From prompting to evidence-based translation: A RAG+prompt system for Japanese-Chinese translation and its pedagogical potential

Wenshi Gu

arxiv: 2605.03387 · v1 · submitted 2026-05-05 · 💻 cs.CL

From prompting to evidence-based translation: A RAG+prompt system for Japanese-Chinese translation and its pedagogical potential

Wenshi Gu This is my paper

Pith reviewed 2026-05-07 16:53 UTC · model grok-4.3

classification 💻 cs.CL

keywords RAGprompt engineeringJapanese-Chinese translationnoun-modifying clause constructionsBLEU evaluationmachine translationlarge language models

0 comments

The pith

A RAG-enhanced prompt system raises BLEU scores for Japanese-Chinese translation of noun-modifying clauses as the example database grows.

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

The paper evaluates a system that adds linguistic analysis and retrieved translation examples to prompts for translating Japanese sentences with noun-modifying clause constructions into Chinese. An analysis module identifies clause types and risk areas, then the top five similar examples are pulled from a knowledge base and included in the prompt to GPT-4o. Tests on 66 sentences show BLEU rising from 24.28 without retrieval to 29.96 with 2,000 examples, with gains increasing alongside database size. This matters because it provides an interpretable way to boost performance on difficult constructions without changing the underlying model. The approach keeps the process auditable by making the added evidence explicit.

Core claim

The RAG+Prompt system improves Japanese-Chinese translation of sentences containing noun-modifying clause constructions by combining an analysis module that outputs inner versus outer NMCC classifications and risk predictions with retrieval of the top-5 similar examples using L2 distance. These elements are inserted into an enhanced prompt for the LLM. On a 66-sentence test set using GPT-4o, mean sentence-level BLEU scores increase steadily with knowledge base size, from 24.28 at zero examples to 29.96 at 2,000 examples.

What carries the argument

The RAG+Prompt pipeline that performs linguistic analysis to label NMCC types and risks, retrieves similar Ja-Zh examples via embedding distance, and augments the prompt without modifying the base model.

If this is right

Translation quality, as measured by BLEU, improves consistently as the knowledge base of examples grows larger.
The method achieves gains while remaining interpretable because the retrieved examples and analysis labels are explicitly added to the prompt.
Performance holds across varying sizes of the example database from 100 to 2,000 entries.
The system does not require changes to the underlying large language model.

Where Pith is reading between the lines

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

This approach could be adapted for other language pairs that share complex grammatical structures.
The explicit examples might help in pedagogical settings by showing learners how similar sentences were translated.
Human judgments or additional metrics beyond BLEU would be needed to confirm real-world translation improvements.
Scaling the knowledge base further or using domain-specific examples could yield even larger gains.

Load-bearing premise

The 66-sentence test set adequately represents real-world Japanese-Chinese translations involving noun-modifying clause constructions, and BLEU scores reflect meaningful quality improvements.

What would settle it

Applying the same system to a larger and more varied collection of sentences from different genres and finding no increase or a decrease in translation quality metrics.

Figures

Figures reproduced from arXiv: 2605.03387 by Wenshi Gu.

**Figure 1.** Figure 1: RAG+Prompt translation pipeline for Ja→Zh sentences containing NMCCs. Note. A1 = NMCC type classification (inner vs. outer); A2 = pre-translation risk prediction (lexical choice / NMCC handling / word order / style/register). Retrieval uses top-k = 5 with L2 distance. 4. Experimental Design This chapter outlines the experimental design, including datasets, retrieval configurations, and evaluation metrics. … view at source ↗

read the original abstract

Large language models perform well on high-resource pairs but are less reliable for Japanese-Chinese sentences containing noun-modifying clause constructions (NMCCs). This study evaluates a retrieval-augmented generation RAG+Prompt translation system that integrates linguistic analysis, embedding-based retrieval, prompt construction, and LLM generation without modifying the base model. The analysis module outputs A1 (inner vs. outer NMCC) and A2 (risk predictions: lexical choice/NMCC handling/word order/style/register); top-k = 5 similar Ja-Zh examples (L2 distance) and A1/A2 are inserted into an enhanced prompt. Using GPT-4o and a 66-sentence test set, we compare six knowledge-base sizes (0/100/200/500/1,000/2,000). Macro-averaged sentence-level BLEU (1-4-gram with brevity penalty; cased; Chinese at the character level) is the sole metric. Mean BLEU increases from 24.28 at 0 (RAG disabled) to 29.96 at 2,000 (+5.68; +23.4%). The upward trend holds across sizes, with larger knowledge bases yielding higher scores. We conclude that the RAG+Prompt translation system improves Ja-Zh translation of sentences containing NMCCs in an interpretable and auditable manner. Limitations include one base model, one metric, and reliance on published texts and commercial APIs; future work will broaden genres, language pairs, and evaluation metrics.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

