Recontextualizing Famous Quotes for Brand Slogan Generation

Technical Summary: Recontextualizing Famous Quotes for Brand Slogan Generation

Problem Statement
The paper addresses the challenge of generating novel, creative, and insightful brand slogans. While Large Language Models (LLMs) have been applied to this task, existing approaches suffer from two primary limitations:

1. Stylistic Redundancy: Outputs often lack a distinct brand persona, resulting in interchangeable slogans that share similar phrasing patterns (e.g., generic motivational phrases).
2. Machine-Generated Aesthetic: Slogans frequently appear overtly machine-produced, lacking the originality, creativity, and insightful impact required for effective advertising.
   The authors attribute these issues to the data-driven nature of current methods, which learn directly from existing slogan corpora. This leads to outputs that mirror training data, suffer from "advertising fatigue" due to repetition, and fail to balance novelty with the familiarity necessary for memorability.

Methodology
The authors propose a modular framework that shifts the generation paradigm from learning from slogans to recontextualizing persona-related famous quotes. The process is decomposed into four interpretable subtasks, rather than relying on an end-to-end LLM generation:

1. Post-Training with Short Text Corpus: A base LLM (QwQ-32B) is fine-tuned using parameter-efficient LoRA on a diverse corpus of short texts, including famous quotes, slang, song lyrics, and memes. This step "reawakens" the model's capacity for generating concise, memorable, and rhetorically diverse text.
2. Quote Matching: The system prompts the LLM to retrieve multiple candidate famous quotes that align with a specific target brand persona (e.g., Joy, Fear, Pride). This ensures the source material possesses the necessary emotional resonance and cultural familiarity.
3. Structure Breakdown: The selected quote is decomposed into fixed segments (preserving the original syntax and rhetorical pattern to maintain recognizability) and editable segments (flexible components for brand integration).
4. Vocabulary Replacement: The system generates replacements for the editable segments while strictly adhering to constraints:
   • Minimal substitution count.
   • Preservation of original cadence and grammatical roles (e.g., noun-to-noun).
   • Maintenance of rhythm and fluency.
5. Remix Generation & Refinement: The LLM integrates the preserved structure with the new vocabulary to generate the final slogan. A refinement step follows to verify logical coherence, semantic clarity, and content safety, discarding outputs that fail these checks.

Key Contributions

• Conceptual Paradigm: The paper introduces a novel approach that remixes famous quotes with brand products, expanding the creative scope beyond existing slogan corpora to inject insightfulness and rhetorical depth.
• Technical Framework: A systematic, modular pipeline is proposed that decomposes slogan generation into controllable subtasks (matching, decomposition, replacement, remixing). This design offers fine-grained control over creativity and structure, ensuring traceability and high-quality, persona-aligned outputs.
• Empirical Validation: The authors provide extensive automatic and human evaluations demonstrating the framework's effectiveness against state-of-the-art LLM baselines.

Results
The proposed method was evaluated against three baselines: GPT-4o, DeepSeek-R1-Distill-LLaMA-70B (DS-L), and DeepSeek-R1-Distill-Qwen-32B (DS-Q) across 40 brands and 5 personas.

• Diversity: The remix-based method achieved significantly higher diversity scores. It attained a Distinct-2 score of 0.840, outperforming GPT-4o (0.585) and DS-L (0.680). It also showed lower Pairwise-BLEU and Self-BLEU scores, indicating less repetition and greater variation among generated slogans.
• Novelty: The method demonstrated stronger novelty, evidenced by lower "Disfluency" and "Unfaithfulness" scores (metrics where lower values indicate a greater departure from conventional expressions memorized by the model).
• Emotional Impact (Hook Strength): In pairwise comparisons using an LLM-as-a-Judge, the proposed method consistently outperformed baselines in "first-glance" emotional impact, particularly in beauty, nutrition, and electronics domains.
• Human Preference: In a crowd-sourced study (Prolific), the method received the highest selection rate (29.5%) for "Hits you first" (immediate attention), surpassing all baselines. However, for "Persona-fit best," the DS-Q baseline was slightly preferred (30.5% vs. 21.5% for the proposed method), suggesting a trade-off between immediate hook strength and precise persona alignment.

Significance
The paper claims that moving beyond slogan-centric training data is crucial for overcoming advertising fatigue. By leveraging the inherent rhetorical structure and cultural resonance of famous quotes, the proposed framework generates slogans that strike a balance between novelty and familiarity. The authors argue that this approach not only mitigates the repetitive nature of data-driven generation but also produces slogans that are more memorable, engaging, and emotionally salient, offering a practical and interpretable solution for real-world advertising systems.