PERCEPT: A Corpus for POS Tagging and Analysis of Persian–English Code-Mixing

Persian-English code-mixing—the practice of blending English words or phrases into Persian social media posts—is a common but under-researched phenomenon. Existing datasets for Persian often focus on sentiment or language identification but lack the fine-grained linguistic annotations necessary for building syntax-aware NLP models. To fill this gap, the authors introduce PERCEPT, a new corpus of 6,800 posts collected from X, Instagram, and Digikala.

The researchers developed an LLM-assisted annotation pipeline using Gemini 3.5 Flash to automatically assign Universal Dependencies (UD) part-of-speech (POS) tags to code-mixed words and label the topics of each post. They validated the reliability of this automated process by comparing it against manual annotations from native Persian speakers, finding high agreement across both POS tagging and topic classification tasks.

Using the PERCEPT corpus, the authors conducted a comprehensive linguistic analysis of code-mixing patterns. They found that nouns are the predominant category for code-mixed words across all three platforms. However, the secondary POS categories vary significantly: verbs are common on X, while adjectives and proper nouns are more frequent on Instagram and Digikala, respectively.

The study also examined the structural and contextual properties of code-mixing. The positional distribution of code-mixed words—where they appear within an utterance—is remarkably consistent across all platforms. In contrast, the "triggering effect" (the tendency for multiple code-mixed words to appear in the same post) is substantially more pronounced on Digikala, likely due to the nature of e-commerce reviews where users frequently list multiple product features or brand names. Finally, topics related to commerce and business consistently exhibit the highest degree of code-mixing.

PERCEPT provides a critical resource for researchers aiming to improve NLP tools for multilingual, code-mixed environments. By offering POS-annotated data, this corpus enables the development of more robust syntactic parsers and POS taggers for Persian-English text. Furthermore, the study’s findings regarding platform-specific mixing behaviors offer valuable insights into how digital environments and topical contexts influence language usage in multilingual communities.