Toolkits for Serbian Language
NLP tools and linguistic resources developed or adapted for Serbian, covering preprocessing, annotation, and analysis.
Development of NLP toolkits and linguistic resources for Serbian, supporting preprocessing, morphology-aware analysis, annotation, and corpus exploration.
Overview
This project groups together toolkits, scripts, and linguistic resources developed or adapted for Serbian language processing. The focus is on practical NLP support for a morphologically rich, under-resourced language, enabling downstream tasks such as classification, extraction, and corpus analysis.
Rather than a single system, this is a collection of interoperable tools used across multiple projects.
Preprocessing & normalization
Core preprocessing components include:
- Text normalization (Unicode, diacritics, casing)
- Tokenization adapted to Serbian specifics
- Handling of:
- Latin and Cyrillic scripts
- informal and social-media text
- punctuation, emojis, and non-standard word forms
These steps were designed to be reusable and configurable depending on the task.
Morphology & lexical resources
Work with Serbian lexical and morphological resources included:
- lemmatization and morphological features
- integration of lexicons (emotion, sentiment, domain-specific)
- use of WordNet-style resources for semantic alignment and enrichment
Special attention was paid to:
- inflectional variation,
- negations, derivations and variations,
- ambiguity introduced by rich morphology.
Annotation & extraction tools
Toolkits also covered annotation and information extraction, such as:
- scripts for manual and semi-automatic annotation inlcudin outputs validation
- named entity extraction (general and domain-specific)
- extraction of linguistic cues (keywords, patterns, markers)
These tools were used to bootstrap datasets and to support iterative annotation workflows.
Corpus analysis & utilities
Additional utilities include:
- frequency and co-occurrence analysis,
- n-gram and subword statistics,
- corpus-level summaries and diagnostics,
- export to formats suitable for ML pipelines.
The emphasis was on transparent, inspectable outputs rather than black-box processing.
Why this matters
For Serbian, high-quality NLP often depends on custom tooling rather than off-the-shelf solutions. This collection of toolkits enabled:
- faster dataset creation,
- consistent preprocessing across projects,
- reproducible experiments,
- and easier adaptation of models developed for high-resource languages.
SRPOL: A Lexicon-Based Framework for Sentiment Strength in Serbian
SRPOL is a lexicon-based sentiment analysis framework for Serbian, designed to estimate both sentiment polarity and sentiment strength in texts from the political and public discourse domain.
The framework is centered around a manually curated and corpus-expanded sentiment lexicon for Serbian, comprising approximately 15,000 lexical entries, enriched with polarity values and sentiment intensity information. In addition to core sentiment-bearing words, SRPOL incorporates contextual sentiment triggers that influence sentiment strength at the sentence level.
The resource was developed to address the lack of high-quality sentiment lexicons for Serbian and to support linguistically informed sentiment analysis in a morphologically rich, under-resourced language.
SRPOL has been evaluated on Serbian textual data from multiple domains and demonstrates competitive performance in sentiment strength detection, making it suitable both as a standalone lexicon-based solution and as a supporting component in hybrid or supervised NLP pipelines.
