A workplace analytics platform plans to enhance its services by providing advanced content monitoring for workplace communication tools. The goal is to help organizations foster healthy, productive communication while mitigating risks associated with toxic or inappropriate interactions.

“Our collaboration with iMerit significantly improved our LLM’s performance. The efficiency gain and enhanced accuracy showcased their commitment to delivering exceptional solutions.”

– Head of Product 

 

Problem 

Our client is developing Large Language Models (LLMs) for toxicity detection and sentiment analysis. The models rely extensively on human feedback for supervised fine-tuning and alignment for accurate toxicity detection and nuanced sentiment analysis in complex, high-context interactions across diverse languages and cultures. The client wants to integrate active human oversight into data operations to enhance content moderation and communication monitoring.

Solution

iMerit team of expert data annotators, solution architects, and NLP specialists assessed workplace conversations, categorizing statements as healthy, neutral, or toxic to identify and flag potentially harmful content across English and Spanish. Additionally, we improved the model performance by identifying edge cases and unexpected contexts, further strengthening the content moderation strategy for the client.

Results

The iMerit team executed four workflows, analyzing over 500k interactions for toxic speech and sentiment. The client achieved an impressive 97% accuracy in identifying and addressing toxicity and negative behavior in workplace conversations while gaining a 30% efficiency boost without compromising data quality. The solution also included language detection and localized analysis, enabling a globally inclusive approach for the client.

 

 

BOTTOM LINE IMPACT

500k

Interactions Analyzed

97%

Annotation Accuracy

30%

Efficiency Boost