Generate Transformative Customer Experiences

High-quality data for conversational chatbots

• CHALLENGE: A global technology leader that develops smart chatbots, sought to iterate on its multipurpose chatbot product and remedy some of its key weaknesses. The bot was known to occasionally produce unhelpful and inaccurate responses, and its presentation of reference information was limited in its utility.
  

  ---

• SOLUTION: We were able to quickly stand up a large-scale program to meet the client’s need for high-quality model response evaluations at incredible scale. We worked with the client to implement their guidelines. Thus far, this program has netted over 3 million response evaluations for accuracy, helpfulness, and reference attribution.
  

  ---

• RESULTS: The client reduced their time to deployment of their clients’ chatbots with accurate and intent-specific data and an improved user experience.  They continue to fine-tunes the model using our evaluation data to improve upon the chatbot’s responses. 

High-quality dialog summarization models for global conversations

• EXPECTATION: Our client sought to build tools to help their clients efficiently absorb and distill large volumes of information shared via spoken dialog in global languages. How can the client build a model that is capable of summarizing natural dialog, generating comprehensive prose summaries, and effectively retaining important information from conversations, and performing effectively for speakers in various countries and regions?
  

  ---

• SOLUTION:  The client required realistic and representative samples of natural dialog, with summaries, in order to build a solution that would work well for English speakers across countries and regions. Accordingly, we collected spoken and chat-based conversations from an AI Training Specialist pool based in locales across the US, UK, India and the Philippines. We delivered over 200 hours of audio with transcriptions and summaries and more than 6,000 naturalistic SMS conversations and summaries to the client.
  

  ---

• RESULT: The client was impressed with the diversity and quality of the dataset. The client was able to use the data to develop high performing dialog summarization models, which they were able to offer to their client organizations to improve operational efficiency.

Transforming online shopping: fine-tuning LLMs for personalized product assistance

CHALLENGE: Frequently, online shoppers stall or abandon a purchase because they cannot find a product or determine whether it will meet their needs. Our client, a global ecommerce leader, seeks to launch an expert shopping assistant that can efficiently and continuously digest their rapidly evolving catalog, answer their customers’ questions effectively, and appropriately represent their brand.

SOLUTION: The client required a dataset of realistic questions and answers about products in their catalog, along with key metadata including product category, shopping stage, and reference URLs. Leveraging Appen RLHF and a diverse team of US-based AI Training Specialists, we delivered 112k product catalog prompts and responses along with requested metadata.

RESULTS: The client noted that they were impressed with our linguists’ edits and additions to the guidelines, which resulted in high quality prompts and responses. The client used the dataset for to fine-tune their LLM for product catalog inquiries and were able to achieve significant improvements in performance across stages of the shopping journey as well as alignment of the model’s responses with their brand voice.

Annotation Tools Powered by Machine Learning

Data annotation faster and at scale with machine-learning assisted annotation tools that provide a comprehensive data-labeling solution. Machine learning assistance is built into our industry leading annotation tools to save customers time, effort, and money – delivering high-quality training data and accelerating the ROI on your AI initiatives.

Robust testing and optimization

Introducing dynamic elements to ensure performance more closely reflects real-world deployment environments.