RAG scaling shows a BLEU lift for Japanese-Chinese NMCC translation but lacks stats to confirm it's not noise.

read the letter

The paper's key result is that increasing the size of the example database from zero to 2,000 sentences raises mean sentence-level BLEU from 24.28 to 29.96 on a 66-sentence test set for translating Japanese noun-modifying clauses into Chinese. They achieve this by adding linguistic analysis of the construction type and risk factors into the prompt along with retrieved examples. This is a solid, targeted application of existing RAG techniques to a real pain point in machine translation. The A1/A2 analysis module provides some interpretability, showing what the system thinks about each input, and the monotonic trend across database sizes is easy to follow. Using published texts for the knowledge base keeps things grounded. The main limitation is the evaluation design. No variance, confidence intervals, or statistical tests are reported, so we can't tell if the improvement exceeds what you'd expect from random variation on such a small set. Relying only on BLEU, especially character-level for Chinese, leaves open questions about actual translation quality. There's also no human evaluation or tests on other models. This paper is for researchers and developers focused on low-resource or structurally complex translation pairs, and possibly for those exploring RAG in educational contexts. A reader working on similar augmentation methods would find the concrete numbers and prompt details useful. It deserves peer review because the experiment is clearly described and the linguistic motivation is sound, even though the current evidence is preliminary. Reviewers could push for better statistical grounding and broader testing.

Referee Report

1 major / 4 minor

Summary. The manuscript describes a retrieval-augmented generation (RAG) system combined with prompting for translating Japanese-Chinese sentences featuring noun-modifying clause constructions (NMCCs). The system uses linguistic analysis to classify NMCCs (A1: inner/outer) and predict risks (A2), retrieves top-5 similar examples via embedding distance from knowledge bases of sizes 0/100/200/500/1,000/2,000, and augments prompts for GPT-4o without modifying the base model. On a 66-sentence test set, macro-averaged sentence-level BLEU (character-level for Chinese) rises monotonically from 24.28 (RAG disabled) to 29.96 (KB=2,000), a +5.68 gain (+23.4%). The authors conclude that the approach improves translation of NMCC-containing sentences in an interpretable, auditable manner with pedagogical potential.

Significance. If the trend is statistically reliable, the work provides evidence that linguistically informed RAG can address specific syntactic challenges in Japanese-Chinese translation without fine-tuning, offering an auditable alternative to black-box prompting. The integration of NMCC analysis and example retrieval adds a concrete, domain-specific contribution to RAG applications in East Asian languages. The pedagogical angle is noted but remains secondary. Limitations to one model, one metric, and small test set reduce immediate impact, but the empirical baseline comparison is a clear strength.

major comments (1)

[Results (comparison of knowledge-base sizes)] Results section (comparison across KB sizes 0–2,000): The central claim of a consistent upward trend and +5.68 BLEU improvement rests on mean scores from 66 sentences without reported per-sentence variances, standard deviations, bootstrap confidence intervals, or any hypothesis test (e.g., paired t-test or Wilcoxon on the 66 differences). Character-level BLEU on Chinese is known to be sensitive to a few difficult items; without statistical grounding it is impossible to rule out that the monotonic pattern and gain arise from sampling noise rather than retrieval quality. This directly undermines confidence in the reported improvement.

minor comments (4)

The 66-sentence test set is small and its selection criteria or genre distribution are not detailed; a clearer description would help evaluate whether it represents typical NMCC usage in real-world Japanese-Chinese translation.
Evaluation relies solely on BLEU; adding at least one neural metric (e.g., COMET) or a small-scale human evaluation would provide a more robust picture of quality gains beyond n-gram overlap.
Only GPT-4o is tested; results with at least one additional model (open or closed) would strengthen claims about the RAG+prompt approach being model-agnostic.
The pedagogical potential is asserted in the title and conclusion but receives limited concrete discussion; expanding this section with example prompts or classroom scenarios would better support the claim.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the major comment regarding the statistical grounding of our results below.

