PaperlayerView exact revision on arXiv

Source-linked AI summary

EntityWeaver: Visual Exploration and Curation of Named-Entity Relationships in Document Collections

Uroš Šmajdek, Ciril Bohak

arXiv:2609.37329v1cs.HC

TL;DR

Entity extraction can yield uncertain links, while visual systems may present those relationships as fixed even when analysts need to validate them. EntityWeaver combines an uncertainty-aware graph with coordinated source-text views and adjustable encoding; expert feedback suggests this approach can enhance exploration and curation of named entities.

  • Problem

    Entity extraction produces ambiguous or probabilistic assignments, yet many visualization systems present relationships as fixed even when analysts need to validate and refine them.

    Source: §1 · p. 1

  • Method

    EntityWeaver links documents, mentions, and entities in a probabilistic graph, exposes link uncertainty, and coordinates graph exploration with source-text inspection.

    Source: §1 · p. 2, §4.1 · p. 6

  • Results

    Expert feedback suggests uncertainty-aware visualization can meaningfully enhance named-entity exploration and curation in multi-document corpora.

    Source: §6 · p. 13

  • Takeaways & Limitations

    The system supports large-scale analysis and fine-grained interpretation by combining flexible filtering and querying with coordinated close and distant reading.

    Source: §7 · pp. 13–14

    Very large corpora can challenge performance, particularly during interactive query refinement or uncertainty-encoding updates.

    Source: §6 · p. 13

Abstract

from arXiv · show

We present an interactive visualization system for exploring named entities and their relationships across document collections, with a strong focus on handling uncertainty and supporting both distant and close reading. The system is built around a graph that links documents, entity mentions, and entities. Uncertainty from mention-to-entity linking is included directly in this graph, so users can see where connections are strong, weak, or ambiguous. A transfer-function control, inspired by approaches in scientific visualization, allows users to adjust how this uncertainty is displayed, making it easy to tune the visualization for different datasets and research questions. The system also provides direct access to the full source texts in a coordinated view, enabling quick context checks, resolving ambiguous cases, and correcting digitization errors. Exploration is further supported through a multi-stage filter query builder, mini-map navigation for large graphs, and export options for downstream analysis. By combining uncertainty-aware graph visualization with direct interaction in the source texts, the system provides a unified workflow that supports both large-scale pattern discovery and curation of named-entity-based document collections. The design choices were supported by the domain experts, who were also involved in the initial system evaluation.

1 Introduction

Entity-extraction uncertainty can make visualized relationships unreliable, yet many systems present them as fixed even when analysts need to validate and refine them. EntityWeaver addresses this gap with uncertainty-aware graph exploration linked to source-text curation.

1 Introduction

  • Entity-extraction uncertainty can make relationships ambiguous or unreliable, while fixed visualizations can misrepresent them during validation and refinement.Analysts need to assess relationship reliability and provenance and justify corrections.Source: §1 · p. 1
  • EntityWeaver represents documents, mentions, and resolved entities in a probabilistic multi-layer graph that exposes linking uncertainty directly.The graph retains uncertainty from NLP-based entity linking as part of the structure used for visual exploration.Source: §1 · p. 2
  • A tunable transfer function lets users adjust uncertainty encoding for exploratory pattern analysis or conservative validation workflows.The visual emphasis can be matched to dataset characteristics and analytic goals.Source: §1 · p. 2
  • Coordinated source-text views let analysts inspect context, resolve ambiguous links, and correct extraction or digitization errors.This combines graph-based distant reading with source-level close reading and curation.Source: §1 · p. 2
  • A multi-stage query builder, minimap navigation, and export mechanisms support iterative refinement and downstream analysis.These interaction techniques complement the uncertainty-aware graph and text views.Source: §1 · p. 2

2 Related Work

Prior work combines structured entity extraction with corpus visualization, but existing approaches address entity-level attributes, uncertainty, and source-grounded curation unevenly. EntityWeaver brings uncertainty into both graph representation and interaction while connecting distant and close reading.

