Possible uses of syntactic complexity annotation in a parallel corpus: The case of French -ant forms in converbal function and their Czech counterparts

Protocol Olga Nadvornikova, Alexandr Rosen & Karolina Pavlikova Faculty of Arts, Charles University, Prague Moznosti vyuziti anotace syntakticke komplexity v paralelnim korpusu: príklad francouzskych tvaru na -ant v konverbalni funkci a jejich ceskych protejsku Possible uses of syntactic complexity annotation in a parallel corpus: The case of French -ant forms in converbal function and their Czech counterparts DOI: 10.5281/zenodo.17912727 Table of contents: 1. Abstract 2. About this document 3. Abbreviations 4. Research questions 5. The source data 6. The tools 7. Generating text-based statistics 7.1. Pre-processing the subcorpus 7.2. Computing statistics 7.3. Results 8. Generating sentence-based statistics 8.1. Pre-processing the subcorpus 8.1.1. Identifying Part 8.1.2. Identifying Ger 8.1.3. Removing outliers 8.2. Form-specific comparisons within and between languages 8.2.1. The goals 8.2.2. The input 8.2.3. The script ALL_SCM_plots.py 8.2.4. The output 8.3. Form-agnostic comparisons between languages 8.3.1. The input 8.3.2. The script ALL_SCM_csfr.py 8.3.3. The output: 8.4. Form-specific comparisons with reference data 8.4.1. The input 8.4.2. The script ALL_SCM_plots_refOnly.py 8.4.3. Results 9. Data availability statement 10. AI disclosure statement 1. Abstract This study explores new research opportunities offered by the InterCorp v16ud parallel corpus, annotated using the Universal Dependencies scheme and enriched with syntactic complexity (SC) measures. The analysis focuses on French sentences containing -ant forms (gerund and present participle) and their Czech translations, with participles restricted to adverbial (converbal) usage for comparability. The results show significant SC variation in literary texts, with Czech translations displaying lower values than French originals. Coefficient of variation and correlation analyses suggest that participles may function as stylistic markers, unlike gerunds. At the sentence level, participles are associated with higher SC than gerunds, though the differences are moderate. The contrastive analysis reveals substantial reductions in clausal SC measures in the Czech translations, probably due to the replacement of subordination by coordination. These shifts affect SC information hierarchy, and occasionally temporal relations. The study underscores the potential of InterCorp v16ud for syntactic research in contrastive linguistics and beyond, while emphasizing the multidimensional nature of SC. https://doi.org/10.14712/23366591.2025.1.7 Časopis pro moderni filologii 107, 2025, c. 1, s. 80–101 2. About this document This document describes the data and tools used and produced during the research presented in the article. It includes some additional materials which are not described in the article, including visualizations as plots and heatmaps. Because the article is in Czech, this document provides some rudimentary background needed for introducing the content. Links to datasets, plots and scripts refer to an external repository, but their names correspond to files in the zipped archive within this project. 3. Abbreviations SC: syntactic complexity SCM: SC metric Ger: gerondif — en chantant Part: participe present — chantant (restricted to CVB forms, i.e. those expressing adverbial subordination) CVB: converb CS: Czech original cs: Czech translation FR: French original fr: French translation UD: Universal Dependencies upos: UD part of speech feats: UD morphosyntactic features deprel: UD dependency relation (syntactic function) maxNPLength: no. of words in the longest NP in a sentence (SCM) maxNPDepth: no. of embeddings in the NP with most levels of embeddings in a sentence (SCM) sLength: no. of words in a sentence (SCM) mdd: mean dependency distance (SCM, average no. of word boundaries between the daughter and its head) maxTreeDepth: maximum no. of clausal embeddings in a sentence (SCM) subRatio: subordination ratio (SCM, subRatio = (no. of T-units + number of dependent clauses) / no. of T-units) T-unit: main clause (not coordinated) 4. Research questions RQ1: Do -ant forms influence the differences in SC between FR sentence and its cs equivalent? If so, in what way? RQ2: Is SC of FR sentences containing a Part higher than that of sentences containing a Ger? The latter tests per sentence are based on