read point-by-point responses

Referee: [Results (comparison of knowledge-base sizes)] Results section (comparison across KB sizes 0–2,000): The central claim of a consistent upward trend and +5.68 BLEU improvement rests on mean scores from 66 sentences without reported per-sentence variances, standard deviations, bootstrap confidence intervals, or any hypothesis test (e.g., paired t-test or Wilcoxon on the 66 differences). Character-level BLEU on Chinese is known to be sensitive to a few difficult items; without statistical grounding it is impossible to rule out that the monotonic pattern and gain arise from sampling noise rather than retrieval quality. This directly undermines confidence in the reported improvement.

Authors: We agree that reporting only mean BLEU scores without measures of variability or formal hypothesis testing weakens the evidential support for the observed improvement. The manuscript currently provides macro-averaged sentence-level BLEU for each of the six knowledge-base sizes but does not include standard deviations, confidence intervals, or statistical tests. In the revised manuscript we will add the standard deviation of the 66 per-sentence BLEU scores for each condition, 95% bootstrap confidence intervals around the means, and the result of a paired Wilcoxon signed-rank test comparing the KB=0 and KB=2,000 conditions. We also note that the monotonic increase in mean BLEU across all six successively larger knowledge bases (0, 100, 200, 500, 1,000, 2,000) supplies qualitative evidence that the trend is unlikely to be an artifact of sampling noise in any isolated comparison. revision: yes

Circularity Check

0 steps flagged

No circularity; purely empirical comparison of RAG variants against baseline

full rationale

The manuscript reports direct BLEU measurements on a fixed 66-sentence test set across six knowledge-base sizes (0 to 2,000). No equations, fitted parameters, or first-principles derivations appear; the central claim is simply the observed monotonic increase in mean sentence-level BLEU when retrieval is enabled. No self-citations are invoked to justify uniqueness or ansatzes, and the evaluation uses a standard external metric (character-level BLEU) without redefining or predicting its own inputs. The derivation chain is therefore self-contained and non-circular.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The work rests on standard RAG components and LLM prompting; the only experimental variable is knowledge-base size, which is varied rather than fitted. No new entities are postulated.

axioms (2)

domain assumption BLEU score is a sufficient proxy for translation quality in this setting
It is used as the sole reported metric without additional justification or human correlation in the abstract.
domain assumption The 66-sentence test set adequately represents NMCC translation challenges
Conclusions about system effectiveness are drawn directly from performance on this fixed set.

pith-pipeline@v0.9.0 · 5577 in / 1310 out tokens · 48802 ms · 2026-05-07T16:53:57.661950+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

14 extracted references · 5 canonical work pages

[1]

translate into Chinese

1 Article From prompting to evidence-based translation: A RAG+prompt system for Japanese–Chinese translation and its pedagogical potential Wenshi Gu (guwenshi@buaa.edu.cn) School of Foreign Languages, Beihang Univeristy.Beijing, China Abstract Large language models perform well on high-resource pairs but are less reliable for Japanese→Chinese sentences co...

2022
[1]

translate into Chinese

Introduction Large language models (LLMs) can understand prompts and follow instructions because they are trained with supervised instruction tuning on human demonstrations and reinforcement learning from human feedback (RLHF). InstructGPT showed that combining instruction tuning with RLHF based on preference rankings improves instruction following (Ouyan...

2022
[2]

The review focuses on representative surveys and empirical studies between 2022 and 2025, emphasizing findings applicable to Ja→Zh translation of sentences containing NMCCs

Literature Review This section reviews two strands of research most relevant to this study: (1) the applications and limitations of LLMs in machine translation and in the pedagogy of foreign languages and translation, and (2) the concept and mechanism of RAG and its empirical evidence in translation and in the pedagogy of foreign languages and translation...

