ROPE: Romanian Political Evasion

Overview

The ROPE Dataset is a curated, human-annotated collection of question-answer interactions between journalists and Romanian Presidents.

The dataset was constructed by extracting official transcripts from the Romanian Presidency websites (presidency.ro and archive.presidency.ro). It spans an extended period between the years 2002 and 2026, covering 5 Romanian Presidents (Ion Iliescu, Traian Băsescu, Klaus Iohannis, Ilie Bolojan, Nicușor Dan). It encompasses various public interaction formats, including press conferences, interviews, and public declarations.

Dataset Structure

The dataset is divided into two splits:

• train: Contains single-annotated question-answer pairs (3,278 examples).
• validation: Contains multi-annotated question-answer pairs (296 examples). Every item in this split was independently evaluated by 3 human annotators.

Schema

Column	Type	Description
index	int32	Unique identifier for the QA pair.
title	string	The official title of the press event.
date	string	The date the event took place (YYYY-MM-DD).
president	string	The name of the Romanian President speaking.
url	string	Direct link to the source transcript on the official Presidency website.
context	string	The format of the interaction (e.g., Interview, Press Conference, Declaration).
question_order	int32	The chronological order of the question within the event.
interview_question	string	The exact, verbatim text of the journalist's question.
interview_answer	string	The exact, verbatim text of the politician's answer.
question_summary	string	A concise summary of the question generated by GPT-5.4 and verified by at least one human annotator.
answer_summary	string	A concise summary of the answer generated by GPT-5.4 and verified by at least one human annotator.
annotator1	string	The fine-grained evasion category assigned by the first human annotator. Set to null in train, as there is only one annotator and their label is reflected in evasion_label.
annotator2	string	The fine-grained evasion category assigned by the second human annotator. Set to null in train, as there is only one annotator and their label is reflected in evasion_label.
annotator3	string	The fine-grained evasion category assigned by the third human annotator. Set to null in train, as there is only one annotator and their label is reflected in evasion_label.
clarity_label	string	The high-level (Level 1) classification of the answer's directness.
evasion_label	string	The specific (Level 2) evasion strategy used (populated in train; set to null in validation because of the task's complexity, meaning any of the individual annotator's valid guesses are considered correct in validation).

Taxonomy of Political Evasion

Figure 3: The taxonomy of response clarity classification (from Thomas et al., 2024).

The dataset utilizes a hierarchical taxonomy.

Level 1: Clarity (High-Level)

This level categorizes the general directness of the response into 3 main classes:

1. Clear Reply: The politician provides a direct, transparent answer to the question asked.
2. Clear Non-Reply: The politician explicitly refuses to answer or directly states their inability to do so.
3. Ambivalent: The politician gives a response that neither directly answers the question nor explicitly refuses it.

Level 2: Evasion Strategies (Fine-Grained)

This level breaks down the Level 1 classes into 9 specific sub-categories:

• Mapped to Clear Reply:

  • Explicit: Directly answers the question without ambiguity.

• Mapped to Clear Non-Reply:

  • Declining to answer: Explicitly refuses to provide the requested information.
  • Claims ignorance: Explicitly states a lack of knowledge on the requested information.
  • Clarification: Genuinely requests further details to understand the requested information before answering.

• Mapped to Ambivalent:

  • Implicit: Answers indirectly, requiring logical deduction from the listener.
  • Dodging: Completely ignores and avoids the question's substance and topic.
  • Deflection: Acknowledges the topic but shifts the focus or argument elsewhere.
  • Partial: Answers only one component of the question while ignoring the rest.
  • General: Stays on topic but uses abstract, vague generalizations instead of specific details to the information requested.

Usage

You can load the dataset using the datasets library:

from datasets import load_dataset

# Load both train and validation splits
dataset = load_dataset("gabrielstefan04/rope")

print(dataset["validation"][0])

Credits & Citation

This dataset and its methodology are inspired by the work done on the English-language QEvasion dataset.

If you use this dataset or its taxonomy in your research, please consider citing the original work establishing this framework:

@inproceedings{thomas-etal-2024-never,
    title = "{''}{I} Never Said That{''}: A dataset, taxonomy and baselines on response clarity classification",
    author = "Thomas, Konstantinos  and
      Filandrianos, Giorgos  and
      Lymperaiou, Maria  and
      Zerva, Chrysoula  and
      Stamou, Giorgos",
    editor = "Al-Onaizan, Yaser  and
      Bansal, Mohit  and
      Chen, Yun-Nung",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-emnlp.300",
    doi = "10.18653/v1/2024.findings-emnlp.300",
    pages = "5204--5233",
    abstract = "Equivocation and ambiguity in public speech are well-studied discourse phenomena, especially in political science and analysis of political interviews. Inspired by the well-grounded theory on equivocation, we aim to resolve the closely related problem of response clarity in questions extracted from political interviews, leveraging the capabilities of Large Language Models (LLMs) and human expertise. To this end, we introduce a novel taxonomy that frames the task of detecting and classifying response clarity and a corresponding clarity classification dataset which consists of question-answer (QA) pairs drawn from political interviews and annotated accordingly. Our proposed two-level taxonomy addresses the clarity of a response in terms of the information provided for a given question (high-level) and also provides a fine-grained taxonomy of evasion techniques that relate to unclear, ambiguous responses (lower-level). We combine ChatGPT and human annotators to collect, validate and annotate discrete QA pairs from political interviews, to be used for our newly introduced response clarity task. We provide a detailed analysis and conduct several experiments with different model architectures, sizes and adaptation methods to gain insights and establish new baselines over the proposed dataset and task.",
}