Wiki-Talks

The Wiki-Talks dataset is a collection of conversational threads extracted from the talk pages on Wikipedia. This dataset captures collaborative dialogue, discussion patterns, and consensus-building among Wikipedia contributors. It is useful for NLP research focused on dialogue, sentiment analysis, and community dynamics.

Details

Currently due to PyArrow incompatibility to the long recursive structures in the dataset there is an intrinsic incompatibility between the dataset formatting and PyArrow, which makes it impossible to load the dataset from HuggingFace. You can still generate the dataset on your own using the code available on GitHub

Description

The Wiki-Talks dataset contains discussion threads from Wikipedia’s talk pages. Each thread includes a conversation title, a main message, and nested replies, capturing the sequence and structure of discussions. This dataset provides a unique view into how contributors collaborate, negotiate meaning, and resolve conflicts when editing content on Wikipedia.

• Curated by: @lflage
• License: CC BY-SA 3.0

Dataset Sources [optional]

• Repository: Wiki-Talks

Uses

The Wiki-Talks dataset is intended for research and development in NLP fields that require structured conversational data. Some use cases include:

• Dialogue Modeling: Training models for conversational agents that need to capture turn-taking and structured discourse.
• Sentiment Analysis: Analyzing language use, tone, and sentiment in collaborative or contentious settings.
• Toxicity and Moderation: Developing models for detecting inappropriate content and toxicity in user interactions.
• Discourse and Pragmatics Research: Studying patterns of collaboration, argumentation, and consensus-building in online communities.

Dataset Structure

The Wiki-Talks dataset is organized in a nested format to capture the hierarchical structure of conversations. Each thread contains:

The threads column contains the nested thread structure.

• title (str): The title of the discussion thread.
• message (str): The main message initiating the conversation.
• replies (list of reply objects): Each reply object contains:
  • message (str): The reply message.
  • replies (list of nested reply objects): Nested replies within the conversation, capturing deeper levels of response and discussion.

thread = { "title": str,
            "message": str,
            "replies: list[reply]}
reply = { "message": str,
           replies: list[reply]}

Dataset Creation

Source Data

The data was collected from publicly available Wikipedia talk pages, more specifically the pages-meta-current.

Data Collection and Processing

Data was selected from Wikipedia's talk pages and processed to structure the threads and replies hierarchically. The dataset includes only messages from public Wikipedia dumps and was processed using WikiTextParser, lxml and python standard libraries. It uses the time signature of a comment on a thread to delimit a comment.

Recommendations

When using this dataset, the following should be considered:

• Bias: As Wikipedia is user-generated, discussions may reflect certain biases based on the contributors' demographics, interests, or viewpoints.
• Limitations: Conversations may sometimes be off-topic or contain incomplete responses due to the open nature of Wikipedia.
• Technical: The dataset can be large and memory-intensive due to nested structures; consider streaming methods for processing.

Citation [optional]

For citation please refer to the original repository.

Contribution

Refer to the repository for how to contribute to expand the dataset