PaperlayerView exact revision on arXiv

Source-linked AI summary

MAS-Bai: Linguistically Informed Adaptive Multi-Agent Translation for Bowen-Script Bai

Jingting Zheng, Hong Ren, Yuehan Cui, Shaolin Zhu, Deyi Xiong

arXiv:2610.23205v1cs.CL

TL;DR

Bowen-script Bai translation is challenging because Chinese-looking characters may be phonetic loans, while limited data and fixed workflows are poorly matched to the task. MAS-Bai combines Bai-specific philological agents, conditional routing, and cycle-consistency; it improves over Standard RAG on the reported metrics.

  • Problem

    Standard translation workflows are poorly matched to Bowen-script Bai, where Chinese-looking characters may have Bai meanings or functions and limited data cannot readily counter pretrained Mandarin priors.

    Source: §1 · p. 1

  • Method

    MAS-Bai routes inputs between retrieval–translation and philological paths, using Bai-specific agents and cycle-consistency to evaluate and refine drafts.

    Source: §1 · pp. 1–2

  • Results

    MAS-Bai achieves 16.23 BLEU and 62.15 COMET, improving over Standard RAG by 8.11 BLEU and 6.85 COMET.

    Source: §5.2 · p. 6

  • Takeaways & Limitations

    The largest gains occur on phonetic-loan ambiguity and divergent Bai word order, with human evaluation and verifier analyses consistent with the translation results.

    Source: §7 · p. 9

    The framework was evaluated within the Bowen-script domains represented in the corpus; applying it to another language would require new linguistic analysis and routing validation.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Large Language Models (LLMs) remain unreliable for extremely low-resource languages whose writing systems conflict with high-resource pretraining priors. We study Bai-Chinese translation in the Bowen script, where Chinese characters are frequently borrowed phonetically rather than semantically, causing severe literal-translation hallucinations. We propose MAS-Bai, a linguistically informed adaptive multi-agent framework for Bowen-script Bai-Chinese translation. MAS-Bai routes each input through either a lightweight retrieval-translation path or a deeper philological path with etymological and syntactic agents, and uses cycle-consistency with a lightweight reverse translator as an inference-time reliability signal. We construct a 6,536-sentence Bai-Chinese corpus and evaluate on a 1,536-sentence test set. MAS-Bai achieves 16.23 BLEU and 62.15 COMET, improving over Standard RAG by 8.11 BLEU and 6.85 COMET. Forward LoRA, expert ablations and controlled routing experiments distinguish the roles of supervised adaptation, Bai-specific expertise and conditional routing. The largest gains occur on Bowen phonetic loans and divergent Bai word order, while human evaluation and reverse-verifier analyses support the automatic results.

1 Introduction

Bowen-script Bai poses a translation challenge because Chinese-looking characters may function phonetically rather than carry their usual Chinese meaning, while training data and fixed workflows are limited. MAS-Bai uses Bai-specific agents and conditional routing, and the paper reports outperforming several translation baselines.

1 Introduction

  • Bowen-script Bai creates translation difficulty because Chinese characters can be borrowed phonetically and express different meanings or grammatical functions than in Chinese.Limited training data and fixed workflows add challenges across inputs ranging from transparent dialogue to opaque phonetic-loan cases.Source: §1 · p. 1
  • Prior computational work examines Bai phonology and spoken translation, but machine translation of the historical logographic Bowen script remains largely unaddressed.The authors also note that general reasoning prompts do not incorporate the etymological strategies needed to decipher Bowen-script Bai.Source: §1 · p. 1
  • MAS-Bai routes inputs between lightweight retrieval–translation and deeper philological paths, using specialized agents and cycle-consistency to refine drafts.Its experts and routing cues are tailored to Bai; the corpus contains 6,536 sentence pairs.Source: §1 · pp. 1–2
  • The paper constructs a 6,536-sentence Bai–Chinese corpus and annotates the 1,536-sentence test set by register and linguistic phenomena.The annotations support stratified evaluation of script opacity and divergent syntax.Source: §1 · p. 2
  • MAS-Bai outperforms zero-shot, ICL, Standard RAG, Forward LoRA, and Static MAS baselines.The authors also describe expert ablations, controlled routing comparisons, stratified evaluation, human evaluation, and cycle-consistency analyses.Source: §1 · p. 2

2 Related Work

Prior low-resource MT and translation-agent work addresses model adaptation, workflow design, or multistage inference. MAS-Bai’s distinction is to use explicit sentence-level Bai/Bowen linguistic cues to select its inference workflow.

