Introduction
Language is a dynamic, evolving entity, yet linguistic corpora often struggle to keep pace with modern usage. In the context of the Serbian language, the foundational collection of similes dates back to the mid-19th century work of Vuk Karadžić. While these historical records are invaluable for understanding folk heritage, they often contain phrases that are now obsolete, while lacking modern expressions like "fast as a Ferrari" or "bankrupt as an endorser."
A significant technical hurdle in updating these records is the inherent ambiguity of the Serbian word kao (like/as). It serves both figurative similes (e.g., "works like a horse") and literal comparisons (e.g., "works as a lawyer"). Standard Natural Language Processing (NLP) tools frequently fail to distinguish between these uses based on syntax alone, as both follow nearly identical grammatical patterns. This research presents a semi-automated pipeline designed to bridge this gap, creating a searchable database of contemporary Serbian similes that reflects 21st-century linguistic reality.
Methodology: A Semi-Automated Pipeline
The researchers developed a four-stage process to identify, extract, and validate similes from the modern web, combining automated data processing with human expertise.
- Document Retrieval: Using the Scrapy framework, the team crawled four diverse Serbian web domains to capture a broad spectrum of language use. These included
burek.com(a popular forum),laguna.rs(a major book publisher),rastko.rs(a digital library of culture), andtarzanija.com(a satirical blog). In total, over 40,000 documents were collected. - Comparison Extraction: A Part-of-Speech (POS) tagger specifically designed for the Croatian and Serbian languages was used to identify potential similes. The extraction followed a specific linguistic model:
(Verb/Adjective) + (Connection word: kao/ko/k'o) + (Optional Adjectives) + (Noun). - Machine Learning Classification: To address the high volume of false positives—literal comparisons that match the grammatical pattern—the team trained classifiers using the Weka toolkit. They tested Naive Bayes, Random Forests, and Support Vector Machines (SVM) on a dataset of 600 examples, utilizing features such as stemmed phrases and the context words surrounding the connection term.
- Human Curation & Crowdsourcing: Recognizing that automated classification has limits, the team built a web interface (ezbirka.starisloveni.com). This platform allows a human curator to approve extracted similes and enables users to manually submit new entries, ensuring the corpus remains accurate and up-to-date.
Data and Cultural Evolution
The study utilized several key datasets to track the evolution of the language:
- Karadžić’s Historical Corpus: A baseline of 333 similes from 1849.
- Web-Crawled Dataset: 40,239 documents from the aforementioned web sources.
- The Final Corpus: A consolidated collection of 787 unique, verified similes.
The analysis revealed a "Semantic Wall": only about 10% of extracted phrases were genuine similes. This finding underscores that for figurative language, syntax is a poor proxy for meaning; current Serbian NLP tools lack the semantic depth to easily separate "like a dragon" from "as a professional."
Furthermore, the data highlighted a profound cultural shift. Modern similes show a significant "de-Ottomanization," with Turkish loanwords common in the 19th century being replaced by global brands and contemporary concepts (e.g., "stomach like an Airbus A330"). There was also a notable mythological shift; the traditional Serbian hala (a mythical spirit of storms) has largely disappeared from common usage, replaced by the Westernized "dragon" (zmaj). Interestingly, the modern "dragon" is no longer the protector-spirit of Serbian folklore but has evolved into the fire-breathing antagonist typical of Hollywood tropes. The low overlap (only 10%) between the 19th-century corpus and the modern web corpus further proves how rapidly figurative language evolves.
Key Results
The project successfully more than doubled the documented Serbian simile count, expanding the formal record from 333 historical entries to 787 contemporary ones. The semi-automated approach proved highly efficient, allowing researchers to filter over 5,000 potential candidates down to 442 valid new similes with significantly less effort than purely manual collection.
In terms of technical performance, the Support Vector Machine (SVM) with a polynomial kernel emerged as the top-performing classifier. It achieved a Precision of 0.804 and an F-Measure of 0.797 in distinguishing figurative similes from literal comparisons.
Limitations and Future Outlook
The primary limitation of the current system is the structural simplicity of the extraction model. The regex-like pattern fails to capture complex similes involving multiple nouns or extended descriptions, such as "bored like a dragon at a fire station."
Future work will focus on improving the extraction of these multi-word figures of speech and expanding the platform to include other short linguistic forms like proverbs and riddles. Additionally, the researchers suggest that this methodology could be applied to other Slavic languages that share similar grammatical structures, potentially leading to a broader understanding of linguistic evolution across the region.
Summary updated