Before using HumanFirst, LimeChat managed their NLU data using spreadsheets. Additionally, LimeChat developed and improved the NLU of Rasa projects using manual techniques.
LimeChat needed a solution to centralize their projects’ NLU data. This centralization would allow the development of re-usable catalogs of intent classifiers. This would help LimeChat speed up the development of projects over time.
LimeChat also wanted to move away from manual data-labeling workflows. It wanted an active learning data-labeling solution that allowed collaboration and conversation-driven development.
Since implementing HumanFirst in their workflow, LimeChat accelerated its conversation-driven development by 8.5x.
To learn more on conversation-driven development see Rasa’s description here.
LimeChat centralized their existing NLU projects’ data using HumanFirst. This allowed them to improve their projects’ NLU at scale. HumanFirst allows them to merge ambiguous intents and split intents to more granular ones.
Using their improved and centralized NLU, LimeChat developed catalogs of intents. LimeChat developed three types of catalogs; core, industry-specific and company-specific. These catalogs are re-used accross their projects. This helps LimeChat speed up development of new Rasa projects.
LimeChat’s catalogs are now improved over time using HumanFirst’s active-learning workflows and raw data generated from their deployed bots.
HumanFirst’s CLI allows LimeChat to sync utterance data from Rasa to HumanFirst. This allows LimeChat to discover new intents and improve existing ones. HumanFirst allowed LimeChat to collaborate on this activity by inviting team members to the platform.
HumanFirst helped LimeChat speed up its conversation-driven development by 8.5x.
LimeChat’s labeling workflow now takes a total of 2.3s per utterance; a workflow that used to take them 19.4s to complete prior to using HumanFirst.
- Efficiency : measured using time spent per utterance
- Data Labeling : time spent labeling an unlabeled utterance to the right intent
- Rasa Sync : syncing NLU improved in HumanFirst back to Rasa (instantaneous with HumanFirst’s CLI)
“We’ve been using Rasa ever since we started LimeChat. HumanFirst’s integration with Rasa has allowed us to continuously improve our Rasa projects with conversational data generated from our deployed bots. Thanks to HumanFirst’s active-learning workflows and collaborative solution, we can now ship new projects and improve existing ones in record time.”
— Nikhil Gupta, CEO @ LimeChat
How HumanFirst speeds up conversation-driven development
HumanFirst allows users to upload raw conversational data generated from various sources. Once data added, HumanFirst utilizes active learning workflows to speed up conversation-driven development.
HumanFirst allows you to cluster similar utterances together. This helps accelerate the selection and labeling of utterances to their appropriate intent.
This semantic clustering can also be used within intents themselves. This helps refactor/split training data to more granular intents at scale. Doing this helps develop robust intent hierarchies.
HumanFirst’s similarity search allows you to train or improve intents at scale. HumanFirst uses their own nearest neighbour algorithm to suggest utterances that fit the intent in question.
Syncing back to Rasa
HumanFirst’s CLI integration makes syncing your work back to Rasa instantaneous. To learn more about HumanFirst’s CLI integration with Rasa click here.
LimeChat is the world’s first level-3 AI chatbot company with a directed focus to boost sales for online D2C brands. Their smart chatbot offers personalized and contextual conversational experiences to customers wherever they are in their buying journey. With integrations across CRMs, store management platforms, payments networks, and logistics platforms, they take great care of end-to-end customer interactions for their client brands. The data they collect, coupled with their data analysis engine, is invaluable for online brands towards their marketing and outreach campaigns.