Identifying Constructive Conflict in Online Discussions through Controversial yet Toxicity Resilient Posts

Readme

20-September-2025

Contact: oseckin[at]iu.edu

This repository contains the datasets required to reproduce the results presented in the paper "Identifying Constructive Conflict in Online Discussions through Controversial yet Toxicity Resilient Posts."

• submissions.parquet: PII redacted Reddit submissions included in the study.
• comments.parquet: PII redacted Reddit submissions included in the study.
• t_manual_annotation.parquet: Manually annotated Reddit posts for toxicity.
• c_manual_annotation.parquet: Manually annotated Reddit posts for controversiality.

Dataset Content for submissions.parquet

Reddit Metadata

• created_utc: Unix timestamp indicating when the Reddit submission was created.
• id: Unique identifier for the Reddit submission.
• subreddit: The name of the subreddit where the submission was posted.
• score: Reddit's upvote/downvote score for the submission.
• train_valid_test: Dataset split indicator showing whether the submission is used for training, validation, or testing.

Toxicity and Controversy Scores

• submission_openai: OpenAI's toxicity score for the submission text (0-1 scale).
• mean_openai: Mean toxicity score across all comments in the submission.
• toxic_ratio: Proportion of comments in the submission that are considered toxic (score ≥ 0.5).
• ta_score: Model-predicted toxicity score generated by the fine-tuned DistilBERT model.
• c_score: Model-predicted controversy score generated by the fine-tuned DistilBERT model.

Topic Analysis

• bertopic: Topic assignment from BERTopic clustering analysis of submission texts.

Linguistic Features

• is_question: Binary indicator (0/1) for whether the submission contains a question mark.
• question_ratio: Proportion of sentences in the submission that are questions.
• gratitude_ratio: Proportion of sentences in the submission that express gratitude.
• gratitude: Binary indicator (0/1) for whether the submission contains gratitude expressions.
• gratitude_count: Count of gratitude expressions found in the submission.
• pos_tag_counts: Dictionary containing counts of each part-of-speech tag in the submission.
• proper_noun_count: Count of proper nouns (NNP, NNPS) found in the submission.
• proper_noun_ratio: Proportion of words in the submission that are proper nouns.
• text_length: Total number of words in the submission.
• lexical_item_count: Count of unique lexical items (nouns, verbs, adjectives, adverbs) in the submission.
• mtld: Measure of Textual Lexical Diversity (MTLD) score indicating vocabulary richness.
• hedge: Binary indicator (0/1) for whether the submission contains hedging language.
• hedge_count: Count of hedging expressions found in the submission.
• hedge_ratio: Proportion of sentences in the submission that contain hedging language.
• polarity: Sentiment polarity score of the submission text (-1 to 1 scale).
• positive_polarity: Binary indicator (0/1) for whether the submission has positive sentiment.
• negative_polarity: Binary indicator (0/1) for whether the submission has negative sentiment.