2022
[3]

Li, Y. (2024). ChatGPT in language learning: A systematic PRISMA review of the first year of research. Computer Assisted Language Learning. Advance online publication. https://doi.org/10.1080/09588221.2024.2345678 Mai, H., Pham, T., & Nguyen, L. (2024). Opportunities and risks of ChatGPT in education: A systematic review. Frontiers in Education, 9, 144567...

work page doi:10.1080/09588221.2024.2345678 2024
[3]

otoko” (男, ‘man’) can be the agent of the main verb “yaku

Methodology This chapter describes the overall methodology and implementation. Building on the research goals above, we construct a RAG+Prompt translation system tailored to Ja→Zh translation of sentences containing NMCCs, providing a unified technical foundation for the experimental evaluation and the pedagogical application. 3.1 Building the RAG+Prompt ...

2023
[4]

Research on the Development of College Japanese Courses Oriented towards 'Engineering + Japanese'

Wang, J., Liu, Z., & Huang, M. (2025). RAGtrans: Retrieval-augmented translation with unstructured multilingual documents. arXiv preprint. https://arxiv.org/abs/2412.04342 Zhao, W., Liu, H., & Sun, M. (2023). Large language models for machine translation: Progress and challenges. Transactions of the Association for Computational Linguistics, 11(1), 897–91...

work page doi:10.1162/tacl_a_00573 2025
[4]

The goal is to assess the effect of the RAG+Prompt translation system on 8 translation quality and to examine how the size of the RAG knowledge base influences performance

Experimental Design This chapter outlines the experimental design, including datasets, retrieval configurations, and evaluation metrics. The goal is to assess the effect of the RAG+Prompt translation system on 8 translation quality and to examine how the size of the RAG knowledge base influences performance. 4.1 Datasets We adopt a two-part experimental s...

2000
[5]

Absolute gain

Average BLEU by RAG knowledge base size. RAG size Average BLEU Absolute gain vs. baseline (RAG disabled)* Relative gain vs. baseline (RAG disabled, %)* 0 (RAG disabled) 24.28 — — 100 24.32 +0.04 +0.2% 200 24.86 +0.58 +2.4% 500 26.77 +2.49 +10.3% 1000 27.50 +3.22 +13.3% 2000 29.96 +5.68 +23.4% * “Absolute gain” = mean BLEU at the given size − mean BLEU at ...

2000
[6]

Gains at small sizes (100/200) are limited (+0.04 and +0.58), while gains at 500 and above become substantial, peaking at 2,000 (+5.68; +23.4%)

to 29.96 (size 2,000), mean BLEU increases steadily, indicating a sustained positive effect of knowledge-base expansion. Gains at small sizes (100/200) are limited (+0.04 and +0.58), while gains at 500 and above become substantial, peaking at 2,000 (+5.68; +23.4%). (2) Relative gains: relative to the RAG-disabled baseline, improvements are modest (0.2%–2....

2023
[7]

ズボンのベルト部分やリストバンド、靴下など、横向きのグリッドをどちらに曲げるべきかのヒントが隠れています

Translation outputs and BLEU under three RAG sizes (Case 1: Lexical choice issues) Table 2 presents the translations and BLEU scores under three RAG sizes. At size 0 (RAG disabled), GPT-4o translated “ズボンのベルト部分やリストバンド、靴下など、横向きのグリッドをどちらに曲げるべきかのヒントが隠れています” as “可以发现关于裤子腰带部分、腕带、袜子等横向网格应该如何弯曲的提示”. The crucial verb “隠れています (are hidden)” was omitted, thereby era...

2000
[8]

空間上の二つの面が交わった時に出来る直線の式を求める、という問題…（a problem that required her to find the equation of the line formed by the intersection of two planes in space）

Translation outputs and BLEU under three RAG sizes (Case 2: NMCC handling issues). RAG size BLEU Target-language output (Chinese) RAG=0 36.04 雪穂正在解决一个关于求解空间中两个平面相交时所形成的直线方程的问题。解题的方法他已经教过，她也理解了。她手中的自动铅笔几乎没有停下过。 RAG=200 37.50 雪穂正在解答一个关于求解空间中两平面相交时形成的直线方程的问题。解题方法他已经教过，她也理解了。她手中的自动铅笔几乎没有停下过。 RAG=2000 48.29 雪穗正在解一道题，题目是求空间中两个面相交时的直线方程式。解法已经教过，她也理解了，因此她手中的自动铅笔几...

