Three datasets, one corpus
Each dataset has its own DOI and can be used on its own. Together they move from the text of each treaty, to the legal content of its clauses, to the relationships between treaties.
Why this corpus
Türkiye has concluded bilateral investment treaties with partners across Europe, Asia, Africa, the Middle East and Latin America over more than six decades. Existing collections cover these treaties unevenly, and in several cases the texts they carry are not the authentic ones. This corpus rebuilds the Turkish treaty record from official sources, checks every text and date against the Official Gazette (Resmî Gazete), and codes each agreement clause by clause.
The result is a complete and verifiable basis for studying how one capital-exporting and capital-importing state has designed its investment treaties over six decades, and how that design has changed across generations of treaty-making.
Dataset 1 · Treaty Corpus
The full texts of Türkiye's bilateral investment treaties, cleaned and authenticated, with treaty-level metadata and a classification of every full protection and security (FPS) clause.
The corpus contains 141 agreements signed between 1962 and 2024, the most recent entering into force on 31 December 2025. Each text was checked against its official source, and the Official Gazette (Resmî Gazete) is treated as the authority for entry-into-force and status data. Signature dates are taken from the treaties themselves.
During cleaning, three texts in the UNCTAD collection turned out not to be the authentic versions (Slovakia 2009, Mali 2018 and Argentina 1992). Their authentic English texts were recovered from the Official Gazette and are the ones included here.
Coverage. 137 of the 141 bilateral investment treaties listed by UNCTAD are included. The other four (Cuba, Tajikistan, Congo-Brazzaville and Kazakhstan, all signed 2024 to 2026) are excluded because no text has been published. Two treaties with Kosovo that UNCTAD does not list are included.
What is inside
| File | Content |
|---|---|
treaty_texts.zip | Cleaned full text of all 141 agreements |
treaty_metadata.csv | Partner state, signature and entry-into-force dates, status, treaty generation |
article_extracts.csv | The text of each treaty divided into its articles |
fps_classification.csv | FPS type of every treaty, with the verbatim clause quoted |
00_REPRODUCIBILITY_PROTOCOL.txt | Six-step protocol for rebuilding all three datasets |
| README, codebook, schema | Documentation of every variable and coding decision |
The Pehlivan FPS Typology
The typology classifies the full protection and security standard by how it is formulated and what it is anchored to, not only by whether it is present.
| Type | Formulation | Treaties |
|---|
Fair and equitable treatment counts as "combined" only when it appears in the same article, back to back with the FPS obligation.
Figures
The left panel counts the 141 agreements by FPS type. The right panel shows how the mix of types changes from one treaty generation to the next.
How to read it: early Turkish treaties use a standalone FPS promise or tie it to fair and equitable treatment (Types A and B). Recent treaties move towards Type C, which anchors protection to the minimum standard of treatment. Type C first appears in 2011.
How to cite
Dataset 2 · Clause-Level Annotation
Every agreement in the corpus coded provision by provision, so that treaty design can be compared across partners and across time.
The coding scheme follows the structure of the UNCTAD IIA Mapping Project. Of the 118 substantive variables, 101 correspond one to one to UNCTAD mapping parameters and 17 are custom variables added for this corpus. Each coding decision is backed by a clause extract, so the legal basis of every value can be checked against the treaty text.
All rows were verified by the author, and a coding log records the decisions taken on difficult provisions.
What is inside
| File | Content |
|---|---|
treaty_annotations.csv | One row per treaty, one column per coded variable |
clause_extracts.csv | The clause text behind each coding decision |
coding_log.csv | Record of coding decisions and their reasons |
explorer.html | Interactive explorer that opens in any browser, offline |
| README, variable list | Definitions and coding rules for every variable |
Figures
The share of treaties containing eight key provisions, compared across four generations: Gen 1 (1962–85), Gen 2 (1986–98), Gen 3 (1999–2009) and Gen 4 (2010–24). The provisions are investor-State dispute settlement, full protection and security, denial of benefits, MFN excluding dispute settlement, qualified fair and equitable treatment, ESG-related language, the essential security exception and the umbrella clause.
How to read it: rising lines mark the arrival of the new Turkish model. Provisions that qualify or limit investor protection become common only in the most recent generation.
Each row is a treaty, ordered from oldest (top) to newest (bottom). Each column is one of 29 provisions. A filled cell means the treaty contains the provision.