2.1 Named Entity Visualization

  • Named-entity visualization brings together entity extraction and large-scale visual exploration of textual corpora.NER detects and classifies entities, while NEL assigns identifiers to instances to support consistent cross-document tracking.Source: §2.1 · p. 2
  • Named-entity visualization systems vary in the granularity and modalities they support, and user-defined entity attributes remain underused in general entity-centric analysis.Coordinated views and interactive filtering are established graph-visualization principles.Source: §2.1 · p. 3
  • Entity-focused systems range from close/distant reading views to corpus biography graphs and queryable knowledge graphs, but often center network structure over entity attributes.The cited systems include Barros Rodrigues et al. (2024), El-Assady et al. (2017), Tamper et al. (2023), and Kusnick et al. (2024).Source: §2.1 · p. 3

2.2 Uncertainty in Network Graphs

  • Uncertainty in text-derived networks arises from OCR, NER, and NEL, and digital-humanities applications of uncertainty-aware network visualization remain limited.Humanities knowledge graphs can be treated as authoritative despite their interpretive or uncertain evidence.Source: §2.2 · pp. 3–4
  • EntityWeaver integrates uncertainty into entity–mention visual encoding and interaction, linking distant exploration to source-context inspection for validation and refinement.This distinguishes its focus from systems limited to network structure or uncertainty annotation.Source: §2.2 · p. 4

3 System Requirements

Six requirements guide EntityWeaver toward flexible exploration and curation of named entities across documents, retaining uncertainty and connecting corpus patterns with textual evidence.

R1 Integrated data exploration and visualization

  • R1 calls for flexible filtering and querying plus direct graph editing so analysts can iteratively refine large-corpus views.Manual subset selection can be time-consuming when the relevant scope is initially unknown.Source: §3 · p. 4

R2 Cross-document entity relations

  • R2 requires linking textual references to the same entity across documents despite variation in forms, spelling, context, or language.The requirement covers both within-document and cross-document references.Source: §3 · p. 4

R3 Extensible entity schema

  • R3 requires users to define, attach, and modify custom entity attributes and use them in visualization and filtering.Source: §3 · p. 4

R4 Retaining the uncertainty of entity data

  • R4 requires retaining multiple candidate entity links and their confidence scores rather than only the best guess.The requirement is motivated particularly by historical or heterogeneous corpora with incomplete metadata and textual variation.Source: §3 · p. 5

R5 Coordinated close and distant reading

  • R5 requires synchronized interaction between graph views, source passages, and metadata panels to connect close and distant reading.Graph navigation should reach underlying text, while textual inspection contextualizes visual analysis.Source: §3 · p. 5

R6 Data curation

  • R6 requires editing links and attributes, resolving or merging entities, and exporting curated data for downstream analysis.The requirement addresses correction, refinement, and enrichment of automatically extracted datasets.Source: §3 · p. 5

4 System Design

EntityWeaver is a web-based named-entity curation and visualization system organized around coordinated text, graph, query, filtering, and curation views.

4 System Design

  • EntityWeaver coordinates interactive text exploration, graph visualization, querying, filtering, and data curation in a web-based system.It is implemented with HTML, CSS, TypeScript, and React.Source: §4 · p. 5
  • Structured models for text, graph nodes, edges, and metadata support storage, retrieval, and modification of graph information.Source: §4 · p. 5
  • A dedicated visualization pipeline computes layouts, applies styling, manages interactions, and renders the network for exploration and analysis.Source: §4 · p. 5

4.1 Data Model

EntityWeaver models documents, textual entity mentions, and resolved entities as distinct node types, retaining candidate entity links and their uncertainty. It also derives virtual entity-to-document links for cross-document analysis.

4.1 Data Model

  • The graph distinguishes document nodes, mention nodes for passage-level entity instances, and entity nodes created through named entity linking.This simultaneous representation supports both individual-document and cross-document entity relations.Source: §4.1 · p. 5
  • Mention-to-document and mention-to-entity edges come from input data, and a mention may retain multiple candidate entity links.Mention-to-entity edges can carry confidence scores; when absent, the system assigns 1/n for n candidate entities.Source: §4.1 · p. 6
  • Virtual entity-to-document edges link each entity to every document where it is mentioned, with confidence set to the maximum among contributing mention-to-entity edges.These edges are derived from existing mention-to-document and mention-to-entity connections.Source: §4.1 · p. 6
  • Figure 2 presents an overview of the internal entity-relation data model.The model represents documents, mentions, and entities as separate node types.Source: §4.1 · pp. 5–6, Figure 2 · p. 6

4.2 Visualization Pipeline

EntityWeaver’s visualization pipeline turns user queries and graph data into an interactive graph through staged querying, view selection, filtering, mapping, layout, and rendering. Users can switch between detailed and abstract graph views and tune visual encodings, including those for uncertainty.

