Structural analysis of opinion columns using ordinal patterns and information-theoretic measures

Source attribution in journalistic text is typically approached through semantic representations, such as term weighting or neural embeddings. These approaches capture topical content, but usually overlook how a text is organized as a sequence. Here, we show whether content-independent ordinal descriptors of such sequences can identify the source that published an opinion column. We converted 1262 columns from 14 Peruvian online newspapers into numerical sequences using two encodings, word length and lexical frequency, and an exploratory part-of-speech (POS) encoding. From non-overlapping blocks with D = 3, we calculated pattern and transition probabilities, permutation entropy, disequilibrium, and statistical complexity. The apparent dominance of pattern 012 is mostly a tie-breaking artifact whose size depends on stopword removal. Under random tie-breaking, its share falls from 23.7% and 48.1% to about 17%, and K-means clusters barely match the newspapers (ARI ≤0.05). Classified against the true sources, the lexical encodings reach accuracies of 0.18–0.29, above the 0.08 majority baseline but below a function-word baseline (0.43), and neither shuffling tokens within each document nor retaining stopwords changes them. Their signal lies in the marginal distribution of the encoded values and in column length, not in word order. Only POS-tag order outperformed all 20 shuffled surrogates with both classifiers, although the effect is small. Therefore, tie handling, surrogate tests, and targets that do not come from the features themselves are needed before ordinal descriptors of text are read as stylistic fingerprints.