2000
[9]

Conclusions, Limitations, and Future Directions We validated the effectiveness of a RAG+Prompt translation system for Ja→Zh translation. The modular pipeline—linguistic analysis → embedding-based retrieval → prompt construction → translation generation—was run under unified conditions using GPT-4o and a 66-item test set. The RAG knowledge base comprised p...

work page doi:10.1016/j.ssaho.2025.101866 2000
[10]

https://doi.org/10.3390/educsci14050233 Baldwin, T. (2004). Making sense of Japanese relative clause constructions. In Proceedings of the 2nd Workshop on Text Meaning and Interpretation (pp. 49–56). Association for Computational Linguistics. https://aclanthology.org/W04-0907/ Chen, X., Zhou, K., & Wang, H. (2024). Retrieval-augmented knowledge integration...

work page doi:10.3390/educsci14050233 2004
[11]

Retrieval-Augmented Generation (RAG) Chatbots for Education: A Survey of Applications

https://doi.org/10.3390/app15084234 Teramura, H. [寺村秀夫]. (1992). 連体修飾のシンタクスと意味—その1— [The syntax and semantics of noun modification—Part 1]. In 寺村秀夫論文集Ⅰ [Collected Papers of Hideo Teramura, Vol. 1]. 東京:くろしお出版. (Original work published 1975 in 日本語・日本文化, No. 4). Thüs, D., Malone, S., & Brünken, R. (2024). Exploring generative AI in higher education: A RAG sy...

work page doi:10.3390/app15084234 1992

[1] [1]

translate into Chinese

1 Article From prompting to evidence-based translation: A RAG+prompt system for Japanese–Chinese translation and its pedagogical potential Wenshi Gu (guwenshi@buaa.edu.cn) School of Foreign Languages, Beihang Univeristy.Beijing, China Abstract Large language models perform well on high-resource pairs but are less reliable for Japanese→Chinese sentences co...

2022

[2] [1]

translate into Chinese

Introduction Large language models (LLMs) can understand prompts and follow instructions because they are trained with supervised instruction tuning on human demonstrations and reinforcement learning from human feedback (RLHF). InstructGPT showed that combining instruction tuning with RLHF based on preference rankings improves instruction following (Ouyan...

2022

[3] [2]

The review focuses on representative surveys and empirical studies between 2022 and 2025, emphasizing findings applicable to Ja→Zh translation of sentences containing NMCCs

Literature Review This section reviews two strands of research most relevant to this study: (1) the applications and limitations of LLMs in machine translation and in the pedagogy of foreign languages and translation, and (2) the concept and mechanism of RAG and its empirical evidence in translation and in the pedagogy of foreign languages and translation...

2022

[4] [3]

Li, Y. (2024). ChatGPT in language learning: A systematic PRISMA review of the first year of research. Computer Assisted Language Learning. Advance online publication. https://doi.org/10.1080/09588221.2024.2345678 Mai, H., Pham, T., & Nguyen, L. (2024). Opportunities and risks of ChatGPT in education: A systematic review. Frontiers in Education, 9, 144567...

work page doi:10.1080/09588221.2024.2345678 2024

[5] [3]

otoko” (男, ‘man’) can be the agent of the main verb “yaku

Methodology This chapter describes the overall methodology and implementation. Building on the research goals above, we construct a RAG+Prompt translation system tailored to Ja→Zh translation of sentences containing NMCCs, providing a unified technical foundation for the experimental evaluation and the pedagogical application. 3.1 Building the RAG+Prompt ...

2023

[6] [4]

Research on the Development of College Japanese Courses Oriented towards 'Engineering + Japanese'

Wang, J., Liu, Z., & Huang, M. (2025). RAGtrans: Retrieval-augmented translation with unstructured multilingual documents. arXiv preprint. https://arxiv.org/abs/2412.04342 Zhao, W., Liu, H., & Sun, M. (2023). Large language models for machine translation: Progress and challenges. Transactions of the Association for Computational Linguistics, 11(1), 897–91...

work page doi:10.1162/tacl_a_00573 2025

[7] [4]

The goal is to assess the effect of the RAG+Prompt translation system on 8 translation quality and to examine how the size of the RAG knowledge base influences performance

Experimental Design This chapter outlines the experimental design, including datasets, retrieval configurations, and evaluation metrics. The goal is to assess the effect of the RAG+Prompt translation system on 8 translation quality and to examine how the size of the RAG knowledge base influences performance. 4.1 Datasets We adopt a two-part experimental s...

