TL;DR
Adapting classical literature for children requires a delicate balance of simplifying language while preserving vivid character traits and concise storylines. This paper introduces InstructChild, a framework that fine-tunes large language models (LLMs) using character personality profiles and narrative triplets, optimized via a novel Chinese readability metric and lookahead decoding. The authors also release Classic4Children, the first parallel dataset mapped from the Four Great Classical Novels of Chinese literature to their child-friendly counterparts.
What Problem Does This Paper Address?
Chinese literary classics (like Journey to the West) are rich in cultural value but are written in classical or semi-classical Chinese, featuring intricate plots and archaic vocabulary that are highly inaccessible to children. Rewriting these text fragments manually is a labor-intensive process.
While modern LLMs excel at translation and text simplification, they fail at Child-friendly Literary Adaptation (CLA). Standard models tend to generate dry, overly verbose explanations that are often longer than the original text, completely ignoring children's reading preferences: vivid character portrayals, concise narrative structures, and an age-appropriate reading level.
Methods
The authors propose InstructChild, a three-stage adaptation pipeline designed to align LLM outputs with children's cognitive and reading preferences:
- Integrative Instruction Tuning: To prevent the model from losing character nuances during simplification, the authors use GPT-4o to extract character personalities based on the Big Five Personality Traits theory along with narrative structure (entity-relation triplets). These profiles are prepended to the original text to fine-tune a base model (Qwen2-7B-Instruct) using Low-Rank Adaptation (LoRA).
- Refinement via Readability Reward (Red-CN): The authors designed a custom metric, Red-CN, which evaluates Chinese text complexity based on the proportion of adverbs/conjunctions, character frequency distributions, and text length. Using Direct Preference Optimization (DPO), they reinforce the model to target the "Goldilocks zone" of readability found in expert-written children's books.
- Lookahead Decoding: During inference, standard token selection is augmented by a lookahead strategy. The model projects $n$ tokens into the future across $L$ candidate paths, selecting the current token that maximizes the long-term Red-CN readability score.
Datasets
The authors introduced the Classic4Children dataset, constructed from the Four Great Classical Novels of Chinese Literature:
- Journey to the West
- Romance of the Three Kingdoms
- Water Margin
- Dream of the Red Chamber
The dataset contains 2,686 training pairs and 300 testing pairs. It was built by manually aligning paragraph fragments from the original classical texts with widely acclaimed, expert-written children's editions.
Key Insights and Ideas
- More Parameters $\neq$ Better Adaptation: Massive proprietary models like GPT-4o and GLM-4 actually underperform on child-directed tasks because their default behavior is to explain comprehensively. This leads to verbose outputs that reduce readability for children.
- Readability is Non-Linear: Simplifying text isn't just about removing rare words. The authors discovered that children's literature maintains a highly specific, clustered density of adverbs and conjunctions. They successfully modeled this as a normalized Gaussian distribution to penalize text that is either too complex or too simple.
- Psychological Anchoring Helps Consistency: Explicitly feeding the model a character's psychological profile (e.g., "openness: 5, innovative and adventurous") acts as a strong behavioral anchor, helping the model maintain distinct, vivid dialogue voices even through heavy text simplification.
Key Results
InstructChild significantly outperformed both closed-source commercial APIs and open-source fine-tuned models across automatic and human evaluations.
Automatic Evaluation
As shown in the paper's main results, InstructChild achieved the highest scores in lexical similarity and readability:
- BLEU-1 / BLEU-2: Achieved 22.93 / 13.09, outperforming GPT-4o (20.91 / 10.42).
- Red-CN (Readability Metric): Reached 73.12, significantly higher than standard open-source fine-tuning baselines like Qwen2-7B (67.31) and commercial models like GLM-4 (61.95).
Human Evaluation
Five independent evaluators scored the models from 0 to 2 on key criteria. InstructChild dominated in child-centric metrics:
- Character Clarity (CC): Scored 1.83, compared to GPT-4o's 1.52.
- Narrative Efficiency (NE): Scored 1.78, vastly outperforming GPT-4o's 0.95, proving it successfully strips away confusing, redundant subplots.
Limitations and Future Work
Despite its strong performance, the authors note a few hurdles for real-world deployment:
- Computational Overhead: The lookahead decoding strategy requires generating multiple future tokens for several candidate paths at every single step, which significantly slows down inference speed. The authors suggest knowledge distillation as a future fix.
- Context Length Constraints: The current dataset operates on paragraph fragments rather than full chapters due to LLM input limitations, which can occasionally cause the model to miss broader narrative arcs.
- Oversimplification Risks: Drastically reducing character relationships can sometimes introduce minor narrative misunderstandings, a trade-off that requires finer balancing in future iterations.
Summary updated