How to read it: the lower part of the chart is visibly denser. Newer treaties are longer and regulate more, adding exceptions, carve-outs and sustainability language to the classic protection core.
The share of the 141 treaties coded "Yes" for each yes/no provision in the dataset, coloured by the section of the treaty it belongs to.
How to read it: the top of the chart holds the near-universal protections found in almost every treaty. The bottom holds rare and recent additions, which is where the corpus records innovation.
Phi correlation between pairs of key clauses across all treaties. A high value means two provisions tend to appear in the same treaties.
How to read it: a cluster of strongly correlated provisions shows the "new model" bundle. These clauses were adopted as a package rather than one by one.
How to cite
Dataset 3 · Diffusion and Genealogy Network
How treaty language moves through Türkiye's treaty programme: which agreements read alike, which families they form, and which earlier treaty each one most closely follows.
Interactive network
Each node is a treaty and each link connects it to the earlier treaty it most closely follows. Drag, zoom and click a node for details.
Every pair of treaties is compared by TF-IDF cosine similarity of their full texts (English stop-words removed, sublinear term frequency, terms in more than 90% of treaties ignored). Treaties are grouped into textual families by average-linkage clustering, cut at a cosine distance of 0.65.
Each treaty is then linked to its most similar predecessor, in a fixed order of signature date and treaty ID, provided the similarity is at least 0.45. The resulting genealogy shows which treaty texts served as templates for later ones.
What is inside
| File | Content |
|---|---|
similarity_matrix.csv | 141 × 141 matrix of cosine similarities |
treaty_pairs.csv | All 9,870 treaty pairs with textual and legal similarity |
treaty_families.csv | Family membership and family labels |
genealogy_edges.csv | Parent and child treaty for each descent link |
network_visualization.html | The interactive network shown above |
Figures
The distribution of cosine similarity across all 9,870 treaty pairs. Dashed lines mark the median, the 0.45 genealogy threshold, and the 0.8 and 0.9 levels of very high similarity.
How to read it: most pairs share a common treaty vocabulary but are not copies of each other. The thin tail on the right holds near-identical texts, the clearest sign of a template being reused.
The full 141 × 141 similarity matrix, with treaties reordered by family and then by year. Brighter cells mean more similar texts.
How to read it: bright squares along the diagonal are the textual families the clustering recovers. Treaties inside a square were drafted from the same model.
The 38 families ranked by the number of treaties they contain, coloured by the generation profile of their members.
How to read it: a few large families account for most of the corpus. They correspond to the main Turkish model texts of successive periods, while many small families capture one-off negotiated texts.
Each point is a descent link: the horizontal axis gives the year the later treaty was signed, the vertical axis the year of the earlier treaty it most closely follows. Colour shows how similar the two texts are.
How to read it: points close to the diagonal show treaties copying a recent predecessor. Points far below it show an old text being reused many years later.
For every treaty pair, textual similarity (horizontal) is plotted against the number of clause-feature packages from Dataset 2 that the two treaties share (vertical).
How to read it: the two measures largely agree (Pearson r = 0.61). Treaties that read alike also tend to contain the same legal provisions, which links the textual network to the legal coding.
How to cite
Code & Reproducibility
Every derived table and figure can be rebuilt from the deposited files. The reproduction toolkit is public on GitHub and is also deposited with Dataset 1, so the data and the code that produced them are preserved together.
Quick start
git clone https://github.com/ouzpehlivan/turkish-bit-corpus.git cd turkish-bit-corpus pip install -r requirements.txt python code/download_data.py # fetch the three datasets from DataverseNO python code/verify.py # 51-check cross-dataset verification suite python code/recompute_ds3.py # rebuild the network from Datasets 1 and 2 python code/fps_rule_check.py # re-derive the FPS typology from the clause text python code/make_figures.py # regenerate the figures
What the checks cover
Row and column counts, the identity of the FPS block across all three datasets, the provenance of every clause extract, synchronisation of the coding log, symmetry of the similarity matrix, family labels re-derived from their members, genealogy edges against the selection rule, the audit ledger, and the data embedded in the interactive files. Recomputing Dataset 3 from Datasets 1 and 2 reproduces the published network exactly.
Steps 2 and 4 of the protocol (article extraction and clause coding) involve legal judgment on the text. They are documented in full rather than automated.