2000

[8] [5]

Absolute gain

Average BLEU by RAG knowledge base size. RAG size Average BLEU Absolute gain vs. baseline (RAG disabled)* Relative gain vs. baseline (RAG disabled, %)* 0 (RAG disabled) 24.28 — — 100 24.32 +0.04 +0.2% 200 24.86 +0.58 +2.4% 500 26.77 +2.49 +10.3% 1000 27.50 +3.22 +13.3% 2000 29.96 +5.68 +23.4% * “Absolute gain” = mean BLEU at the given size − mean BLEU at ...

2000

[9] [6]

Gains at small sizes (100/200) are limited (+0.04 and +0.58), while gains at 500 and above become substantial, peaking at 2,000 (+5.68; +23.4%)

to 29.96 (size 2,000), mean BLEU increases steadily, indicating a sustained positive effect of knowledge-base expansion. Gains at small sizes (100/200) are limited (+0.04 and +0.58), while gains at 500 and above become substantial, peaking at 2,000 (+5.68; +23.4%). (2) Relative gains: relative to the RAG-disabled baseline, improvements are modest (0.2%–2....

2023

[10] [7]

ズボンのベルト部分やリストバンド、靴下など、横向きのグリッドをどちらに曲げるべきかのヒントが隠れています

Translation outputs and BLEU under three RAG sizes (Case 1: Lexical choice issues) Table 2 presents the translations and BLEU scores under three RAG sizes. At size 0 (RAG disabled), GPT-4o translated “ズボンのベルト部分やリストバンド、靴下など、横向きのグリッドをどちらに曲げるべきかのヒントが隠れています” as “可以发现关于裤子腰带部分、腕带、袜子等横向网格应该如何弯曲的提示”. The crucial verb “隠れています (are hidden)” was omitted, thereby era...

2000

[11] [8]

空間上の二つの面が交わった時に出来る直線の式を求める、という問題…（a problem that required her to find the equation of the line formed by the intersection of two planes in space）

Translation outputs and BLEU under three RAG sizes (Case 2: NMCC handling issues). RAG size BLEU Target-language output (Chinese) RAG=0 36.04 雪穂正在解决一个关于求解空间中两个平面相交时所形成的直线方程的问题。解题的方法他已经教过，她也理解了。她手中的自动铅笔几乎没有停下过。 RAG=200 37.50 雪穂正在解答一个关于求解空间中两平面相交时形成的直线方程的问题。解题方法他已经教过，她也理解了。她手中的自动铅笔几乎没有停下过。 RAG=2000 48.29 雪穗正在解一道题，题目是求空间中两个面相交时的直线方程式。解法已经教过，她也理解了，因此她手中的自动铅笔几...

2000

[12] [9]

Conclusions, Limitations, and Future Directions We validated the effectiveness of a RAG+Prompt translation system for Ja→Zh translation. The modular pipeline—linguistic analysis → embedding-based retrieval → prompt construction → translation generation—was run under unified conditions using GPT-4o and a 66-item test set. The RAG knowledge base comprised p...

work page doi:10.1016/j.ssaho.2025.101866 2000

[13] [10]

https://doi.org/10.3390/educsci14050233 Baldwin, T. (2004). Making sense of Japanese relative clause constructions. In Proceedings of the 2nd Workshop on Text Meaning and Interpretation (pp. 49–56). Association for Computational Linguistics. https://aclanthology.org/W04-0907/ Chen, X., Zhou, K., & Wang, H. (2024). Retrieval-augmented knowledge integration...

work page doi:10.3390/educsci14050233 2004

[14] [11]

Retrieval-Augmented Generation (RAG) Chatbots for Education: A Survey of Applications

https://doi.org/10.3390/app15084234 Teramura, H. [寺村秀夫]. (1992). 連体修飾のシンタクスと意味—その1— [The syntax and semantics of noun modification—Part 1]. In 寺村秀夫論文集Ⅰ [Collected Papers of Hideo Teramura, Vol. 1]. 東京:くろしお出版. (Original work published 1975 in 日本語・日本文化, No. 4). Thüs, D., Malone, S., & Brünken, R. (2024). Exploring generative AI in higher education: A RAG sy...

work page doi:10.3390/app15084234 1992