Hol-PCFG cuts parsing parameters by 99.94% and tops results in six languages

Researchers published a paper on arXiv in July 2026 describing Hol-PCFG (Holographic Neural PCFG), a new approach to unsupervised sentence parsing. It achieves the best results in six languages while cutting the number of rule-scoring parameters by 99.94% compared with baseline neural PCFG models.

What is unsupervised syntactic parsing

Syntactic parsing builds a tree showing how the words in a sentence group together: subject, predicate, noun phrases. The classic tool is the probabilistic context-free grammar (PCFG), which assigns probabilities to grammar rules, for example, "a noun phrase can consist of an article and a noun."

In the unsupervised variant, the model is not trained on ready-made annotated trees. It infers structure from raw text on its own. That matters because treebank annotation is expensive and exists for only a few dozen languages.

Modern neural PCFGs achieve strong results but rely on opaque neural network blocks to score each rule. A rule's probability is simply a number produced by a "black box," with no interpretable mathematical form.

How Hol-PCFG works

Hol-PCFG recasts rule scoring as algebraic modeling of relations. The authors borrow the idea from Holographic Embeddings (Nickel et al., 2016), a method used to score triples in knowledge graphs, where the task is to predict whether the statement "object A is linked to object B through relation R" is true.

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

Key results:

• Best unsupervised parsing scores in six languages
• 99.94% fewer rule-scoring parameters than the baseline Neural PCFG
• More stable training: less variance between runs
• Parsing of Japanese directly from characters, without morphological segmentation
• Character-level quality comparable to morpheme-based models

Why Japanese without morphology is a nontrivial result

Japanese is written without spaces between words: the text runs as a continuous stream of kanji and kana. Traditional parsers first run the text through a morphological analyzer that splits it into morphemes, and only then parse the syntax.

Hol-PCFG works directly from characters, skipping that step, and still matches the quality of models that use ready-made morpheme segmentation. For languages with rich morphology or without mature preprocessing tools, which means most of the world's languages, this approach opens the way to more universal syntactic parsers.

What this means

Hol-PCFG is a rare case where interpretability does not come at the expense of performance: replacing neural network blocks with algebraic operations on vectors shrinks the model almost 2,000-fold in parameter count while improving quality. Where model transparency, resource constraints or support for languages with minimal infrastructure matter, the approach could become a real alternative to heavy neural parsers.