samples of 10K sentences. Per text, higher correlations of Ger/Part with average SCMs mean higher contribution of Ger/Part to the text’s SC. 5. The source data From InterCorp v.16ud (https://wiki.korpus.cz/doku.php/en:cnk:intercorp:verze16ud) we extracted the subcorpus of FR-cs fiction, 103 texts ➔ FRcs_Fiction_texts.xlsx For the text-based statistics, the average SCMs of the subcorpus were compared with those for the French and Czech translation of the Bible, extracted again from InterCorp v.16ud ➔ bible.xlsx 6. The tools ● The KonText corpus tool (https://www.korpus.cz/kontext/) for extracting the subcorpus and generating concordances ● Standard spreadsheet editor to generate tabular formats used to generate statistics ● Python scripts to generate statistics (see below) 7. Generating text-based statistics 7.1. Pre-processing the subcorpus The subcorpus was used to generate a French and a parallel Czech list of the 103 texts. For each text in each language, the text’s basic metadata, average SCMs and frequency data about the Ger and Part forms were added. ➔ CMFipm_bare.xlsx ➔ FRcs_Fiction_texts.xlsx ➔ csFR_Fiction_texts.xlsx 7.2. Computing statistics scatter.py – this script generates scatter plots for ipm of Ger/Part as x and SCMs as y AnalyseTextsSCM.py – what this script does: 1. Calculates variability: ○ Standard Deviation: Shows how spread out the values are. ○ Coefficient of Variation: Standard deviation divided by the mean (useful for comparing variability between measures with different scales). 2. Calculates correlations: ○ Pearson correlation matrix to identify which pairs of measures have strong, moderate, or no correlation. 3. Visualizes correlations: ○ Heatmap of the correlation matrix for a quick overview. 4. Plots distributions: ○ Histogram for each measure to observe its distribution and check for skewness, outliers, etc. 7.3. Results ➔ scatter_plots.pdf ➔ SCM_heatmap.png ➔ correlation_results.xlsx ➔ Plots_texts 8. Generating sentence-based statistics 8.1. Pre-processing the subcorpus The subcorpus was used to generate two 10K random samples including Part and Ger, each consisting of 10K French segments (mostly sentences) and their parallel Czech counterparts. For each segment in each language, the basic metadata about the text and the SCMs of the segment were added. The Part and Ger forms were identified using regular expressions (see below). (The SCMs have been computed before InterCorp 16ud was released. They are represented as metadata for each sentence and text, and can be displayed and processed like other metadata in the KonText tool.) 8.1.1. Identifying Part [ mark_lemma!="en" & e_deprel="advcl|aux:tense|cop|aux:pass|aux:caus|advcl:cleft" & upos="VERB|AUX" & feats="Tense=Pres" & feats="VerbForm= Part" & word=".*ant" & word!="[n|N]onobstant" ] 8.1.2. Identifying Ger [ mark_lemma="en" & ( (feats="Tense=Pres" & feats="VerbForm=Part") | (aux_type="tense|caus|pass" & aux_feats="Tense=Pres" & aux_feats="VerbForm=Part") & (feats="VerbForm=Inf " | (feats="Tense=Past" & feats="VerbForm=Part") ) ) ] ➔ 2025-02-06_ICv16ud_FRfic_GerSCMs.xlsx ➔ 2025-02-06_ICv16ud_FRfic_PartSCMs.xlsx 8.1.3. Removing outliers The samples were stripped of sentences showing extreme values that could distort the resulting analyses. Following Mačutek et al. (2019) we defined outliers on the basis of a boxplot as values lying below q1−5 IQR and above q3−5 IQR, where IQR is the interquartile range. Any sentence that fell among the outliers in at least one SCM was excluded. There were 395 such sentences in the Part sample and 267 in the Ger sample). In both samples, the most frequent source of such cases was a single text, La Belle du Seigneur by Albert Cohen (122 instances in Part and 39 in Ger), including sentences up to 16,210 words long with several -ant forms. Potential outliers among the SCMs for the Czech translations were ignored. 