POCAI Research
MLM — Meaning Language Model
Understand meaning before predicting language
Personal research project
Research proof of concept. Every claim is measured on identical input, and the cost is reported next to the benefit.
A research POC that asks what is happening, who is involved and what information is missing — then scores a Transformer's next-token candidates against that structure.
Problem & business context
A conventional language model only asks which token statistically comes next. MLM tests whether first understanding meaning — roles, places, time — makes predictions better, and measures the cost honestly.
My role & contribution
Designed and built the research prototype.
What I built or designed
- Pipeline: language, lexicon, grammar, morphology and syntax → Universal Phrase DNA → context graph → expected semantic slot → linguistic scoring of the Transformer's candidates
- Every component can be switched off for fair comparison on the same model and hardware
- Language packs, multimodal input (text, image, video) and AWS EC2 deployment