TurEngMix: A Text Corpus and Benchmark for Turkish-English Code-Mixed Language Identification and Named Entity Recognition

Paper: link · Code: link · Dataset: (this page)

Abstract

Natural language processing systems underperform on code-mixed text, particularly for low-resource language pairs. Turkish-English poses a further challenge: it lets English stems combine with Turkish suffixes to form single mixed-language tokens. We introduce TurEngMix, a corpus of 5.5K noisy, naturally occurring social media posts (486,974 tokens) rich in Turkish-English code-mixing. From this corpus, we construct a new Turkish-English benchmark for code-mixed language identification (LID) and named entity recognition (NER), comprising 15K expert-annotated tokens. Evaluating both decoder LLM and fine-tuned encoder baselines, we find that monolingual Turkish and English tokens are labeled reliably, but all models have high error rates on mixed-language tokens for both LID and NER. For morphologically integrated tokens, NER error rates were 5.2× and 6.3× higher for GPT-4o and Qwen, respectively. This highlights how morphological integration remains a challenge. We release the corpus, annotations, and code to support future computational and sociolinguistic research on Turkish-English code-mixing.

Dataset Description

This release contains the TurEngMix corpus and the TurEngMix annotated benchmark with language ID and named entity labels for each word token.

• ~5500 full text posts
• ~15K annotated tokens
• Source: Turkish-English code-mixed social media text

Dataset Structure

This release contains two related resources derived from the same collection pipeline:

1. TurEngMix Corpus

The full, unannotated corpus of naturally occurring Turkish-English code-mixed posts.

File	TurEngMix_Corpus.csv
Posts	5,549
Tokens	486,974
Post length	3–500 words (mean 87.75, median 57)
Format	One row per post: topic, entry, word_count

This is the source pool the benchmark below was sampled from. It is not token-annotated and is intended for pretraining, further sampling, or corpus-level sociolinguistic study.

2. TurEngMix Annotated Benchmark

250 posts from the corpus above, tokenized and expert-annotated at the token level for language identification (LID) and named entity recognition (NER).

File	TurEngMix_Annotated_Benchmark.csv
Posts	250
Sentences	321
Tokens	15,012
Format	One row per word token: 'doc_id', 'sent_id' 'tok_id', 'token', 'lid', 'integrated', 'ner'

Columns:

Column	Description
doc_id	Post identifier (e.g. post_001)
sent_id	Sentence index within the post
tok_id	Token index within the sentence (0-indexed, sequential)
token	The token text
lid	Language ID label (see schema below)
integrated	MIXED if the token is an English-origin stem with Turkish morphological suffixes attached (e.g. influencer + -lar + -ımız); blank otherwise
ner	Named entity label in BIO format (see schema below)

LID label schema (adapted from Solorio et al. 2014):

Label	Meaning
TR	Turkish
EN	English
MIXED	Token contains both an English-origin stem and Turkish morphology, and is not a named entity
NE	Named entity (any language of origin; overrides TR/EN/MIXED)
AMBIGUOUS	Could plausibly belong to either language given context
OTHER	Any other language

NER label schema (BIO tagging, CALCS 2018 guidelines): O (not an entity) plus B-/I- tags for: PER (person), ORG (organization), LOC (location), GROUP (sports teams, bands), PROD (product), TITLE (creative works), EVENT, TIME, OTHER.

3. TurEngMix Benchmark Posts

The 250 posts from the benchmark annotated for type of code mixing present in each post.

File	TurEngMix_Benchmark_Labelled_Posts.csv
Posts	250
Format	One row per post: 'post_id', 'post_text', 'mixed_language_token','embedded_english_phrase', 'isolated_english_token'

Columns

Column	Description
borrowed_suffix	True if post contains at least one morphologically integrated token
embedded_english_phrase	True if post contains ≥2 consecutive English tokens
isolated_english_token	True if post contains an English token with no adjacent English tokens

Loading the data

import pandas as pd

corpus = pd.read_csv("TurEngMix_Corpus.csv")
benchmark = pd.read_csv("TurEngMix_Annotated_Benchmark.csv")
benchmark_posts = pd.read_csv("TTurEngMix_Benchmark_Labelled_Posts.csv")

# Reconstruct a single post's tokens in order
post = benchmark[benchmark["doc_id"] == "post_001"].sort_values(["sent_id", "tok_id"])

---

Intended Use

This dataset is released strictly for:

• Non-commercial research
• Academic use
• Model evaluation and benchmarking purposes

Prohibited Uses

• Commercial use of any kind
• Redistribution or re-hosting of the dataset
• Training models for commercial deployment
• Attempting to identify individuals in the dataset

Ethical Considerations

• Dataset may reflect biases present in social media
• No PII was deliberately collected or annotated
• Intended for research use only

Citation

BibTeX:

  @misc{dogan2026turengmixtextcorpusbenchmark,
      title={TurEngMix: A Text Corpus and Benchmark for Turkish-English Code-Mixed Language Identification and Named Entity Recognition}, 
      author={Ilayda Dogan and Phuong-Anh Nguyen-Le and Julia Mendelsohn},
      year={2026},
      eprint={2609.06963},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.06963}, 
}

Contact

For questions or access issues:

• idogan@umd.edu
• nlpa@umd.edu