Does network complexity help organize Babels library?

In this work, we study properties of texts from the perspective of complex network theory. Words in given texts are linked by co-occurrence and transformed into networks, and we observe that these display topological properties common to other complex systems. However, there are some properties that seem to be exclusive to texts; many of these properties depend on the frequency of words in the text, while others seem to be strictly determined by the grammar. Precisely, these properties allow for a categorization of texts as either with a sense and others encoded or senseless.

💡 Research Summary

The paper investigates whether complex‑network measures can be used to distinguish meaningful texts from meaningless or ciphered ones, using the fictional setting of Borges’s Library of Babel as a motivating backdrop. The authors convert each text into a word‑co‑occurrence network: each distinct word becomes a node, and an undirected edge connects two words that appear consecutively in the text. The construction is case‑insensitive and treats punctuation as a hard break, so only true adjacency contributes to the graph.
To create a “senseless” counterpart, the original edge list is circularly permuted by a random offset τ, preserving each node’s degree (i.e., the frequency‑derived connectivity) while scrambling the actual connections. This method guarantees that any metric that depends solely on degree distribution will be identical for the original and the ciphered version, allowing the authors to isolate structural signatures that go beyond Zipf‑law frequency effects.
The study analyzes a broad corpus: classic literature (Homer’s Iliad, Cervantes’ Don Quixote, Kafka’s Metamorphosis), the Universal Declaration of Human Rights in 17 languages, a set of programming code fragments (Fortran), and the enigmatic “Vöynich manuscript.” For each text and its permuted version the authors compute (i) the degree distribution P(k), (ii) the average clustering coefficient C, and (iii) the average shortest‑path length L.
Key findings:

1. Degree distribution follows a power‑law (P(k) ∼ k⁻ᵞ) for all texts, both original and permuted. This reflects Zipf’s law and shows that degree alone cannot discriminate meaning.
2. Clustering coefficient is markedly higher in natural‑language texts than in their permuted counterparts or in formal code. High C indicates strong transitivity: words that co‑occur with a common neighbor tend also to co‑occur with each other, a property of grammatical structure.
3. Average path length remains essentially unchanged between original and permuted versions, suggesting that L by itself is not a useful discriminator.
4. The most striking result is an empirical relationship C ≈ L⁻³·⁵⁶ that holds across all natural‑language networks examined, regardless of language, genre, or size. This power‑law correlation between clustering and path length is absent in the Vöynich manuscript and in all programming‑code networks, indicating that it is a signature of “sense‑bearing” texts.
5. Table I presents detailed statistics (number of nodes W, edges E, mean degree k, C, Cc, L, Lc) for each corpus. Notably, the same declaration translated into different languages shows substantial variation in C, reflecting language‑specific syntactic patterns, yet the C–L scaling persists.
   The authors argue that the C–L correlation can serve as a first‑order filter for the Library of Babel problem: any book whose word‑network obeys the scaling is likely to contain coherent grammatical structure, whereas books that violate it (including the Vöynich manuscript) are probably nonsensical or heavily encrypted.
   Limitations are acknowledged: the network only captures immediate adjacency, ignoring longer‑range syntactic dependencies, semantic relations, or higher‑order n‑gram structures. Moreover, the circular permutation cipher is relatively simple; more sophisticated encryption (e.g., word substitution plus reordering) may disrupt the C–L relationship differently. Future work could incorporate dependency trees, semantic similarity edges, or dynamic network measures to improve robustness.
   In summary, by applying complex‑network theory to textual data, the paper identifies a novel, structure‑based metric— the clustering‑path length scaling— that distinguishes meaningful natural language from random or formal code. This contributes both a methodological tool for large‑scale text classification and a conceptual insight into how grammatical constraints manifest as topological signatures in word‑co‑occurrence networks.