CrossLit: Connecting Visual and Textual Sensemaking for Literature Review

Paper Title

CrossLit: Connecting Visual and Textual Sensemaking for Literature Review

Publication Info

• Topic area: Integrated tools for literature review sensemaking
• Keywords: Literature review, sensemaking, visual-text integration, academic workflows, schema building, hybrid queries, metadata visualization, narrative construction, cross-modal synchronization, human-AI collaboration

Background and Problem

• Problem / challenge: Existing tools for literature review either focus on visual interfaces for metadata exploration or textual interfaces for narrative construction, but they lack integration, forcing researchers to switch between tools and disrupting workflow continuity.
• Significance: Literature reviews are essential yet cognitively demanding tasks in academic research. Integrating visual and textual modalities could streamline the process, reduce cognitive load, and improve the quality of reviews.
• Motivation and related work: Prior tools either provide macroscopic overviews (e.g., citation networks) or support thematic organization and writing. However, they fail to bridge the gap between exploration and synthesis, leaving researchers to manually reconcile insights across modalities. This paper addresses the need for a unified system that synchronizes visual and textual modalities.

Solution

• Proposed approach: CrossLit, a system that integrates visual and textual interfaces for literature review, enabling seamless transitions between exploration, organization, and narrative construction.
• Novelty:
  1. Bidirectional synchronization between visual and textual modalities, ensuring structural consistency across both.
  2. Hybrid query functionality combining metadata-driven and content-driven discovery.
  3. Support for schema building as a critical intermediate stage in literature review.
  4. Integration of LLMs for adaptive text generation and semantic coherence.
• Procedure and key techniques:
  • Visual editor: Allows spatial organization of papers using metadata axes, hierarchical grouping, and citation visualization.
  • Text editor: Supports hierarchical structuring of narratives with inline citations and block-based editing.
  • Synchronization: Maintains one-to-one correspondence between visual and textual representations, with LLMs adapting content across modalities.
  • Hybrid queries: Combines citation-based and semantic search to discover relevant papers, reranked using a cross-modal encoder.

Results

• Concrete findings:
  • CrossLit achieved a System Usability Scale (SUS) score of 73.2, indicating above-average usability.
  • Participants retrieved an average of 18.31 papers per session, with 62.2% sourced via hybrid queries.
  • Hybrid queries balanced structural coherence (citation density: 0.0358) and semantic relevance (similarity: 0.3818), outperforming single-modality baselines.
• Advantage over baselines:
  • Enabled more frequent and fine-grained iterations compared to traditional workflows.
  • Integrated discovery features reduced reliance on multiple tools, consolidating exploration, grouping, and writing into one system.
• Experiments / evaluation:
  • User study with 16 researchers (HCI domain, 4.5 years average experience) using CrossLit for 45 minutes on their own topics.
  • Data collected via think-aloud protocol, screen recordings, and post-task interviews.
  • Sequential pattern analysis revealed distinct modality usage: visual for early-stage exploration and textual for later-stage narrative construction.
• Limitations and future work:
  • Scalability: Current visualization supports ~100 papers; future work should address larger datasets.
  • Synchronization: Users expressed concerns about preserving crafted content during automated updates.
  • Advanced phases: Longitudinal studies are needed to evaluate support for later review stages, such as evaluation and collaboration.

Summary

CrossLit integrates visual and textual modalities to support the literature review process, enabling seamless transitions between exploration, schema building, and narrative construction. Its bidirectional synchronization and hybrid query functionality allow researchers to iteratively refine their reviews while balancing metadata-driven and content-driven discovery. A user study demonstrated its effectiveness in supporting fine-grained iterations and schema development, though challenges remain in scalability and preserving user-crafted content. Future work should explore adaptive synchronization and cross-layer integration to enhance support for complex sensemaking tasks.