TLDR: A new web application called “Claim Advisor” uses generative AI, including in-context learning and fine-tuning of LLMs like Phi-3, to accelerate the creation of product claims. It offers semantic search for existing claims, generates and optimizes new claims, and ranks them using synthetic consumer simulations, significantly improving efficiency and claim appeal compared to traditional methods.
The process of creating compelling product claims is a cornerstone of consumer marketing, directly influencing purchasing decisions and building brand trust. However, this task is often time-consuming, expensive, and requires rigorous scientific and legal substantiation. A new research paper introduces “Claim Advisor,” a web application designed to significantly accelerate and enhance the creation of product claims using generative artificial intelligence (AI).
Traditionally, developing product claims involves extensive manual searching of existing claims and visuals, followed by multiple rounds of consumer testing to gauge appeal. This iterative process can span weeks or months and demand substantial financial resources. The paper highlights that many product advertisements often lack sufficient scientific citations, underscoring a critical need for more credible and evidence-based strategies in claim creation.
Claim Advisor leverages the power of large language models (LLMs) to address these challenges. It functions as a minimum-viable-prototype (MVP) web application with three core capabilities:
Semantic Search and Identification
The application can semantically search and identify existing claims and visuals that resonate with consumer preferences. This is achieved using text embedding techniques, like OpenAI’s TEXT-EMBEDDING-ADA-002 model, to compute semantic similarity. For visuals, it integrates the CLIP model, which aligns text and image embeddings, allowing for multimodal searches. Users can even combine text and image queries, controlling the influence of each component. This feature allows companies to quickly repurpose legally approved claims and visuals from their “Claim Log” data or identify promising candidates from “MaxDiff” studies, which are consumer preference assessments.
Claim Generation and Optimization
Claim Advisor can generate and optimize new claims based on a product description and a consumer profile. It employs prompt engineering and in-context learning, incorporating insights from past market research and MaxDiff studies. The system learns from examples of successful claims, either based on their performance scores or semantic similarity, to synthesize new, optimized claims. Through an iterative process, the application has shown remarkable improvements, with claims generated in the third round achieving 100% “highly appealing” performance, a significant leap from the 20% achieved by human-designed claims in the first round.
Also Read:
- Enhancing Language Models with Specialized Knowledge through Iterative Reinforcement Learning
- Evaluating AI’s Role in Connecting Patients to Clinical Trials
Claim Ranking and Simulation
To virtually screen claims before costly actual market research, Claim Advisor can rank generated or manually created claims using simulations via synthetic consumers. This is done by fine-tuning a lightweight version of Microsoft’s Phi-3 model using Low-Rank Adaptation (LoRA) techniques. The model is trained to mimic the MaxDiff process, where it selects the “best” and “worst” claims from a small set. This approach proved more effective than asking LLMs to rank all claims in a single pass, aligning better with real-world consumer preferences. The fine-tuned Phi-3 models, particularly the 14B parameter version, consistently outperformed larger commercial models like ChatGPT-3.5, GPT-4, and GPT-4o in ranking accuracy and top-N coverage, even with fewer examples.
The research highlights that while prompt engineering is crucial for guiding LLMs, excessive instruction can sometimes reduce output diversity. The study also emphasizes the benefits of using open-source models like Phi-3, offering greater control, transparency, and lower inference costs compared to proprietary alternatives. Despite some limitations, such as the use of proprietary datasets and prompts, the authors have shared their codebase and dummy data files to encourage further research and reapplication of this innovative pipeline across various industries. You can find more details about this work in the full research paper available here.
Claim Advisor has been successfully deployed as an MVP web application within a consumer packaged goods (CPG) company, demonstrating its potential to significantly enhance the efficiency and creativity of product researchers in developing impactful product claims.
