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

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

HULAT2-UC3M submits three automatic runs to MER-TRANS 2026 (Spanish Easy-to-Read simplification track), with two runs built on a LangGraph-based multi-agent workflow combining Gemini 2.5 Flash and RigoChat-7B-v2, and one linear generate–evaluate–regenerate baseline. The multi-agent system uses Event–Condition–Action (ECA) routing over internal quality signals to select, edit, or merge candidate outputs. The signal-guided multi-agent approach (RUN1 SARI = 44.05) outperforms the linear baseline (RUN3 SARI = 38.51), ranking 6th of 19 Spanish submissions.

Problem

LLM-based text simplification for accessibility (Easy-to-Read) risks producing fluent but unfaithful outputs that omit information, add unsupported content, or alter numerical/factual details. Standard metrics (BLEU, SARI, BERTScore) capture only partial quality dimensions and do not measure factual consistency, readability, or user-oriented adequacy. No prior system addresses the need for governed, traceable, multi-signal routing across diverse simplification strategies in a Spanish E2R competition setting.

Method

Multi-agent workflow (RUN1/RUN2): Implemented in LangGraph with stateful nodes and conditional edges.

  1. Pre-analysis: Rule-based feature extraction detects sentence length, punctuation complexity, numbers, dates, negations, conditions, and glossary terms to characterize risk profile.
  2. Parallel generation: Three generators run concurrently — Generator A (Gemini 2.5 Flash, conservative Plain Language), Generator B (RigoChat-7B-v2, structural simplification), Generator C (Gemini 2.5 Flash, E2R-oriented targeting CEFR A2).
  3. Candidate evaluation: Each candidate is scored on five signal groups: semantic/factual preservation, readability (Fernández-Huerta), syntactic clarity, lexical simplicity (difficult-word ratio, glossary hits), and robustness (repetition, script changes).
  4. ECA routing: Rules select among V1 (best candidate as-is), V2 (base + minimal edit from alternative), or V3 (controlled merge of complementary candidates); targeted retries triggered when signals fall below thresholds.
  5. Final validation: Critical signals (factual consistency, semantic preservation, robustness) recomputed; decision and trace recorded.

RUN2 adds an optional lexical agent before pre-analysis that consults a domain glossary and Spanish medical lexical simplification resources to propose substitutions for difficult terms.

Baseline (RUN3): Linear generate–evaluate–regenerate loop using RigoChat-7B-v2 with LoRA adapter (trained on 126 aligned Spanish Constitution articles), explicit A2/E2R prompts, and up to two regeneration attempts when internal readability/MeaningBERT checks fail.

Key Contributions

  • A fully governed LangGraph multi-agent architecture for Spanish E2R generation with traceable ECA routing and parallel candidate diversification.
  • An empirically-grounded signal taxonomy (semantic/factual, readability, syntactic, lexical, robustness) operationalized as inference-time routing constraints, not post-hoc metrics.
  • A LoRA-adapted RigoChat-7B-v2 baseline with internal quality-control loop, providing a controlled linear comparison point.
  • Demonstration that signal-guided multi-agent routing outperforms linear generate–evaluate–regenerate by ~14.4% relative SARI in the same team's submissions.
  • Evidence that glossary-based lexical support does not automatically improve reference-based scores, highlighting the tension between domain lexical substitution and reference alignment.

Results

  • RUN1 (base multi-agent): BLEU-Orig 35.57, BLEU-Gold 17.57, SARI 44.05, BERTScore 0.9263 — ranked 6th / 19 Spanish submissions.
  • RUN2 (+ lexical agent): SARI 43.10 (−0.95 vs. RUN1), ranked 9th.
  • RUN3 (linear baseline): SARI 38.51, BLEU-Orig 2.40, ranked 16th.
  • RUN1 improves over RUN3 by +5.54 SARI points (~14.4% relative).
  • The DoNothing baseline achieves the highest BLEU-Orig and high BERTScore but the lowest SARI, confirming source-overlap metrics are insufficient proxies for simplification quality.

Limitations

  • Results are specific to one shared-task domain (civic/democratic participation texts) and one language (Spanish); generalizability is untested.
  • Official metrics (BLEU, SARI, BERTScore) do not capture factual consistency, document-level coherence, or actual reader comprehension — the paper explicitly notes this gap.
  • Adding lexical support (RUN2) hurt reference-based scores; segment-level and document-level analysis needed to understand when lexical resources help.
  • LoRA adapter for RUN3 trained on only 126 articles (Spanish Constitution), limiting coverage.
  • No human evaluation of output quality; user-oriented adequacy and accessibility impact remain unmeasured.
  • ECA threshold calibration and signal weights are heuristic and not systematically optimized.

Relevance to Harnesses / Meta-Harnesses

This paper is a concrete instantiation of a multi-agent harness for controlled NLG: LangGraph provides the stateful orchestration layer, while ECA rules act as the meta-level governor that routes, retries, merges, and validates outputs from specialized sub-agents. The architecture is directly analogous to meta-harness design patterns — parallel fan-out of diverse generators, signal-aggregated candidate evaluation, conditional branching, and traceable decision logs — applied to an accessibility NLP task rather than software engineering. For researchers tracking harnesses, it demonstrates how domain-specific quality signals can be embedded as hard routing constraints within a graph-based workflow, and how a governed multi-agent approach can be compared rigorously against a linear pipeline baseline within a shared evaluation setting.