Conversational AI Localization, Language Engineering
Virtual Assistant Prompt Classification and Language Pair Creation
The Challenge
After creating their virtual assistant, our client, one of the largest global communication and technology companies, approached DataForce seeking high-quality prompt training data. The virtual assistant was meant to mimic a human by connecting and conversing with customers across different touchpoints, digitally. However, it needed to be trained to incorporate the company’s internal corporate nomenclature and taxonomy, operating procedures, and contractual information. In order to accomplish this task, DataForce needed to deliver high-quality, original material while balancing both compliant and non-compliant guidelines, including ADA compliance, in the training and deliverables.
• • • •The Solution• • • •
To train the virtual assistant, DataForce generated paired input prompts and output responses from the assistant’s point of view. Inputs were non-compliant based on the customers’ questions while outputs were compliant with the guidelines provided by the client.
- Data: The number of input-output pairs totaled more than 15,000 and covered two primary channels—web chat and mobile.
- Guidelines: Our client requested coverage of all guideline types, including greetings, apologies, questions, and statements.
- Intent: We identified key consumer group intent types per the client’s specifications while also using and learning their internal taxonomy as a reference. There were nine common intent types, including “Add a Line,” “Change Plan,” and “Upgrade.”
- Staffing: This engagement included a dedicated project manager, linguists, creative writers, and English copywriters based in the United States.
Due to the size and dynamic nature of the input-output generation, DataForce delivered the completed project in four milestones with the final delivery made ahead of schedule:
- Milestone 1 = 10% accepted batch (POC)
- Milestone 2 = 30% accepted batch
- Milestone 3 = 30% accepted batch
- Milestone 4 = 30% final batch delivery
To ensure that quality assurance (QA) was measured properly and that redundancy of prompt generation was mitigated, DataForce requested feedback from the client’s Product Lead before commencing each subsequent milestone.
Outcome: The client performed an internal QA on procedures and comparisons and found the final deliverable had 98% correct phrasing marked as compliant/non-compliant, exceeding the original goal of 90%.
DataForce has a global community of over 1,000,000 members from around the globe and linguistic experts in over 250 languages. DataForce is its own platform but can also use client or third-party tools. This way, your data is always under control.