8.2. Form-specific comparisons within and between languages 8.2.1. The goals 1. Within-language comparisons: ○ For French, comparing values for Ger vs. Part for the same SCM; for Czech, comparing values for the segments parallel to the French Ger vs. Part 2. Between-language comparisons: ○ For each SCM (in both French and Czech), comparing the values across the parallel sentences, done separately for Ger and Part. 8.2.2. The input ● Two spreadsheets: for Ger and for Part. ● Each spreadsheet contains 6 columns for French SCMs and 6 for the corresponding Czech SCMs with the names exactly as listed below, approximately 10K each. ● maxNPDepth_fr, subRatio_fr, sLength_fr, maxNPLength_fr, mdd_fr, maxTreeDepth_fr, ● maxNPDepth_cs, subRatio_cs, sLength_cs, maxNPLength_cs, mdd_cs, maxTreeDepth_cs ● Cells that contain an underscore (_) are treated as missing. ➔ ger.xlsx ➔ part.xlsx 8.2.3. The script ALL_SCM_plots.py How to run it: python your_script_name.py gerunds.xlsx participles.xlsx --n_iterations 1000 --ci 95 --output_dir plots (could be *.csv instead of *.xlsx) The steps: 1. Reads two input CSV files: One for French Ger and one for French Part 2. For each SCM (six in total): ○ Within-language comparisons (Ger vs. Part): ■ For French: compares the French column for the SCM between the Ger and Part files. ■ For Czech: compares the Czech column for the SCM between the Ger and Part files. ○ Between-language comparisons (French vs. Czech): ■ For the Ger dataset: compares the French and Czech columns for the SCM. ■ For the Part dataset: does the same comparison. 3. Statistical Analyses: For each comparison (within‑ and between‑language): ● Computes extended descriptive statistics (mean, standard deviation, min, Q1, median, Q3, max). ● Computes the simple difference of means between the two groups. ● Uses bootstrap resampling (by default, 1,000 iterations) to obtain a (default 95%) confidence interval (CI) for the difference of means. ● Generates a “binary” violin plot (i.e. side‑by‑side violins) to visualize the distributions of the two groups. 8.2.4. The output 1. For each SCM, the script prints a table with the descriptive statistics for each group plus the difference of means and its bootstrap CI. 2. It saves violin plot images (one per comparison) to an output directory (by default a folder called “plots”). ➔ ALL_SCMv5_merged.xlsx ➔ plots 8.3. Form-agnostic comparisons between languages For each SCM, the script compares French versus Czech (using the corresponding _fr and _cs columns). No distinction is made now between Part and Ger.. 8.3.1. The input A single input file (XLSX or CSV) must contain the following 12 columns: maxNPDepth_fr, subRatio_fr, sLength_fr, maxNPLength_fr, mdd_fr, maxTreeDepth_fr, maxNPDepth_cs, subRatio_cs, sLength_cs, maxNPLength_cs, mdd_cs, maxTreeDepth_cs. Cells with an underscore (_) are treated as missing. 8.3.2. The script ALL_SCM_csfr.py How to run it: python your_script_name.py input_file.xlsx --n_iterations 1000 --ci 95 --output_dir plots The steps: For each measure the script computes: ● Extended descriptive statistics for each language (mean, standard deviation, min, Q1, median, Q3, max). ● The absolute difference of means along with its bootstrap confidence interval. ● Cohen’s d (as before). ● A significance test (t‑test if both groups pass a Shapiro–Wilk normality check, otherwise Mann–Whitney U). ● Both the absolute difference of means and Cohen’s d are included in the results. ● A split violin plot is generated for each measure comparing French and Czech. The y‑axis is set to log‑scale. 8.3.3. The output: For each SCM, the terminal displays: ● Descriptive statistics for French and Czech. ● The absolute difference of means with its bootstrap CI. ● The relative difference of means (in percent) with its bootstrap CI. ● Cohen’s d and the significance test result. ➔ PartFrCsSCM.out 8.4. Form-specific comparisons with reference data For each SCM, the script compares French Ger and Part versus reference values from a 10K sample from the subcorpus not constrained to Ger or Part. In this task, the _cs values do not represent Czech parallels but the reference data. 8.4.1. The input There are two files, one for Ger and one for Part, each including SCM values for segments including the specific form and from the reference sample. ➔ ref_ger.xlsx ➔ ref_part.xlsx 8.4.2. The script ALL_SCM_plots_refOnly.py – the script is similar to the one used in Part 3, but now it is run twice on each of the input file 8.4.3. Results ➔ ref_ger_out.xlsx ➔ ref_part_out.xlsx ➔ plots_refOnly 9. Data availability statement All data, scripts, and results are available under the Creative Commons Attribution 4.0 International License (CC BY 4.0). 10. AI disclosure statement The scripts were generated in January 2025 by ChatGPT, model version GPT-4o1. Parts of the script documentation generated by the model were also used in this document. However, all those generated results were reviewed and edited by the authors, who bear full responsibility for their consistency and accuracy.