Abstract
This article is devoted to the problem of defining a genre in computer linguistics and searching for parameters that could formalize the concept of a genre. All kinds of existing typologies of genres rely on different types of features, whereas in the practice of NLP, any modern applications are adapted to learning on big data, and therefore - on text features that do not require additional non-automatic markup. Based on such text-internal features, in this article we focus on differentiation of various genres and their grouping on the basis of a similar distribution of features. The description of the contribution of various types of features to the final result and their interpretation are given, and also an analysis of how such features can be used to further adaptation of NLP models is provided. The materials of the "Taiga" corpus with genre annotation are used as experimental data.
Introduction
In the field of Natural Language Processing (NLP), the concept of "genre" often acts as a "ghost in the machine." While its impact on model performance is undeniable—for example, a morphological parser trained on standard literary corpora frequently fails when applied to tweets or search queries—a formal, computational definition remains elusive. Traditional genre typologies rely on manual, expensive, and often static labels.
This research proposes a shift toward "systematics": formalizing genre based on text-internal features that can be extracted automatically. By moving away from rigid human-defined categories and toward "genre clusters"—groups of texts sharing similar linguistic distributions—we can better adapt machine learning models to diverse content types.
Methodology and Data
The study utilizes the "Taiga" corpus, a massive open-source webcorpus containing 5 billion words of Russian. The experimental dataset comprised 78,000 texts across 39 prose and 39 poetry genres, alongside an unlabeled set of 40,000 texts from news, Twitter, and literary sources.
To classify these texts, a Support Vector Machine (SVM) was employed using six feature categories:
- Lexical: Word lemmata and n-grams.
- Morphological: POS tags and morphotag n-grams.
- Syntactic: Frequency of connections and properties of syntactic graphs (e.g., density, diameter).
- Readability: Metrics like Flesch-Kincaid.
- Structural: Text/word length, paragraph counts, and social media markers (hashtags/smilies).
- Symbolic: Character-level n-grams (1-5 characters).
Results and Discussion: The Power of Genre Clusters
The analysis revealed that automated features are highly effective at identifying broad genre clusters. Notably, symbolic (character-level) features proved surprisingly powerful, distinguishing children's stories with 100% accuracy and detective fiction with 76%. This suggests that the "texture" of characters provides a strong genre signal.
While syntactic features struggled with subtle sub-genre distinctions (e.g., "novels" vs. "short stories"), they were excellent at separating broad categories. Using only syntactic graph properties, the system clustered news, fiction, and social media with 88% accuracy. Broad clustering on web sources achieved near-perfect F1-scores for Twitter (1.00) and Fiction (0.98), though overlap remained between News and Popular Science.
Conclusion
The findings support the transition from traditional genre labels to "genre clusters" defined by internal linguistic signs. This approach is particularly valuable for transfer learning and model adaptation. Future work aims to use these genre-correlating parameters as a "pre-processor" to automatically tune the hyperparameters of morphological and syntactic parsers, ensuring the most optimized model is selected for any given text type.
Summary updated