Abstract
This article proposes a hybrid model for the estimation of the complexity of legal documents in Russian. The model consists of two main modules: linguistic feature extractor and a transformer-based neural encoder.
The set of linguistic metrics includes both non-specific metrics traditionally used to predict complexity, as well as style-specific metrics developed in order to deal with the peculiarities of official texts. The model was trained on a dataset constructed from text sequences from Russian textbooks.
Training data were collected on either subjects related to the topic of legal documents such as Jurisprudence, Economics, Social Sciences, or subjects characterized by the use of general languages such as Literature, History, and Culturology. The final set of materials used contain 48 thousand selected text blocks having various subjects and level-of-complexity identifiers.
We have tested the baseline fine-tuned BERT model, models trained on linguistic features, and models trained on features in combination with BERT predictions. The scores show that a hybrid approach to complexity estimation can provide high-quality results in terms of different metrics.
The model has been tested on three sets of legal documents.