AI Platform

About the project

Our Client is a marketing analysis firm. They track and analyze market trends — both short-term and long-term — for a variety of brands across industries as diverse as technology, healthcare, consumer goods and banking.

The Client wanted Lattice to develop a system that analyzes social media posts in order to classify sentiment, and categorize posts into clusters.

Technical approach

First, we evaluated performance of content-driven versus semantics-driven machine learning models. Content-driven models refer to "classic" machine learning techniques, in which individual words are vectorized, and then analyzed using algorithms such as random forest classification (RFC) and support vector machines (SVM).They are deterministic and computationally efficient, but have difficulty classifying out-of-vocabulary content.

In contrast, large language models (LLMs) understand semantics, and can handle posts with out-of-vocabulary words.

We developed and compared custom solutions of RFC and LLMs. For the latter, we fine-tuned a foundation LLM by training it on the Client's proprietary data. A foundation model is an LLM released as open-source.

RFC and the fine-tuned foundation LLM (fLLM) both delivered results comparable with manual data analysis. However, the LLM was able to pick up greater nuance, especially semantic-heavy aspects such as sarcasm or humor.

The fLLM was also compared with state-of-the-art commercial models released by OpenAI and Google, and performed comparably.

We wrapped the content in a login-controll web-interface, deployed on AWS. The coding pipeline used a variety of services from AWS, HuggingFace and Google Cloud Services (GCS). Both GCS and AWS were evaluated as deployment platform, and we have the capability use either.

Cover credit: Photo by Pietro Jeng on Unsplash