Hol-PCFG cut syntactic parsing parameters by 99.94% and took the lead in six languages

Researchers published a paper on arXiv in July 2026 about Hol-PCFG (Holographic Neural PCFG) — a novel approach to unsupervised syntactic parsing that achieves state-of-the-art results on six languages while reducing the number of parameters for evaluating grammatical rules by 99.94% compared to baseline neural PCFG models.

What is Unsupervised Syntactic Parsing

Syntactic parsing is the construction of a parse tree that shows how words in a sentence combine into groups: subject, predicate, noun phrases. A classic tool is probabilistic context-free grammars (PCFG). They assign probabilities to grammatical rules: for example, "a noun phrase can consist of an article and a noun."

In the unsupervised variant, the model is not trained on pre-annotated parse trees — it derives the structure from raw text on its own. This is valuable: syntactic parse tree annotation is expensive and exists for only a few dozen languages.

Modern neural PCFGs achieve high performance, but use opaque neural network modules to evaluate each rule. A rule's probability is simply a number produced by a "black box," without an interpretable mathematical form.

How Hol-PCFG Works

Hol-PCFG translates the rule evaluation task into algebraic modeling of relations. The authors draw on the idea from Holographic Embeddings (Nickel et al., 2016) — it was used to evaluate triples in knowledge graphs, where one must predict whether the statement "object A is related to object B through relation R" is true.

In the new model, each nonterminal of the grammar — symbols like NP, VP, S — is represented as an embedding vector constrained to the surface of a torus. The probability of the rule "S → NP VP" is computed via circular correlation of vectors for the left and right child symbols. This gives each rule a closed mathematical form that explicitly reflects the grammar structure, rather than a neural network output.

Key results:

• State-of-the-art unsupervised syntactic parsing performance on six languages
• 99.94% reduction in parameters for rule evaluation compared to baseline Neural PCFG
• More stable training: lower variance across runs
• Japanese parsing directly from characters — without morphological segmentation
• Character-level quality comparable to morpheme-based models

Why Japanese Without Morphology is a Non-trivial Result

Japanese is written without spaces between words: text flows as a continuous stream of kanji and kana characters. Traditional parsers first run the text through a morphological analyzer that segments it into morphemes, and only then perform syntactic parsing.

Hol-PCFG works directly from characters, bypassing this step, and maintains quality comparable to models using pre-segmented morphemes. For languages with rich morphology or without mature preprocessing infrastructure — which is the majority of the world's languages — this approach opens a path toward more universal syntactic parsers.

What This Means

Hol-PCFG is a rare example where interpretability does not come at the cost of performance: replacing neural network modules with algebraic operations on vectors compresses the model by nearly 2000x in parameter count while simultaneously improving quality. For tasks where model transparency, computational constraints, or support for languages with minimal infrastructure are important, this approach could become a real alternative to heavy neural parsers.