mawaskow/irish_forestry_incentives

Dataset Card for "Irish Forestry Incentives"

Dataset Summary

This dataset consists of sentences extracted from Irish national forestry policies, labeled for presence and type of incentive in a sentence. Policy incentives incentivise/ disincentivise actions and enable policies to enact change. Sentences have been labeled as either "Non-incentive" or as one of six incentive types.

The original and intended use of this dataset was to train binary and multiclass sentence classifiers, based on different transformations or groupings of the dataset. The dataset was created by scraping English-language policies from the Irish government website, extracting the text from their PDFs, cleaning and splitting that into sentences, then our labeling cycle.

Dataset Structure

Data Instances

Typical instances consist of a dictionary structure containing: "text", the sentence, and "label", its corresponding label. Below are examples of a non-incentive entry and an incentive entry.

{
  'text': "A policy sentence wholly unrelated to incentives."
  'label': "Non-Incentive"
}

{
  'text': "A policy sentence about fining people who participate in deforestation."
  'label': "Fine"
}

Data Fields

• text: string of sentence extracted from policy pdf and cleaned
• label: string from set ["Non-Incentive", "Credit", "Direct_payment", "Fine", "Supplies", "Tax_deduction", "Technical_assistance"]

Label Definitions

Our labels and definitions come from this paper.

Binary

• Non-Incentive: The sentence does not contain a policy incentive, or simply mentions an incentive keyword like grant or loan
• Incentive: The sentence contains a policy incentive, especially the creation or implementation of an incentive

Multiclass

• Credit: Loans, insurance
• Direct payment: Cash, grants
• Fine: Penalty payment
• Supplies: Material support, equipment
• Tax deduction: Reduced tax liability
• Technical assistance: Training, experts

Data Splits

There are no recommended splits of the dataset-- in our work, we first transformed and separated our dataset into binary [incentive/ non-incentive] and multiclass [only incentives] datasets, then trained and evaluated models across 10 splits of each dataset [train_test_split(r_state=range(10), stratify=labels)]. Our train/dev/test sets were split 60/20/20.

	train
Input Sentences	1419
Average Sentence Length	43.9

Dataset Creation

Curation Rationale

We originally wanted to apply the pipeline of this paper to Irish forestry policies, but their dataset and models were unavailable. We decided to replicate their pipeline in our new context, as well as make our resulting dataset available for future use. The details and code for model training are available in the policy-classifier repository.

Source Data

Initial Data Collection and Normalization

The data collection details including the scraper, search keywords, anti-keywords, and the text extraction, cleaning, and normalization can all be found in the policy-classifier repository.

Who are the source language producers?

The source language producers are the various authors of the policies contained on the website of the Irish government.

Annotations

Annotation process

The labeling of the sentences was peformed locally using the Doccano platform. Our labeling cycle consisted of first using a keyword search to gather around 1000 sentences to label by hand, then using five sentences from each class as a query, performing a sentence similarity search to find sentences common across multiple queries. These sentences then had their labels validated to complete the human-in-the-loop process. At this point, we grouped sentences by Levenshtein distance to get rid of duplicate entries, resulting in our final dataset.

Who are the annotators?

Besides the sentence similarity search in the human-in-the-loop automatic labeling step, the labeler was the author. For validation of our labelling, we brought in a policy researcher to label a stratified subsample of our final dataset. They annotated 10.7% of the sentences, including 24.5% of the incentives. To evaluate our agreement, we calculated the Cohen’s kappa score of the resulting binary and multiclass datasets. The agreement of the binary labelling, computed across the entire dataset, was 0.631 (“substantial agreement”); for the multiclass labelling, we evaluated agreement across the subset of examples where both annotators labelled the sentence as some kind of incentive, resulting in a score of 0.859 (“near perfect agreement”).

Personal and Sensitive Information

The dataset contains no identity categories or personal information.

Considerations for Using the Data

Social Impact of Dataset

We believe that the use of this dataset to train incentive classification models can help automate the time-consuming but important work of policy analysts looking for policy solutions.

Discussion of Biases

The dataset has a notable but expected skew towards non-incentives, with incentives making up only 18.5% of the dataset. The classes of incentives were also themselves unbalanced, with Supplies at 30.1% of all incentive sentences, Technical assistance at 27.9%, Direct payment at 23.0%, then Fine at 8.6%, Credit at 7.1%, and Tax deduction at 3.3%.

Other Known Limitations

Due to the non-political expertise of the single labeler, there may be incorrect labels. Please refer to our discussion of inter-annotator agreement with a policy researcher's annotations on a subset of the data in our "Who are the annotators?" section.

Additionally, the parsing of text from PDFs and the separating of texts into distinct sentences do not always result in single, clean sentences and this is true of some sentences in the dataset.

Additional Information

Dataset Curators

M.A. Waskow, Insight Centre for Data Analytics, University of Galway

Licensing Information

Copyright 2025 M.A. Waskow

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Citation Information

[Bibtex citation not yet available]