LLMs for Low-Resource Machine Translation

  • MAS-Bai selects an inference workflow using explicit Bai/Bowen cues, rather than adapting model parameters or modules to sentence-level philological complexity.Its cues guide selection of external philological agent topologies at inference time.Source: §2 · pp. 2–3

Multi-Agent Systems and Workflow Design

  • Earlier multi-agent systems often used static topologies, while newer approaches search or refine workflows using task-level signals; MAS-Bai routes by sentence-level phonetic-loan and syntax cues.The architecture’s Meta-Agent routes inputs to retrieval–translation or philological paths.Source: §2 · p. 2, Figure 1 · p. 3

LLM Agents for Translation

  • Translation-agent systems explore memory, reflective feedback, long reasoning, and online adaptation, whereas MAS-Bai uses Bai/Bowen cues to select external philological workflows.The cited examples include document-level memory, dual-learning feedback, literary chain-of-thought translation, and parameter internalization.Source: §2 · pp. 2–3

3 Bai–Chinese Parallel Corpus

The paper constructs a sentence-level Bowen-script Bai–Chinese corpus with Modern Chinese references and annotates its test set to support evaluation across registers and linguistic phenomena.

3 Bai–Chinese Parallel Corpus

  • The corpus contains 6,536 aligned Bai–Chinese sentence pairs compiled from an officially published Dabenqv collection and a community-held handwritten Bowen-script manuscript.The construction process includes digitization, translation, sentence-level alignment, and quality control.Source: §3 · p. 3
  • The 1,536-sentence test set is annotated by register and Bai-specific phenomena, including phonetic loans and word-order patterns that diverge from standard Chinese.The register categories are dialogue-like utterances, formal narrative passages, and lyric or verse lines.Source: §3 · p. 3

4 Inference-Time Adaptive Multi-Agent Collaboration Framework

MAS-Bai combines specialized philological agents with conditional pathway selection and a cycle-consistency loop. The system uses linguistic cues to choose a translation path and can revise its topology when a draft appears inconsistent with the source.

4 Inference-Time Adaptive Multi-Agent Collaboration Framework

  • The framework consists of a philological agent pool, a Meta-Agent that conditionally routes inputs, and a cycle-consistency self-evolution loop.Together, these components provide an adaptive workflow for Bowen-script Bai translation.Source: §4 · p. 3, §4.2 · p. 4

4.1 Philological Agent Pool

  • The philological agent pool separates character decipherment, syntactic reordering, contextual retrieval, and Chinese generation.The Etymologist distinguishes semantic morphemes from phonetic loans, while the Syntactician produces an intermediate representation approximating Chinese word order; the Translator synthesizes these cues and retrieved examples into a draft.Source: §4.1 · pp. 3–4

4.2 Conditional Meta-Architecting Mechanism

  • The Meta-Agent uses counts of ambiguous Bowen loans and divergent-order syntax markers to route inputs to an efficiency-oriented or quality-oriented path.The deterministic rule selects QUALITY when a(x) + s(x) > 2 and EFFICIENCY otherwise; the threshold was selected on a 100-sentence development sample and fixed for testing. The lightweight path avoids unnecessary expert calls, while the deeper path adds etymological and syntactic analysis.Source: §4.2 · p. 4

4.3 Inference-Time Self-Evolution via Cycle-Consistency

  • A reverse translator maps draft Chinese back into Bai, and the verifier compares the reconstruction with the original using semantic similarity and Bai/Bowen character overlap.The consistency threshold is τ = 0.85, with a maximum of T = 3 refinement iterations; failed checks prompt topology updates, and the highest-scoring candidate is returned if none reaches the threshold.Source: §4.3 · p. 4

5 Experiments

Experiments evaluate translation quality, system components, routing behavior, verifier reliability, register and linguistic subsets, and human judgments. On a 1,536-sentence test set, the results compare MAS-Bai with baselines and examine its targeted linguistic capabilities.

