Pipitone, N., Alami, G.H.: Legalbench-rag: a benchmark for retrieval-augmented generation in the legal domain (2024). URL https://arxiv. org/abs/2408.10343 12

2024 arXiv
[17]

URL https://arxiv

Colombo, P., Pires, T.P., Boudiaf, M., Culver, D., Melo, R., Corro, C., Martins, A.F., Esposito, F., Raposo, V.L., Morgado, S., et al.: Saullm-7b: A pioneering large language model for law, 2024. URL https://arxiv. org/abs/2403.0388353

2024 arXiv
[19]

In: Proceedings of the Nineteenth International Conference on Artificial Intelligence and Law, pp

Paul, S., Mandal, A., Goyal, P., Ghosh, S.: Pre-trained language models for the legal domain: a case study on indian law. In: Proceedings of the Nineteenth International Conference on Artificial Intelligence and Law, pp. 187–196 (2023)

2023
[20]

Journal of Big Data10(1), 127 (2023)

Abdallah, A., Piryani, B., Jatowt, A.: Exploring the state of the art in legal qa systems. Journal of Big Data10(1), 127 (2023)

2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [18]

In: Proceedings of the Natural Legal Language Processing Workshop 2022, pp

Malik, V., Sanjay, R., Guha, S.K., Hazarika, A., Nigam, S.K., Bhattacharya, A., Modi, A.: Semantic segmentation of legal documents via rhetorical roles. In: Proceedings of the Natural Legal Language Processing Workshop 2022, pp. 153– 171 (2022)

work page 2022

[2] [21]

In: Aletras, N., Chalkidis, I., Barrett, L., Goant ,˘ a, C., Preot,iuc-Pietro, D

Malik, V., Sanjay, R., Guha, S.K., Hazarika, A., Nigam, S.K., Bhattacharya, A., Modi, A.: Semantic segmentation of legal documents via rhetorical roles. In: Aletras, N., Chalkidis, I., Barrett, L., Goant ,˘ a, C., Preot,iuc-Pietro, D. (eds.) Proceedings of the Natural Legal Language Processing Workshop 2022, pp. 153–171. Association for Computational Ling...

doi:10.18653/v1/2022.nllp-1.13 2022

[3] [1]

Foundations and trends®in information retrieval4(1-2), 1–174 (2009)

Robertson, S., Zaragoza, H.: The probabilistic relevance framework: Bm25 and beyond. Foundations and trends®in information retrieval4(1-2), 1–174 (2009)

work page 2009

[4] [2]

In: Burstein, J., Doran, C., Solorio, T

Devlin, J., Chang, M.-W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North Ameri- can Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers...

work page 2019

[5] [3]

Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)

2018

[6] [4]

Advances in neural information processing systems 33, 9459–9474 (2020)

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., K¨ uttler, H., Lewis, M., Yih, W.-t., Rockt¨ aschel, T.,et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems 33, 9459–9474 (2020)

2020

[7] [5]

In: 11 International Conference on Smart Computing and Informatics, pp

Arora, Y., Sharma, G., Maurya, S., Bansal, N., Jain, P., Rehman, A., Shelke, N.: Building legal intelligence: Designing and developing a chatbot with llm. In: 11 International Conference on Smart Computing and Informatics, pp. 1–12 (2025). Springer

work page 2025

[8] [6]

arXiv preprint arXiv:2302.13971 (2023)

Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozi` ere, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

arXiv 2023