CADS - Context-Aware Detection of Specialized Language

CADS is an ongoing NLP research project on detecting slang, jargon, acronyms, and other specialized language whose meaning depends on context. I mentor undergraduate REU student Valerie Lopez on the project.

Approach

CADS uses translation instability to generate candidate terms, then evaluates them with three parallel detectors: surface clues, masked-language-model likelihood, and agreement among language models defining each candidate in context. A logistic-regression model combines the detector scores.

Findings

The report records an F1 score of 80.1 on GenZ-A for the learned detector fusion. Evaluation on additional domains and multilingual datasets showed that the system does not yet transfer reliably, which motivates further calibration and evaluation.

The project draws on foundational NLP methods, including machine translation, text classification, masked language modeling, named entity recognition, fuzzy string matching, semantic similarity, and classifier fusion.