Why General-Purpose Language Models Struggle with Legal Text

Legal language has structural and semantic properties that general-purpose models were not trained to handle. For high-stakes legal NLP applications, the choice between fine-tuning, domain adaptation, and prompting is a consequential engineering decision — not a default.

Can You Tell If Text Was Written by an LLM?

The "em-dash debate" misses the point entirely. Detecting LLM-generated text is a challenging classification problem, and the "folk methods" people use to do it are somewhere between unreliable and useless.

Making a Monolingual Model Bilingual with Domain Adaptation

You have an English BERT model that works well on legal text. Your corpus is bilingual. Here is how domain adaptation on a parallel EU law corpus can produce a model with strong masked language modeling performance in both languages — and why legal text makes this work better than you might expect.

Your Model Learned the Wrong Distribution

A model that performs well on your test set has demonstrated one thing: it learned the training distribution. Whether that distribution matches production is a separate question, and most evaluation pipelines don't ask it.

Not Every NLP Problem Needs a Frontier Model

Frontier LLMs are remarkable. They are also generative, expensive, slow, and prone to hallucination. For a significant class of NLP tasks, a well-trained BERT-style model is more accurate, cheaper by orders of magnitude, and easier to audit.

The Label Is a Hypothesis

Label design is the most consequential modeling decision in a supervised learning project. Treat it as a mechanical annotation task and you'll build a model that learns the wrong thing with high confidence.

You Have the Data. Now What?

Data availability is not problem readiness. Organizations that move from 'we have internal data' to 'let's train a model' without a research plan tend to build technically functional systems that don't change how anyone works.

Why Your NLP Model's Accuracy Number Is Probably Misleading

Accuracy, F1, and AUC are useful summaries, but they can obscure the most important questions about whether your model is actually measuring what you think it's measuring. A case for applying construct validity thinking to NLP evaluation.

The Question Before the Model: Research Design for ML Practitioners

The most expensive ML mistakes happen before a single line of code is written. A framework for problem framing borrowed from social science methodology — and why it produces better systems than jumping straight to model selection.

Domain Adaptation Is Not Fine-Tuning: A Practical Distinction That Matters

Fine-tuning and domain adaptation are often used interchangeably, but they solve different problems and require different approaches. Getting the distinction wrong is one of the more expensive mistakes in applied NLP.