5.1 Experimental Setup

  • A 5,000-pair Knowledge Base supported retrieval and reverse-translator training, while a separate 1,536-sentence test set was reserved for evaluation.The evaluation reported BLEU, ChrF++, and COMET; COMET was interpreted alongside human assessments because the source language falls outside typical metric-training distributions.Source: §5.1 · p. 5

5.2 Main Results

  • 16.23 BLEU, 35.74 ChrF++, and 62.15 COMET: MAS-Bai exceeds Standard RAG by 8.11 BLEU, 11.18 ChrF++, and 6.85 COMET.It also exceeds Forward LoRA by 2.78 BLEU, 4.54 ChrF++, and 4.35 COMET.Source: Table 2 · p. 6, §5.2 · p. 6

5.3 Attribution of Performance Gains

  • Removing the Etymologist lowered BLEU from 16.23 to 11.45, while removing the Syntactician lowered it to 13.08.With the expert pool and generation settings fixed, MAS-Bai also exceeded Static MAS by 1.72 BLEU and 3.92 COMET using 38% fewer tokens, indicating an additional contribution from conditional routing.Source: §5.3 · pp. 6–7, Table 4 · p. 7

5.4 Inference-Time Adaptation

  • The Meta-Agent selected the efficiency-oriented path for 72% of simple dialogue-like inputs and the quality-oriented path for 85% of complex inputs.Against adjudicated human labels, it achieved 88.5% accuracy on 200 sentences, with annotator agreement of Cohen’s κ = 0.84.Source: §5.4 · p. 7

5.5 Stratified Results by Register and Linguistic Phenomenon

  • MAS-Bai’s COMET gains over Standard RAG rose from 4.05 on dialogue-like inputs to 6.40 on formal narrative and 9.85 on lyric or verse lines.Gains were 25.40 COMET on the Bowen Script Subset and 21.70 on the SOV Syntax Subset, targeting phonetic loans and word-order divergence, respectively.Source: §5.5 · p. 7, Table 5 · p. 7

5.6 Robustness of Cycle-Consistency Verification

  • Removing self-evolution reduced BLEU from 16.23 to 12.87 and COMET from 62.15 to 54.91.The disjoint-data verifier reached 15.88 BLEU and 61.34 COMET, while scores increased with verifier training data and cycle-consistency correlated with human adequacy (Spearman’s ρ = 0.68, p < 0.01).Source: §5.6 · pp. 7–8

5.7 Human Evaluation

  • MAS-Bai reduced the Bowen hallucination rate from 76.0% for zero-shot GPT-4o and 38.0% for Standard RAG to 8.0%, while achieving the highest adequacy and fluency scores.The human evaluation covered 100 test sentences, with adequacy and fluency averaged across three annotators.Source: Table 6 · p. 8, §5.7 · p. 8

6 Discussion

The discussion attributes gains to supervised adaptation, Bai-specific expertise, and adaptive routing, and describes a broader design principle whose demonstrated scope remains Bowen-script domains represented in the corpus.

6 Discussion

  • Forward LoRA, expert ablations, and routing comparisons distinguish contributions from supervised adaptation, Bai-specific knowledge, and conditional pathway selection.The findings characterize linguistic specialization and adaptive orchestration as complementary roles.Source: §6 · p. 8
  • The broader design principle links interpretable linguistic difficulty cues to specialist capabilities and uses those cues to select and revise inference pathways.Applying it elsewhere requires language-specific analysis, cue inventories, specialist definitions, and routing validation; the present evidence covers Bowen-script domains represented in the corpus.Source: §6 · p. 8
  • Cycle-consistency provides an internal screening signal, while adaptive routing reduces average token consumption by 38% relative to Static MAS.The discussion reports that stronger reverse models yield higher translation scores and verifier scores correlate with human adequacy.Source: §6 · p. 8

7 Conclusion

The conclusion presents MAS-Bai and its corpus, reports comparisons against translation baselines, and limits the demonstrated effectiveness to the represented Bowen-script domains.

7 Conclusion

  • The 6,536-sentence Bai–Chinese corpus and MAS-Bai were evaluated on a 1,536-sentence test set against zero-shot, ICL, Standard RAG, Forward LoRA, and Static MAS baselines.The conclusion states that MAS-Bai outperforms these baselines.Source: §7 · p. 9
  • Expert ablations and controlled routing comparisons indicate complementary contributions from Bai-specific philological knowledge and conditional pathway selection.The largest gains occur on phonetic-loan ambiguity and divergent Bai word order; human evaluation and verifier analyses are consistent with the translation results.Source: §7 · p. 9
  • The findings demonstrate MAS-Bai’s effectiveness within the Bowen-script domains represented in the corpus; transfer to other low-resource languages or shared-glyph systems remains future work.Source: §7 · p. 9

Limitations

MAS-Bai is evaluated within the Bowen-script domains represented in its corpus, and its language-specific design requires new analysis and routing validation for other languages. The evaluation also has boundaries in its split design, cycle-consistency signal, human sample, and inference cost.

Limitations

  • MAS-Bai was developed and evaluated within the Bowen-script domains represented in its corpus; applying it elsewhere requires new linguistic analysis and routing validation.The sentence-level split also does not provide a source- or document-held-out evaluation.Source: Limitations · p. 9
  • Cycle-consistency is a heuristic whose performance depends on the coverage and biases of the reverse translator Mrev.Human evaluation covers 100 sentences and three annotators.Source: Limitations · p. 9
  • Self-evolution adds LLM calls, although adaptive routing reduces unnecessary expert computation.Source: Limitations · p. 9

Ethics Considerations

The study documents its authorized source materials and human-centered corpus workflow, while noting restrictions on manuscript access and evaluation boundaries. Appendix diagnostics describe retrieval overlap and assess routing behavior and efficiency.

Ethics Considerations

  • Community-held manuscript content will not be released under an unrestricted open license; access requests may be considered case by case with rights-holder or custodian approval.Where permissions allow, prompts, configuration files, evaluation scripts, split metadata, annotation schemas, and redacted or derived examples may be released.Source: Ethics Considerations · p. 9

A Detailed Corpus Construction and Token Statistics

  • The corpus combines an officially published Dabenqv collection and an authorized community-held handwritten manuscript, digitized, translated, sentence-aligned, and quality-checked.Native Bai speakers translated manuscript sentences, a Bai-language researcher reviewed them, and automated clerical assistance did not determine semantic adequacy or sentence-pair inclusion.Source: Ethics Considerations · p. 9, A Detailed Corpus Construction and… · p. 12

B Retrieval Overlap Diagnostics

  • Only 18 of 1,536 test sentences (1.2%) exceeded a Knowledge Base similarity threshold of 0.9, indicating limited near-duplicate overlap.Mean and median maximum similarities were 0.68 and 0.65.Source: B Retrieval Overlap Diagnostics · p. 12

C Routing and Efficiency Diagnostics

  • The routing and efficiency diagnostics report pathway selection for simple and complex inputs, agreement with human complexity labels, and token use relative to translation quality.Source: C Routing and Efficiency Diagnostics · p. 12

C.1 Pathway Selection Analysis

The appendix examines how MAS-Bai routes inputs of differing linguistic complexity, compares its computational cost and quality with Static MAS, and describes verifier comparison settings.

C.1 Pathway Selection Analysis

  • The Meta-Agent selects the efficiency-oriented pathway for 72% of simple inputs and the quality-oriented pathway for 85% of complex inputs.Simple and complex categories were based on human-informed linguistic criteria; the deeper philological pathway is concentrated on inputs with stronger Bai/Bowen cues.Source: §C.1 · p. 12, Figure A1 · p. 13
  • Against human complexity labels, routing achieved 88.5% overall accuracy, with 86.7% recall for complex inputs and 90.2% recall for simple inputs.Two Bai-language researchers labeled a random sample of 200 sentences, with Cohen’s κ = 0.84; Table A2 reports the routing metrics.Source: §C.1 · p. 12, Table A2 · p. 13

C.2 Efficiency and Quality Trade-off

  • MAS-Bai reduces inference token consumption by approximately 38% relative to Static MAS while achieving higher overall translation quality.The figure contrasts sentence-level COMET with token consumption; Static MAS always runs the full expert chain, while MAS-Bai uses two operating regions.Source: §C.2 · p. 13, Figure A2 · p. 13

D.1 Verifier Separation and Strength

  • Verifier comparisons include a disjoint setup with 4,000 retrieval pairs and a non-overlapping 1,000-pair reverse-model training set; strength settings vary Mrev training data while holding the retrieval base fixed.Table A3 reports the comparisons.Source: §D.1 · p. 13, Table A3 · p. 13

D.2 Threshold and Iteration Sensitivity

The appendix tests nearby acceptance thresholds and a larger iteration cap, then explains the selected setting as a limit on refinement calls.

D.2 Threshold and Iteration Sensitivity

  • Performance was stable for nearby values of τ when T = 3.Source: §D.2 · p. 14
  • Increasing the iteration cap to T = 5 improved BLEU, while COMET changed little.Table A4 reports sensitivity to the acceptance threshold and iteration cap.Source: §D.2 · p. 14, Table A4 · p. 14
  • The authors retained the pre-specified T = 3 setting because it limits the maximum number of refinement calls.Source: §D.2 · p. 14
Loading 2610.23205v1…