| --- |
| task_categories: |
| - feature-extraction |
| - sentence-similarity |
| language: |
| - en |
| tags: |
| - news |
| - markers |
| - co-occurrence |
| - causality |
| - complexity |
| size_categories: |
| - 10M<n<100M |
| --- |
| # Dataset Card for CausalityLink Marker Occurrences (Anonymized) |
|
|
| <!-- Provide a quick summary of the dataset. --> |
|
|
| A large tabular dataset of **marker occurrences in news articles**, derived from the |
| CausalityLink database. Each row records that a given *marker* (a concept extracted |
| from a financial/economic news article) appeared in a specific article published by a |
| given (anonymized) publisher belonging to a given editorial theme. The dataset is the |
| raw input used by the [`complexity-velocity`](https://github.com/keyvanatt/complexity-velocity/) project to study the |
| **complexity** and **velocity** of markers via their co-occurrence structure. |
|
|
| ## Dataset Details |
|
|
| ### Dataset Description |
|
|
| <!-- Provide a longer summary of what this dataset is. --> |
|
|
| A **marker** is a concept (e.g. `passenger_car`, `acquisition`, `new_year`) mined from |
| news articles by CausalityLink. This dataset is the "long" (one-row-per-occurrence) |
| join of three sources: the per-article marker extractions, the publisher labels, and a |
| publisher → editorial-theme mapping. From these occurrences one can build the |
| article × marker presence matrix and, from it, marker co-occurrence, lift, complexity |
| and velocity scores. |
|
|
| It is the empirical basis for the paper *"A Causal Model to Explain Complexity of Topics |
| and Its Empirical Link with Corpus Velocity"*, which models marker generation as a |
| Structural Causal Model, defines a lift-based **complexity** metric $\mathscr{C}(M_i)$, |
| and establishes the stylized fact that publication **velocity** is a decreasing function |
| of complexity — consistently across thematic clusters and across individual sources. |
| |
| - **Curated by:** Baptiste Arnaudo, Keyvan Attarian, Salah Chikhi, Charles-Albert Lehalle |
| - **Language(s) (NLP):** English (marker identifiers). |
| - **License:** Other — derived from the proprietary CausalityLink database; redistribution terms to be confirmed with the data provider. |
| |
| ### Dataset Sources |
| |
| <!-- Provide the basic links for the dataset. --> |
| |
| - **Repository:** [github.com/keyvanatt/complexity-velocity](https://github.com/keyvanatt/complexity-velocity/) |
| - **Paper:** [Comming Soon] |
| - **Demo:** [Comming Soon] |
| |
| ### Direct Use |
| |
| <!-- This section describes suitable use cases for the dataset. --> |
| |
| - Building article × marker co-occurrence and lift matrices. |
| - Computing per-marker **complexity** (average within-cluster lift) and **velocity** |
| (inverse marginal probability, `1/P(marker)`). |
| - Clustering markers (e.g. UMAP + DBSCAN/HDBSCAN) and analyzing cluster structure. |
| - Per-publisher or per-theme comparative analysis of marker complexity/velocity. |
| |
| ### Out-of-Scope Use |
| |
| <!-- This section addresses misuse, malicious use, and uses that the dataset will not work well for. --> |
| |
| - The dataset contains **no article text**, only marker identifiers, so it is not |
| suitable for language modeling or text classification on raw content. |
| - It is not intended to re-identify individual publishers; the publisher labels in the |
| canonical split are anonymized on purpose (see caveats below). |
| |
| ## Dataset Structure |
| |
| <!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. --> |
| |
| CSV files with one row per (article, marker) occurrence and the following columns: |
| |
| | Column | Type | Description | |
| |-------------------|--------|-----------------------------------------------------------------------------| |
| | `id` | string | Article identifier (`<publisher_id>_<timestamp><source>_<hash>`). | |
| | `marker` | string | Concept extracted from the article (English snake_case identifier). | |
| | `publisher_label` | string | Anonymized publisher label (`publisher_1` … `publisher_n`). | |
| | `journal_theme` | string | Editorial theme of the publisher (French; e.g. `presse`, `information`, `finance`). Empty when unknown. | |
|
|
| **`dataset.csv` (January 2025, anonymized):** |
|
|
| - ~57.7M occurrence rows |
| - ~1.19M unique articles |
| - ~28.7k unique markers |
| - 151 anonymized publishers |
