Conversation Derailment Forecasting with Graph Convolutional Networks

Enas Altarawneh, Ameeta Agrawal, Michael Jenkin, Manos PapagelisACL 2023

TL;DR

This paper introduces a novel graph convolutional neural network (GCN) model, FGCN, to predict online conversation derailment. By incorporating user dynamics and public perception alongside textual content, FGCN significantly outperforms existing state-of-the-art methods on benchmark datasets, offering a more proactive approach to online moderation.

What Problem Does This Paper Address?

Online conversations, especially on platforms like social media and forums, are prone to "derailment" – a shift away from the original topic due to inappropriate or abusive comments. This can lead to confusion, frustration, and a negative user experience. While existing methods often rely on sequence models that treat conversations as linear text streams, they overlook crucial aspects like user interactions and the broader perception of comments. The challenge is to forecast derailment before it causes significant harm, enabling timely intervention. This paper aims to bridge this gap by developing a model that captures the complex, multi-party dynamics of online discussions.

Methods

The core of this work is the Forecasting Graph Convolutional Network (FGCN). The authors propose a graph-based approach to model conversations, moving beyond simple sequential processing. Here's a breakdown of their method:

  • Input Representation: For each turn in a conversation, the model takes three types of input:
  • Textual Input: The text of the utterance, encoded using BERT embeddings.
  • User ID Input: Each user is represented by a unique, randomly initialized vector, processed by a Bi-LSTM. This captures user-specific tendencies.
  • Public Perception Input: A score derived from up-votes minus down-votes on an utterance, also processed by a Bi-LSTM. This captures how the community perceives the comment.
  • Graph Construction: Three types of homogeneous graphs are constructed, one for each input type (text, user ID, public perception). Vertices represent conversation turns, and edges represent relationships between turns.
  • User-to-User Relationship Edges: Edges connect turns based on temporal dependency (parent-child) and user-to-user interactions.
  • Edge Weights: These are learned using a similarity-based attention module.
  • Feature Transformation: A two-step graph convolution process transforms the initial feature vectors. This step aggregates local neighborhood information and then applies a further transformation to capture user-dynamic dependent features.
  • Forecasting Derailment: The transformed features from all three graph types are concatenated for each turn. These concatenated turn representations are then combined to form a conversation representation, which is fed into a classifier (linear layer, fully connected network, sigmoid activation) to predict the derailment label.

The novelty lies in the integration of user dynamics (via user IDs) and public perception (via voting scores) into a graph structure, alongside textual information, to model conversation flow.

Datasets

The authors evaluate their model on two publicly available datasets:

  • Conversations Gone Awry (CGA): Extracted from Wikipedia Talk Pages, this dataset contains conversations with toxicity scores and manual annotations to identify personal attacks. It provides text and user IDs.
  • Reddit ChangeMyView (CMV): Sourced from the ChangeMyView subreddit, this dataset identifies derailed conversations based on moderator deletions. It includes text, user IDs, and crucially, a public vote score (up-votes minus down-votes) for each utterance.

The paper notes that the CGA dataset does not provide public perception data, so FGCN-TS and FGCN-TSU variants were only tested on CMV.

Key Insights and Ideas

  • Beyond Text: The most significant insight is that simply analyzing text is insufficient for accurately forecasting conversation derailment. Incorporating user-specific information (like user IDs) and community feedback (public perception scores) dramatically improves prediction accuracy.
  • Graph Power: Modeling conversations as graphs, rather than linear sequences, effectively captures the complex, multi-party interactions and dependencies between users and their utterances.
  • Proactive Moderation: The ability to forecast derailment early (higher forecast horizon) is crucial for proactive moderation. This allows platforms to intervene before significant harm occurs, rather than just reacting to existing abuse.
  • Flexibility of FGCN: The FGCN framework is shown to be flexible, allowing for the integration of various data types. This suggests potential for future enhancements by including other signals like explicit emotion or sentiment.

Key Results

The FGCN model, particularly its variants incorporating user ID and public perception data, consistently outperformed baseline models on both CGA and CMV datasets.

  • Accuracy Improvement: The paper reports improvements of 1.5% and 1.7% on the CGA and CMV benchmark datasets, respectively, for the FGCN model compared to state-of-the-art.
  • Best Performing Variant: The FGCN-TSU model, which utilizes all three input types (text, user ID, public perception), achieved the best results.
  • Forecast Horizon: The dynamic training variant, FGCN-T+, demonstrated the earliest overall forecasting of derailment, with a mean forecast horizon of 4.12 for CMV and 2.95 for CGA. This means, on average, derailments were detected nearly 3-4 turns before they occurred, offering a substantial window for intervention.
  • Static vs. Dynamic Training: While statically trained models generally performed well, dynamically trained models (indicated by a "+" suffix) showed superior performance, especially in achieving earlier detection.

Limitations and Future Work

  • Minimum Conversation Length: Graph models, including FGCN, require a minimum number of utterances (stated as four or more) to form meaningful connections. Shorter conversations that derail might be better modeled by simpler sequential approaches.
  • Asynchronous Conversations: The models are best suited for asynchronous conversations where a time lag between turns allows for moderation after forecasting.
  • Future Directions: The authors suggest exploring the integration of explicit emotion or sentiment values into derailment forecasting models. They also highlight the potential benefit of incorporating other user-specific signals. The ethical implications of deploying such systems, including fairness and user notification, are also noted as important considerations for future work.

Summary updated

BibTeX
@inproceedings{altarawneh2023conversation,
  author = {Enas Altarawneh and Ameeta Agrawal and Michael Jenkin and Manos Papagelis},
  title = {{Conversation Derailment Forecasting with Graph Convolutional Networks}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure