A Mind Map-Based Evaluation Method for the Structural Creativity of Large Language Models

This article proposes a mind map-based evaluation method for the structural creativity of large language models (LLMs), aiming to solve the problem that traditional evaluation methods are difficult to systematically and objectively measure AI creativity. By converting LLM-generated text into mind maps, this method conducts quantitative evaluation from dimensions such as structural complexity, connection density, innovation index, and semantic coherence, providing a new perspective and tool for AI creativity evaluation. It has advantages like structuredness and strong interpretability, while also having limitations such as conversion accuracy issues.

With the improvement of LLM capabilities, accurately evaluating their creativity has become an important issue. Traditional methods focus on text fluency, grammatical correctness, etc., but lack systematic and objective measurement of creativity. Creativity includes multiple dimensions such as fluency, flexibility, and originality. Existing evaluations have problems like high cost of manual judgment and automatic metrics (e.g., BLEU) being difficult to capture the essence of creativity. Therefore, there is a need to develop new structural creativity evaluation methods.

Mind maps were proposed by Tony Buzan, with a central theme as the core, organizing information through branch structures, and having characteristics such as hierarchical structure, connection links, keyword focus, and visual diversity. The advantages of applying them to AI evaluation include: the structured nature allows independent analysis of creativity dimensions, visualization facilitates understanding and verification, and the hierarchical structure is suitable for evaluating content organization and logic.

The evaluation framework includes text-to-mind map conversion (theme identification, hierarchy extraction, connection discovery, node generation, implemented using NLP technology), and four evaluation dimensions: 1. Structural complexity (number of nodes, number of branches, depth, branch balance); 2. Connection density (number of internal/cross connections, connection diversity); 3. Innovation index (proportion of novel nodes, proportion of unique connections, concept fusion degree); 4. Semantic coherence (theme consistency, logical fluency, global coherence). The evaluation process is: prompt design and content generation → automatic mind map construction → multi-dimensional index calculation → comprehensive scoring and ranking.

Compared with traditional methods, the advantages of this method include: structured evaluation (capturing three-dimensional structures), strong interpretability (visualized results are easy to understand), multi-dimensional quantification (avoiding the limitations of single indicators), and human-machine collaboration (combining machine efficiency with human judgment). It has broad application prospects, and weights can be adjusted to meet the needs of different scenarios.

The method has limitations such as conversion accuracy (errors may occur in text-to-mind map conversion), domain adaptability (need to adjust indicators for different domains), dynamic evaluation (current static analysis, need to explore dynamic trajectories), and cross-modal expansion (can be extended to multi-modal scenarios). In the future, it is necessary to improve conversion algorithms, develop domain-adaptive methods, and explore dynamic evaluation and cross-modal expansion.

The mind map-based evaluation method provides a new tool and perspective for AI creativity research, converting abstract creativity into quantifiable structural indicators, improving the objectivity and repeatability of evaluation, and helping to understand the creative thinking process of AI. With the development of LLMs, this method is expected to promote further development in the field of AI creativity evaluation.