As Cool as a Cucumber: Towards a Corpus of Contemporary Similes in Serbian

Nikola Milosevic, Goran NenadicarXiv 2016 · arXiv

Introduction

Similes—figurative comparisons such as "white as snow" or "works like a horse"—serve as vital cultural and linguistic markers. However, these expressions are not static; they evolve alongside the culture that uses them. In the context of the Serbian language, traditional collections have remained largely stagnant since the mid-19th century, most notably with the work of Vuk Karadžić published in 1849. While some similes have persisted, many have vanished or evolved (e.g., the shift from kurjak to the modern vuk for "wolf").

This research addresses the need for a contemporary Serbian simile corpus. The primary technical challenge lies in automated extraction: while identifying the comparative keyword kao ("as" or "like") is trivial, distinguishing figurative similes ("works like a horse") from literal descriptions ("works as a lawyer") requires bridging a significant semantic gap.

Methodology

The researchers developed a semi-automated four-step pipeline to modernize the Serbian simile collection:

  1. Document Retrieval: Custom crawlers built with the Scrapy framework targeted diverse domains, including literary sites (laguna.rs, rastko.rs), general forums (burek.com), and blog portals (tarzanija.com).
  2. Comparison Extraction: Using the HunPos Part-of-Speech (POS) tagger, the system identified patterns matching (Verb/Adjective) + (kao/ko/k’o) + (Noun Phrase). To handle the phrase-level limitations of the tagger, noun phrases were modeled as sequences of adjectives ending in a noun.
  3. Simile Classification: To filter literal comparisons, the team trained machine learning classifiers (Naive Bayes, Random Forests, and SVM) using features such as stemmed words and local context.
  4. Data Curation: A web portal enabled human experts to validate extracted candidates and allowed for crowdsourced additions.
The four-step methodology for building the Serbian Simile Corpus.

Datasets and the "Evolution Gap"

The study utilized several key datasets:

  • Traditional Baseline: 333 similes from Vuk Karadžić’s 1849 collection.
  • Web-Mined Candidates: Approximately 5,000 candidate phrases extracted from over 40,000 crawled documents.
  • Training Set: A balanced set of 600 labeled examples (300 true similes, 300 literal comparisons).

A striking finding was the "Evolution Gap": only 9 similes overlapped between the 19th-century collection and the modern web-mined dataset. This highlights a massive shift in figurative language usage over the last 150 years and underscores the necessity of this project.

Key Results and Performance

The classification stage was critical for maintaining corpus quality. Among the tested algorithms, the Support Vector Machine (SVM) with a polynomial kernel proved most effective.

Performance metrics for the three machine learning classifiers tested.

As shown in the table above, the SVM achieved a Precision of 0.804 and an F-Measure of 0.797, significantly outperforming the other models in identifying true figurative similes. The final resulting "Serbian Simile Corpus" contains 852 approved similes, more than doubling the existing resource and providing a contemporary snapshot of the language. This corpus has been assigned an International Standard Language Resource Number (ISLRN) and is publicly accessible.

Conclusion and Future Directions

This work successfully demonstrates a hybrid approach to corpus building, combining automated extraction with human curation. However, limitations remain. The current "termination rule" for noun phrases can miss complex similes involving prepositional phrases (e.g., "bored like a dragon at a fire station").

Future work will focus on integrating deeper semantic resources like WordNet to better handle lexical ambiguity. Additionally, the researchers aim to maintain the corpus as a "living" document through continued crowdsourcing, ensuring it evolves alongside the Serbian language.

Summary updated

BibTeX
@misc{milosevic2016as,
  author = {Nikola Milosevic and Goran Nenadic},
  title = {{As Cool as a Cucumber: Towards a Corpus of Contemporary Similes in Serbian}},
  year = {2016},
  eprint = {1605.06319},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1605.06319},
}
Figure