Skip to content

Chinese Short-Form Creative Content Generation via Explanation-Oriented Multi-Objective Optimization

🕒 Published (v1): 2025-11-19 13:05 UTC · Source: Arxiv · link

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

MAGIC-HMO formalizes Chinese short-form creative NLG (e.g., baby naming) as a Heterogeneous Multi-objective Optimization (HMO) problem, jointly optimizing explicit user-specified objectives (EUOs) and implicit explanation reliability objectives (IROs). A training-free three-agent pipeline—manager (MOM), generator (MOG), evaluator (MOE)—iteratively refines both the creative output and its accompanying explanations. The core insight, grounded in Brunswik's Lens Model, is that for extremely brief outputs, structured explanations must serve as observable cues for constraint verification.

Problem

Outcome-oriented multi-aspect controllable generation methods assume the generated text itself provides sufficient signal for verifying objective satisfaction. This assumption collapses for Chinese short-form CNLG tasks (names, slogans) where outputs are so compact that multiple latent user constraints (cultural significance, Bazi & Wuxing compatibility, parental expectations) cannot be inferred from the surface form alone, making reliable multi-objective evaluation and optimization impossible without supplementary cues.

Method

The task is formalized as HMO: jointly maximize \(F^{\text{EUO}}(y, e \mid U; \theta_t)\) and \(F^{\text{IRO}}(y, e \mid U; \theta_t)\), where \(y\) is the short-form output and \(e = \{e_1, \ldots, e_m\}\) are per-constraint explanations serving as additional observable cues.

Stage 1 — Information Enhancement (MOM + MOE): - MOM performs task-type identification, EUO parsing with preference weight estimation \(W_{\text{exp}}\), and key-information extraction. - Retrieval: MOM issues a style-reshaped embedding query over a knowledge base (e.g., classical Chinese poetry corpus CPoetry); MOE reranks candidates and revises the query iteratively until an optimal \(I_{\text{rk}}\) is found. - MOM constructs per-EUO descriptions \(I_{\text{desc}}\) and fine-grained explanatory requirements \(I_{\text{reqs}}\); MOE evaluates and refines them.

Stage 2 — Explanation-oriented Dynamic Iterative Optimization (MOG + MOE): - MOG generates draft \((y, e)\) conditioned on user query \(U\) and expanded information \(I\). - MOE evaluates IROs first: factuality via rule-based checking against \(I\); completeness and clarity scored 0–2 using \(I_{\text{reqs}}\). Aggregate implicit score: \(\theta_{\text{imp}} = W_{\text{imp}} \cdot S_{\text{imp}}\). - Unreliable explanations trigger corrective feedback to MOG; once IROs pass threshold, MOE scores EUOs with weighted aggregation. Iteration continues until convergence or a maximum iteration budget.

Key Contributions

  • First formalization of Chinese short-form CNLG as HMO, distinguishing EUOs and IROs as heterogeneous objective types.
  • MAGIC-HMO: training-free three-agent framework (MOM/MOG/MOE) with explanation-oriented iterative optimization and dynamic IRO-first evaluation ordering.
  • CBNames benchmark for the Chinese Baby Naming task.
  • CPoetry dataset (classical Chinese poetry) for knowledge-grounded retrieval to mitigate domain hallucination.
  • New evaluation metrics designed for short-form multi-objective creative generation.

Results

  • MAGIC-HMO significantly outperforms six strong baselines (including decoding-based methods MacLaSa, COLD, MUCOCO; fine-tuning; reinforcement learning; and agent-based LLM-D) on the CBNames benchmark across multiple LLM backbones.
  • (Note: specific numeric scores are not included in the provided paper excerpt.)

Limitations

  • Evaluation is confined to a single task domain (Chinese Baby Naming); generalization to other short-form CNLG tasks (advertising slogans, brand names) is claimed but not empirically demonstrated in the provided text.
  • IRO factuality checking is rule-based and depends on the quality and coverage of the retrieved knowledge base; gaps in CPoetry propagate to hallucinated explanations.
  • The iterative pipeline incurs multiple LLM calls per output; latency and token cost scale with the number of objectives and refinement iterations.
  • Explanation reliability does not guarantee EUO satisfaction—MOE's EUO scoring is still LLM-based and subject to the same hallucination risks the framework seeks to mitigate.

Relevance to Harnesses / Meta-Harnesses

MAGIC-HMO is a three-role agent harness (manager → generator → evaluator) with a formal iterative feedback loop, instantiating a canonical harness pattern: a meta-level coordinator (MOM) that decomposes a task, routes structured context to a worker (MOG), and triggers a critic (MOE) whose scored feedback drives refinement cycles. The explicit termination criterion (threshold on \(\theta_{\text{imp}}\) and \(\theta_{\text{exp}}\)) and the IRO-first evaluation ordering are harness-design choices that directly control loop convergence—a concern central to meta-harness engineering. The framework also demonstrates how domain-specific knowledge retrieval can be embedded as a sub-pipeline within a harness, and how heterogeneous objective types require staged (not simultaneous) evaluation, a lesson applicable to multi-objective harness design in other domains.