Reader-based Turkish sentence difficulty: stimulus features, trial-level comprehension responses, pairwise preference judgments, and analysis code

This dataset accompanies a manuscript, currently submitted for publication, on reader-based measurement of Turkish sentence reading difficulty. The study examines whether morphological, syntactic, and lexical complexity predict comprehension difficulty beyond surface sentence length. The dataset contains the raw field-study data and the analysis code needed to regenerate the quantitative results reported in the manuscript.

THIS VERSION

Table numbering aligned with the PLOS ONE version of the manuscript; full regression output added (Section 15); a pair-matching check added (Section 16); a licence file added; several stale table labels in the analysis script corrected. The analysis itself is unchanged: no reported value differs from the previous version.

FILES

saha_calismasi_veri_seti_800.xlsx - Sentence pool with expert-checked linguistic features (morpheme density, constituent-phrase count, non-finite verbs, foreign/proper-name count, rare-word ratio, word count). The sheet "selected_sentences" is used; the 787 sentences administered in the field study are a subset.

2_soru_yanitlari-10.csv - Trial-level comprehension responses from the web-based field study (1,480 participants; 22,490 responses), including question type, chosen answer, accuracy, and response time.

3_karsilastirma_tercihleri-4.csv - Trial-level pairwise preference judgments (2,960 judgments).

reproduce_analysis.py - A single self-contained Python script that regenerates the manuscript's tables and in-text statistics directly from the raw files, in sixteen labelled sections.

requirements.txt - Python dependencies.

reproduction_results.txt - Console output of a complete run of the analysis script, for reference.

README.md - Instructions and a description of the analysis pipeline.

LICENSE - Reuse terms: CC BY 4.0 for the data and documentation, MIT for the analysis code.

COVERAGE

The script regenerates Tables 1-4 and 6-13 of the manuscript and the in-text statistics. Table 5 is the one exception: it describes the word-count matching of the planned sentence-pair pool of 1,970 pairs, and only the 1,656 pairs actually administered appear in the archived response files. Section 16 of the script therefore runs the same matching check on the administered pairs, which agrees with Table 5 to within half a percentage point. This is explained in full in README.md.

HOW TO REPRODUCE

Create a fresh virtual environment, install the dependencies (pip install -r requirements.txt), place the three raw data files in one directory, and run: python reproduce_analysis.py --data-dir /path/to/data

The script prints every result section by section, labelled with the corresponding manuscript table. Runtime is under two minutes. The output is deterministic: repeated runs, from any directory, produce byte-identical results, so the archived reproduction_results.txt can be compared directly against your own run.

Note on dependencies: scipy is pinned below 1.16 because statsmodels 0.14 cannot be imported alongside newer scipy releases. Installing from requirements.txt avoids the problem.

PRIVACY

Participants are identified only by an arbitrary integer ID. No names, contact details, or other directly identifying information were collected.

ETHICS

The study was approved by the Scientific Research and Publication Ethics Committee of the Faculty of Science and Engineering, Afyon Kocatepe University (decision no. 2026/4, 25 February 2026).

FUNDING

Scientific and Technological Research Council of Türkiye (TÜBİTAK), 1001 programme, grant no. 124E056.