Facets, Taxonomies, and Syntheses: Navigating Structured Representations in LLM-Assisted Literature Review

TL;DR

This work introduces DimInd, an interactive, LLM-assisted system for literature review that guides researchers through four linked representations: a paper collection, a faceted literature review table, a facet taxonomy, and a facet synthesis. By enabling user-defined and system-suggested columns, hierarchical taxonomies, and controllable narrative generation with provenance, the approach aims to scale sensemaking across large paper collections. A within-subject study with 23 computer science researchers shows DimInd reduces cognitive load and improves organization relative to a ChatGPT-assisted baseline, while also highlighting the need for preserving researcher agency and careful management of information overload. The findings suggest that combining structured representations with conversational tools can support scalable, rigorous literature reviews, though future work should expand controls, multi-facet synthesis, and fine-grained discovery to further enhance efficacy and usability.

Abstract

Comprehensive literature review requires synthesizing vast amounts of research -- a labor intensive and cognitively demanding process. Most prior work focuses either on helping researchers deeply understand a few papers (e.g., for triaging or reading), or retrieving from and visualizing a vast corpus. Deep analysis and synthesis of large paper collections (e.g., to produce a survey paper) is largely conducted manually with little support. We present DimInd, an interactive system that scaffolds literature review across large paper collections through LLM-generated structured representations. DimInd scaffolds literature understanding with multiple levels of compression, from papers, to faceted literature comparison tables with information extracted from individual papers, to taxonomies of concepts, to narrative syntheses. Users are guided through these successive information transformations while maintaining provenance to source text. In an evaluation with 23 researchers, DimInd supported participants in extracting information and conceptually organizing papers with less effort compared to a ChatGPT-assisted baseline workflow.

Figures (7)

• Figure 1: Columns can be added to the literature review table in two ways: A) User-defined columns precisely specify a faceted information need and allow additional context for steering LLM assistance; B) System-suggested columns offer collection-aware recommendations for columns that can be added with a single click.
• Figure 2: In DimInd, users review large paper collections by navigating and analyzing information across various structured representations. Each cell in the literature review table is a snippet of faceted information from a paper (evidence snippet). Clicking on a snippet shows a popover with additional detail (evidence summary), with a button that can further open the paper PDF in an integrated paper reader with attributed paragraphs highlighted (evidence source). Faceted columns are transformed into distinct hierarchical taxonomies (facet taxonomy), which can be explored, refined, and used to controllably generate a narrative summary with citations (facet synthesis).
• Figure 3: The facet taxonomy. Each category shows the number of included papers (A). Users can manually refine the taxonomy through drag-and-drop interactions (B) or add additional categories (C). If at least one category is selected, the taxonomy can be summarized into prose (D).
• Figure 4: Selecting specific categories in the facet taxonomy: 1) highlights cells for the included papers in the literature review table, allowing users to quickly delineate between and browse the selected (and not selected) papers; 2) controls the structure and papers included in the generated summary.
• Figure 5: Users can view additional detail while exploring the synthesized representations: 1) Clicking an evidence snippet in the facet taxonomy shows the full evidence summary; 2) Clicking a citation in the facet synthesis shows an in-situ citation card. From either, users can click See in Table to scroll directly to the corresponding row in the literature review table.