TL;DR
This paper introduces a novel visualization for argument search results, transforming the traditional list-based view into an interactive "topic space." This allows users to explore the diverse aspects of controversial topics covered by arguments, significantly improving the efficiency of finding relevant information and building informed stances.
What Problem Does This Paper Address?
In today's information landscape, discerning truth and forming well-rounded opinions on controversial topics is increasingly difficult due to the prevalence of fake news and biased sources. While argument search engines like args.me aim to provide a balanced view of pro and con arguments, simply ranking them by relevance isn't enough. Users need to understand the breadth of topics and sub-aspects discussed within the search results to truly grasp a complex issue. The existing text-based result lists make it hard to discover these diverse aspects, hindering users from forming an informed stance.
Methods
The core contribution is a new visualization interface for args.me. This interface represents arguments in a barycentric coordinate system, effectively creating a "topic space."
- Topic Modeling: To define the "aspects" of a topic, the authors explored two methods:
- Wikipedia-based: Using term distributions from Wikipedia's list of controversial issues.
-
LDA Topic Modeling: Applying Latent Dirichlet Allocation to the arguments themselves. The Wikipedia-based approach was found to be more convincing and is the default.
-
Barycentric Coordinates: Arguments are visualized as points (glyphs) in this topic space. Each vertex of the polygon represents a key aspect. An argument's position is determined by its weighted combination of these aspects.
- Argument Glyphs: These colored circles represent individual arguments (green for pro, red for con). Their size indicates the number of arguments aggregated at that point, and their color reflects the majority stance of the aggregated arguments. A small arrow on the glyph points to the argument's main covered aspect.
- Interaction: Users can hover over glyphs to see detailed aspect coverage and highlight corresponding arguments in the text list. Conversely, hovering over a text argument highlights its glyph. Crucially, users can select arguments by clicking or brushing in the topic space, which then filters the textual result list, allowing for rapid exploration and refinement of search results.
Datasets
The paper utilizes the existing args.me index, which contains approximately 291,440 arguments. For determining topical aspects, they leverage:
- Wikipedia's list of controversial issues: Used for the Wikipedia-based topic model.
- Internal
args.meargument corpus: Used for LDA topic modeling.
No new datasets or benchmarks were introduced; the focus is on visualizing existing data.
Key Insights and Ideas
- Beyond Relevance: The paper strongly argues that for complex topics, aspect coverage is a critical ranking criterion, often more important than simple relevance, recency, or user ratings.
- Visualizing the "Unknown Unknowns": The topic space visualization helps users discover aspects of a topic they might not have even considered, which is crucial for forming a comprehensive understanding.
- Interactive Exploration: The combination of hovering and brushing/clicking provides a powerful and intuitive way to navigate and filter argument search results, making the process significantly more efficient.
- Bridging Text and Visualization: The tight integration between the textual result list and the visual topic space creates a synergistic experience, where each view enhances the other.
Key Results
The authors demonstrate the effectiveness of their visualization through two case studies:
- Topic Space Exploration (Query: "feminism"): With 659 arguments, the visualization revealed that aspects like "Abortion" and "United Nations" were covered, which might have been missed in a purely relevance-ranked list. Arguments related to these aspects were ranked #43 and #46, highlighting the visualization's ability to surface less obviously ranked but relevant arguments.
- Search Refinement (Query: "assassination of dictators"): A user looking for arguments about "Muammar Gaddafi" could simply click on the corresponding glyph in the topic space. This instantly filtered the results, bringing the relevant argument to the top (which would have been ranked #34 otherwise).
The paper qualitatively shows that the visualization speeds up argument search notably by enabling efficient exploration and refinement.
Limitations and Future Work
- Aspect Model Limitations: While the Wikipedia-based model is effective, the authors note that the visualization limits the display to the eight most frequent aspects in the search results, merging others into an "other" category. This could potentially obscure less frequent but still important aspects.
- Ambiguity of Barycentric Coordinates: The paper acknowledges that barycentric coordinates can sometimes place arguments with different aspect distributions at similar locations, necessitating hover interactions for disambiguation.
- Future Work: The authors suggest that future research on modeling the aspects of an argument could further enhance the usability of the visualization. This implies exploring more sophisticated aspect extraction or representation techniques.
Summary updated