Skip to content

HULAT2 at MER-TRANS 2026: Governed Multi-Agent Simplification for Spanish Easy-to-Read Generation

🕒 Published (v1): 2026-07-02 16:18 UTC · Source: Arxiv · link

Why this paper was selected

LangGraph-based multi-agent pipeline for text simplification; reusable agentic NLP scaffold pattern

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

HULAT2-UC3M presents a LangGraph-based multi-agent pipeline for automatic Spanish Easy-to-Read (E2R) text simplification, submitted to the MER-TRANS 2026 shared task. The system uses parallel candidate generation with two LLMs (Gemini 2.5 Flash and RigoChat-7B-v2), Event–Condition–Action routing over internal quality signals, and controlled editing to produce accessible text. The signal-guided multi-agent architecture (SARI 44.05) outperformed a linear generate–evaluate–regenerate baseline (SARI 38.51) by ~14.4% relative, ranking 6th out of 19 Spanish submissions.

Problem

LLM-based text simplification for accessibility (Easy-to-Read) lacks mechanisms to govern output quality across multiple dimensions simultaneously—semantic preservation, factual consistency, readability, syntactic clarity, and lexical difficulty—since a single generation pass cannot reliably balance these competing constraints, especially for languages like Spanish with domain-specific E2R standards (UNE 153101:2018 EX).

Method

A stateful LangGraph workflow orchestrates specialized agents with conditional routing:

  1. Pre-analysis module: Rule-based feature extraction detects structural and factual-risk signals (sentence length, numbers, dates, negations, conditions, glossary terms).
  2. Parallel generation: Three generators run simultaneously—Generator A (Gemini 2.5 Flash, conservative Plain Language), Generator B (RigoChat-7B-v2, structurally simplified), Generator C (Gemini 2.5 Flash, E2R-oriented targeting CEFR A2).
  3. Candidate evaluation: Each candidate is scored across five signal groups: semantic/factual preservation, readability (Fernández-Huerta), syntactic clarity, lexical simplicity, and robustness.
  4. ECA router: Event–Condition–Action rules select among three strategies—V1 (best candidate unchanged), V2 (base + minimal edit from second candidate), V3 (controlled fusion of complementary candidates)—or trigger targeted retries.
  5. Merger-editor: Applies minimal controlled edits under strict semantic-risk constraints.
  6. Final evaluator: Rechecks factual consistency and semantic preservation before output.

RUN2 adds an optional lexical agent (glossary + Spanish medical lexical resources) before pre-analysis. RUN3 is a standalone baseline: RigoChat-7B-v2 with LoRA fine-tuning (trained on 126 aligned Spanish Constitution E2R articles), prompt engineering, and a generate–evaluate–regenerate loop (up to 2 regenerations), governed by Fernández-Huerta readability and MeaningBERT checks.

Key Contributions

  • LangGraph-based multi-agent workflow with parallel diverse generation strategies (conservative PL, structural simplification, E2R-oriented) for Spanish accessibility.
  • ECA-style routing over a multi-dimensional internal signal taxonomy (semantic, factual, syntactic, lexical, robustness) replacing single-metric optimization.
  • Empirical comparison between signal-guided multi-agent routing and linear generate–evaluate–regenerate baseline on the same task.
  • Demonstration that lexical resource injection (glossary, medical simplification data) does not automatically improve reference-based scores (SARI), suggesting careful calibration is needed.
  • Traceable intermediate decisions logged throughout the workflow for qualitative analysis and rule calibration.

Results

  • RUN1 (base multi-agent): BLEU-Orig 35.57, BLEU-Gold 17.57, SARI 44.0543, BERTScore 0.9263 — ranked 6th/19 in the MER-TRANS 2026 Spanish leaderboard.
  • RUN2 (+ lexical agent): SARI 43.1049 — ranked 9th/19; lexical support did not improve over RUN1.
  • RUN3 (linear LoRA baseline): SARI 38.5136, BLEU-Orig 2.40 — ranked 16th/19.
  • RUN1 vs. RUN3: +5.54 SARI absolute, ~14.4% relative improvement.
  • DoNothing baseline (identity copy) achieves highest BLEU-Orig and near-top BERTScore but lowest SARI, confirming BLEU-Orig/BERTScore are insufficient simplification metrics.

Limitations

  • Official metrics (BLEU, SARI, BERTScore) do not capture accessibility-oriented quality dimensions such as factual consistency, document coherence, terminology adequacy, or actual reader comprehension.
  • Lexical support (RUN2) did not improve reference-based scores; the interaction between glossary substitution and reference alignment is not fully explained.
  • LoRA adapter for RUN3 trained on only 169 aligned articles (126 train / 43 test), limiting generalization.
  • No human evaluation or user studies conducted; the gap between automatic scores and user-oriented adequacy remains unquantified.
  • Segment-level and document-level readability/consistency analysis deferred to future work.
  • CEFR A2 used as an operational heuristic target, not validated against actual reader comprehension.

Relevance to Agentic AI / LLM Agents

This paper is a concrete applied instance of a governed multi-agent LLM workflow where agent specialization, parallel generation, and rule-based routing replace single-model prompting for a quality-sensitive NLP task. The ECA routing paradigm—detecting signals, applying conditional rules, triggering retries or merges—is an operational blueprint for building controllable agentic pipelines where output quality must satisfy multi-dimensional constraints simultaneously. The comparison between multi-agent routing and a linear generate–evaluate–regenerate loop provides empirical evidence that structured orchestration with diverse generators outperforms sequential self-refinement, a result relevant to any agentic task requiring iterative output improvement. LangGraph's stateful, traceable execution model is highlighted as a practical framework for deploying production agentic workflows with auditable decision paths.