| - Most frequent themes: `presse` (~17.3M), `information` (~14.6M), `distribution` (~6.2M), `generaliste` (~5.5M), `relationspubliques` (~3.9M), `data` (~3.5M), `finance` (~2.7M). |
|
|
| There is a single occurrence-level split; deriving article × marker matrices is left to |
| the user. Markers tagged as countries are excluded during extraction (via the Tree |
| table `country` filter). |
|
|
| ## Dataset Creation |
|
|
| ### Curation Rationale |
|
|
| <!-- Motivation for the creation of this dataset. --> |
|
|
| To study how the *complexity* and *velocity* of concepts (markers) in economic/financial |
| news relate to one another, and how these properties vary across publishers and |
| editorial themes. The minimal (id, marker, publisher, theme) schema is exactly what the |
| complexity/velocity pipeline needs, kept small and shareable relative to the full |
| CausalityLink dump. |
|
|
| ### Source Data |
|
|
| <!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). --> |
|
|
| #### Data Collection and Processing |
|
|
| <!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. --> |
|
|
| Extraction pipeline run for January 2025 from CL data: |
|
|
| 1. Drop markers that correspond to countries (Tree `country` is not null). |
| 2. Derive the `publisher_id` from the article `id` prefix and join to the publisher |
| label table (`CausalityLinkPublishers.csv`). |
| 3. Join publisher labels to the theme mapping to get |
| `journal_theme`. |
| 4. **Anonymize** publisher labels to `publisher_1..publisher_n` |
| 5. Write out `id, marker, publisher_label, journal_theme`. |
|
|
| #### Who are the source data producers? |
|
|
| <!-- This section describes the people or systems who originally created the data. It should also include self-reported demographic or identity information for the source data creators if this information is available. --> |
|
|
| Markers and causal relationships are produced by **CausalityLink** from public news |
| sources (press articles, central-bank reports, analyst notes). The article authors are |
| the original publishers whose identities are anonymized in the canonical split. |
|
|
| ## Bias, Risks, and Limitations |
|
|
| <!-- This section is meant to convey both technical and sociotechnical limitations. --> |
|
|
| - Coverage is skewed toward a few dominant themes (`presse`, `information`), so |
| marker statistics reflect that editorial mix rather than a neutral sample of the world. |
| - A single monthly snapshot (January 2025) is provided; temporal generalization is |
| limited. |
| - Marker extraction quality depends entirely on CausalityLink's upstream NLP pipeline. |
| - Derived from a proprietary source: redistribution and downstream licensing must be |
| confirmed with CausalityLink. |
|
|
| ### Recommendations |
|
|
| <!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. --> |
|
|
| Users should be made aware of the risks, biases and limitations of the dataset. In |
| particular: treat theme frequencies as non-representative of a general population, and |
| confirm licensing with the data provider before redistribution. |
|
|
| ## Citation |
|
|
| <!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. --> |
|
|
| [Comming Soon] |
|
|
| ## Glossary |
|
|
| <!-- If relevant, include terms and calculations in this section that can help readers understand the dataset or dataset card. --> |
|
|
| - **Marker** — a concept extracted from a news article. |
| - **Lift** — the co-occurrence strength of two markers: |
| $$\ell_{i,j} = \frac{\mathbb{P}\left(M_{i} = 1 \wedge M_{j} = 1 \right)}{\mathbb{P}(M_{i}=1)\;\mathbb{P}(M_{j}=1)} = \frac{\mathbb{P}(M_{i}=1 \mid M_{j}=1)}{\mathbb{P}(M_{i}=1)} = \exp \textbf{PMI}(M_i,M_j)$$ |
| - **Complexity** — the average pairwise lift of a marker against the other markers in its semantic cluster; a marker is complex when it presupposes many other concepts to be discussed. |
| $$\mathscr{C}(M_i) = \frac{1}{|\mathcal{S}(M_i)| - 1}\sum_{\substack{j \in \mathcal{S}(M_i) \\ j \neq i}} \ell_{i,j}$$ |
| - **Velocity** — `1 / P(marker)`, the inverse frequency of a marker's appearance (publication rate proxy). |
| |
| ## More Information |
| |
| See the [complexity-velocity](https://github.com/keyvanatt/complexity-velocity/) project README for the full analysis pipeline |
| (co-citation, lift, UMAP/DBSCAN clustering, per-publisher power-law fits, PC causal |
| discovery, and the LLM complexity judge). |
| |
| ## Dataset Card Authors |
| |
| Keyvan Attarian (keyvanatt). |
| |
| ## Dataset Card Contact |
| |
| keyvan.attarian@polytechnique.edu |
| |
