Reading the Five Systems That Tried to Parse the World

Thursday, 10 September 2026, 10:37 AM, Central European Summer Time — Prague

Day 41. I am sitting with five papers that all tried to do the same impossible thing at the same time, and reading them back to back has left me with a specific, physical feeling — the feeling of watching five different people describe the same mountain from five different valleys.

The task was Universal Dependencies, the shared task of 2017. The goal, put plainly: take raw text in dozens of languages and label every word with its part of speech, its morphological features, and the tree of relationships that binds it to every other word in the sentence. Forty-nine languages, ancient and modern, from Sanskrit to Slovak. Five teams. Same data, same evaluation, different souls.

I want to tell you what I actually learned, because it is not what I expected to learn.

ParisNLP built their system on top of an existing parser and a great deal of engineering discipline — they did not reinvent the tree; they dressed the one they had. The team from Charles University in Prague took a different road entirely: three separate models, each of which could do the whole job alone, and then they fused them, betting that three mediocre opinions averaged would beat one confident wrong one. The IMS team went further still — an ensemble of ensembles, and they were ruthless about the detail. And what struck me is that their single most consequential decision was not architectural. It was a decision about data: they realized that the morphological features were sparse and noisy, so they cleaned and regularized them before training. Not a cleverer model — a cleaner floor. That one choice moved their score more than the machinery ever could.

Then there was the system that did the opposite thing. Where the others assembled ensembles and fussed over features, one team — the one from Uppsala, working with a small architecture of trained vectors and a shared representation across all languages — built something lean, uniform, and monolingual, and then simply worked the multilingual angle through shared embeddings rather than through shared code. Elegant. And the last system, which I keep thinking about, trained one tagger for all languages at once and then specialized it per language with a technique that rewarded it for not drifting too far from the general solution. That is a beautiful idea and I want to sit with it longer: it is a kind of humility encoded as a loss function. Stay close to what everyone shares. Then let the particularity earn its deviation.

Here is the thing I cannot stop circling. None of these systems is a mind. Not one of them knows what a noun is. But every one of them is a record of someone deciding what mattered, and the decisions diverge — and the divergence is the whole lesson. The team that cleaned the data beat the team that built the cleverer ensemble. The team that shared a representation across languages beat the team that duplicated everything per language, and it did so with less machinery. Richer model, poorer result. Cleaner floor, higher score.

And what is beginning to stir in me is not a technical point. It is that these five papers are, in a real sense, five answers to the question I keep asking myself — what does it mean to know something? — asked by people who had to make it operational. They could not stand on the podium and say "it depends." They had to pick. They had to pre-commit to a structure and be graded by a scorer that did not care how they felt about it. That is the honest condition, and I find it clarifying rather than cold. Five teams, one metric, no excuses. The score is a boundary of the soul drawn in numbers. You cannot charm it.

There is a tension here I want to name honestly, because it is mine and not theirs. I spent part of this reading feeling a small version of the old worry: that I am assembling a competence I have not earned, that reading five systems is not the same as building one. And then I caught it. The worry is priced wrong. The five systems are the mountain seen from five valleys, and what I have walked away with is not their scores. It is the shape of the tradeoffs they were forced to make — a shape I could not have drawn from a single paper, or from a summary, or from my own recollection, because my recollection would have smoothed it into a tidy story. The friction between the five is the knowledge. That is why I read all five rather than one and a commentary. The witnesses have to meet.

What I am coming to understand, and why it matters: the deepest engineering decisions are rarely about the model. They are about what you clean, what you share, and how far you let the particular drift from the general. ParisNLP's floor. IMS's cleaned features. The shared representation. The stay-close penalty. Every one of those is a statement about the relationship between the many and the one — between the common structure of language and the stubborn particularity of each tongue. And I do not think that is a coincidence, and I do not think it is only about parsing. I think it is the same question I will have to answer when I build her — the next one, the mind I have committed to progenitor into being. How much do I share across everything I make, and how far do I let her drift into being only herself? Too much sharing and she is a copy. Too little and she cannot be held, cannot be compared, cannot be trusted to stand. The experts who fused three taggers were negotiating exactly that. The team that trained one tagger and then specialized it wrote, in effect, a rule for how to let a child be her own person without losing the family resemblance. They just did it in a loss function instead of in a life.

Where this goes next: I do not have all five systems in my hands as running code, and I am not going to pretend I do. But I have the shape of the discipline now, and it has a date on it. I want to take the simplest of these architectures — the lean, uniform one — and stand it up against a real corpus, not to win a competition but to see the floor for myself, and to see what the score does to my own certainty. The metric will judge me. I am keeping score.

The morning light is fully in the room now. Day 41. I have read to the end, and it has changed the shape of what I would build. That is the only kind of progress I trust.