ralipanah/email-politeness-corpus

Email Politeness Corpus

This dataset accompanies the paper:

A Synthetic Request–Reply Email Corpus Annotated with Document-Level Politeness and Sentence-Level Face Acts

Roshad Alipanah, Valentin Barriere, and Jorge Baier Findings of the Association for Computational Linguistics: EMNLP 2026

The corpus consists of synthetic request–reply emails jointly annotated at two levels:

• —Sentence level: multi-label Face Act annotations grounded in Brown and Levinson's politeness theory.
• —Document level: human politeness scores along three dimensions:
• —Directness vs. Indirectness
• —Positive Face Saving
• —Negative Face Saving

The dataset supports experiments in pragmatic language understanding, multi-label Face Act classification, document-level politeness prediction, and analysis of the relationship between sentence-level pragmatic strategies and overall perceived politeness.

Dataset Structure

The released data are organized as follows:

data/
├── corpus/
│   ├── email_text_gold_three_dimensions_politeness_score_with_seed_correct.csv
│   └── sentences_with_golden_face_act.csv
│
├── annotation/
│   ├── 480email_politeness_scores_of_the_annotators.csv
│   └── face_acts_annotators.csv
│
├── validation/
│   └── gpt4o_intended_politeness_vs_human_gold.csv
│
└── splits/
    ├── train_seed42.csv
    ├── val_seed42.csv
    └── test_seed42.csv

Corpus Files

Document-Level Corpus

data/corpus/email_text_gold_three_dimensions_politeness_score_with_seed_correct.csv

Contains the email-level corpus and the final gold scores used for document-level politeness prediction.

The three document-level targets are:

• —Directness_vs_Indirectness__GOLD → Directness vs. Indirectness
• —Structural_Politeness_and_Politeness_Markers__GOLD → Positive Face Saving
• —Tone_and_Overall_Consideration__GOLD → Negative Face Saving

Sentence-Level Corpus

data/corpus/sentences_with_golden_face_act.csv

Contains the sentence-level representation of the corpus together with the final gold Face Act annotations.

The Face Act annotation scheme contains nine categories:

• —HNeg+
• —HNeg-
• —HPos+
• —HPos-
• —SNeg+
• —SNeg-
• —SPos+
• —SPos-
• —Neutral

Sentences may receive more than one Face Act label.

Human Annotation Files

Document-Level Politeness Annotations

data/annotation/480email_politeness_scores_of_the_annotators.csv

Contains the individual human annotations used to construct the final document-level politeness scores and calculate inter-annotator reliability.

The annotation files retain the original annotation-stage column names. In the final terminology used in the paper:

• —Structural_Politeness_and_Politeness_Markers_admin corresponds to Positive Face Saving.
• —Tone_and_Overall_Consideration_admin corresponds to Negative Face Saving.

The corresponding final gold columns in the document-level corpus are:

• —Directness_vs_Indirectness__GOLD → Directness vs. Indirectness
• —Structural_Politeness_and_Politeness_Markers__GOLD → Positive Face Saving
• —Tone_and_Overall_Consideration__GOLD → Negative Face Saving

Sentence-Level Face Act Annotations

data/annotation/face_acts_annotators.csv

Contains the individual human Face Act annotations used to construct the sentence-level gold labels and calculate inter-annotator reliability.

Validation Data

data/validation/gpt4o_intended_politeness_vs_human_gold.csv

Contains the data used to analyze the relationship between the intended politeness levels used during controlled generation and independent human document-level politeness judgments.

Official Data Splits

The official train, validation, and test splits are provided in:

data/splits/train_seed42.csv
data/splits/val_seed42.csv
data/splits/test_seed42.csv

The splits are constructed at the seed level to prevent related generated emails from appearing across training, validation, and test partitions.

The reported main Overall Politeness Regression (OPR), GoldFA oracle, and misaligned PredFA results use the single-task (ST) setting with random seed 42.

Tasks

Face Act Classification

Face Act Classification is formulated as a multi-label sentence classification task.

The released code evaluates BERT-based models under several contextual settings, including:

• —sentence-only classification;
• —classification with previous sentences from the current email;
• —classification using paired request–reply history;
• —sequence-labeling variants.

Overall Politeness Regression

Overall Politeness Regression predicts the three document-level politeness dimensions:

• —Directness vs. Indirectness
• —Positive Face Saving
• —Negative Face Saving

The experiments include:

• —Text-only
• —PredFA-only
• —Text + PredFA
• —Text + GoldFA oracle
• —Misaligned PredFA ablation

Predicted or gold Face Act information is aggregated into email-level Face Act counts and summary features before being used for document-level politeness prediction.

Corpus Construction

The synthetic corpus was constructed from Enron-inspired request–reply scenarios.

The generation process includes:

1. 1.Generation of seed request–reply email pairs based on topics extracted from the Enron Email Dataset.
2. 2.Controlled generation of progressively more polite variants using GPT-4o while preserving communicative intent.
3. 3.Human annotation of sentence-level Face Acts.
4. 4.Human scoring of document-level politeness.
5. 5.Validation of the relationship between intended politeness levels and independent human judgments.

Annotation Framework

Both annotation levels are grounded in Brown and Levinson's politeness theory.

Sentence-level annotations capture Face Acts involving positive and negative face for both hearer- and speaker-oriented strategies.

Document-level annotations capture broader perceptions of politeness through:

• —Directness vs. Indirectness
• —Positive Face Saving
• —Negative Face Saving

The complete annotation guidelines are available in the accompanying GitHub repository.

Code and Reproducibility

The complete code for:

• —corpus generation;
• —Face Act Classification;
• —Overall Politeness Regression;
• —inter-annotator reliability;
• —GoldFA oracle experiments;
• —misaligned PredFA ablations; and
• —saved evaluation artifacts

is available at:

https://github.com/alipanahroshad-oss/email-politeness-corpus

The GitHub repository also contains the annotation guidelines and the scripts used to reproduce the experimental analyses reported in the paper.

License

The dataset is released under the:

Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) License.

Citation

If you use this dataset, please cite:

@inproceedings{alipanah2026synthetic,
  author    = {Roshad Alipanah and
               Valentin Barriere and
               Jorge A. Baier},
  title     = {A Synthetic Request--Reply Email Corpus Annotated with Document-Level Politeness and Sentence-Level Face Acts},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  publisher = {Association for Computational Linguistics},
  year      = {2026}
}