Hierarchical Compositional Hypergraphs Encode Document Structure for Classification
This paper introduces a Hierarchical Compositional Hypergraph (HCH) that encodes document structure through ordered token, sentence, and paragraph layers, demonstrating that combining these structural features with standard TF–IDF yields statistically significant improvements in text classification accuracy and macro-F1 over lexical baselines alone.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the vast landscape of computer science, there is a constant struggle to teach machines how to read. For decades, the most reliable way to help a computer understand a document has been to count words. This method, known as a "bag of words," treats a text like a jar of marbles: it counts how many red marbles (the word "game") or blue marbles (the word "team") are inside, but it ignores the order in which they were poured or how they are arranged. While this approach is surprisingly effective, it misses the architecture of human language. It cannot tell the difference between a sentence that says "the dog bit the man" and one that says "the man bit the dog," because the jar contains the exact same marbles. To solve this, researchers have turned to more complex structures, trying to capture not just the words, but the sentences, paragraphs, and the way these parts fit together to form a whole.
A team of researchers at Semnan University in Iran has proposed a new way to map this structure, one that treats a document not as a flat list of words, but as a layered construction. They call their creation a Hierarchical Compositional Hypergraph. Imagine a document as a building. The bricks are the individual words, the rooms are the sentences, the floors are the paragraphs, and the entire structure is the document itself. In their model, every single brick is tracked as a unique occurrence, and the connections between them are recorded with precision. Unlike older methods that might flatten these layers into a single, messy web, this new approach keeps the layers distinct. It uses special connections, which the researchers call hyperedges, to show exactly which words make up a sentence, which sentences make up a paragraph, and how the order of these elements matters. This allows the computer to see the blueprint of the text, not just the pile of materials.
The researchers tested this idea on a classic challenge known as the 20 Newsgroups dataset, a collection of nearly 19,000 real-world messages sorted into 20 different topics, ranging from sports to science. They stripped away any extra information that might give the computer an unfair advantage, such as email headers or footers, leaving only the pure text. They then compared their new structural map against the standard word-counting methods. The results were clear: the structural map alone was not enough to beat the traditional methods. When the computer relied only on the building blueprint without looking at the specific words, it performed worse than the simple word counters. This finding was crucial, as it ruled out the idea that structure alone could replace the need to understand vocabulary.
However, when the researchers combined the structural map with the traditional word counts, something interesting happened. The hybrid model, which used both the "what" (the words) and the "how" (the structure), performed slightly better than the best traditional method alone. In the final test, this combined approach achieved an accuracy of roughly 70 percent, edging out the standard method by a small but measurable margin. The researchers found that the most valuable part of their new structure was the order of the words. Knowing that one word follows another within a sentence provided a significant boost. In contrast, knowing that two words appeared in the same paragraph or the same sentence without regard to their order added very little new information, as these details often overlapped with what was already known.
The study concludes that while this new way of mapping text is not a magic bullet that solves all problems, it is a useful tool. It acts as a quiet complement to existing methods, adding a small layer of predictive power by respecting the way humans actually organize their thoughts into sentences and paragraphs. The researchers emphasize that this is not a replacement for the deep learning models that currently dominate the field, nor is it a perfect solution. Instead, it is a precise, interpretable addition that proves that paying attention to the shape of a document can help a computer understand it a little better. The work stands as a controlled proof that the architecture of language holds information that simple word counts miss, provided that the method used to capture it is careful not to lose the unique identity of every single word in the process.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.