4.2 Visualization Pipeline

  • The pipeline processes user inputs through querying, view selection, filtering, mapping, layout creation, and rendering, recomputing a stage only when a preceding stage changes.The stages transform raw data and user inputs into a network graph.Source: §4.2 · p. 7

4.2.1 Data Querying

  • Rule-based query groups target node types and execute sequentially; groups support string, numeric, range, and categorical conditions combined with AND/OR logic.The resulting nodes satisfy relevant query groups or connect to nodes returned by an earlier group.Source: §4.2.1 · p. 7

4.2.2 View Selection

  • The D–M–E view retains mention-level detail, whereas the D–E view removes mention nodes to expose cross-document relations with less visual clutter.The D–E representation uses virtual entity-to-document edges in place of mention nodes.Source: §4.2.2 · p. 7

4.2.3 Data Filtering

  • Selection filters highlight a chosen node or edge, while node-focus filters retain only a specified node and its directly connected neighborhood.Both can run concurrently without removing the focused node’s edges.Source: §4.2.3 · p. 8

4.2.4 Data Mapping

  • Nodes use color for categorical or numerical attributes, glyphs for a core categorical attribute, and border color to coordinate focus across views.Users can assign categorical colors manually, while numerical color uses a transfer-function widget.Source: §4.2.4 · pp. 8–9
  • The mapping stage encodes node attributes and mention–entity uncertainty, with a tunable transfer function for numerical attributes and confidence relationships.The transfer-function mechanism supports continuous encodings without predefined interpretations and can suppress low-confidence edges or highlight ambiguous links.Source: §4.2.4 · pp. 8–10

4.2.5 Layout Creation

  • Layout positions can be set manually or computed with ForceAtlas2, whose Web Workers implementation keeps the interface responsive during computation.The layout stage positions nodes and edges to represent relational structure while preserving readability and reducing clutter.Source: §4.2 · p. 7, §4.2.5 · p. 10

4.2.6 Rendering

  • Sigma.js uses GPU-based WebGL rendering for dense graphs, but rasterized output does not support vector-based external post-processing or editing.In-system editing tools let users adjust the visualization directly.Source: §4.2.6 · p. 10

4.3 User Interface

EntityWeaver coordinates graph exploration, source-text reading, and node curation in three linked views. Mapping and navigation tools support inspection across dense graphs and the underlying dataset.

4.3 User Interface

  • The central graph supports network overview and node or edge inspection, while search, filters, selection, deletion, and a minimap aid navigation in dense graphs.Source: §4.3 · pp. 11–12, Figure 8 · p. 12
  • The document editor highlights mentions and synchronizes selections and hover between source text and graph in both directions.Source: §4.3 · p. 11
  • The node-curation panel exposes metadata and connections and synchronizes edited attributes across the editor and graph.Source: §4.3 · p. 11
  • The edge color mapper supports uncertainty mapping informed by an edge-value distribution, while the dataset table enables attribute-based sorting.Source: §4.3 · pp. 11–12

5 Expert Evaluation

A Digital Humanities scholar informed system requirements through consultation and two rounds of prototype evaluation. Feedback shaped filtering, uncertainty visualization, close reading, and support for varied entity metadata.

5 Expert Evaluation

  • A Digital Humanities scholar with more than 15 years of corpus-analysis experience collaborated on the system’s design.Source: §5 · p. 12
  • Consultation identified the labor cost of curating high-recall entity sets, leading the team to support broad candidate sets, flexible filtering, and web-based access.Source: §5 · p. 12
  • Prototype testing found node and edge editing cumbersome and dense graphs difficult to filter, prompting expanded requirements and content filtering.Source: §5 · p. 12
  • A second evaluation highlighted difficulty judging uncertain links and frequent source-text checks, motivating clearer confidence visualization and coordinated close and distant reading; requests also reinforced support for additional entity types and metadata.Source: §5 · p. 13

6 Discussion

Expert feedback supports the value of uncertainty-aware exploration and curation, while the discussion identifies limits in evaluation breadth, collaboration, export, scalability, and workflow integration.

6 Discussion

  • Expert feedback highlights the value of alternative entity candidates and continuous uncertainty control in historical and heterogeneous corpora.Source: §6 · p. 13
  • Evaluation with one domain expert does not establish generality across fields; additional expert sessions, controlled studies, or longitudinal deployments could assess broader usability and applicability.Source: §6 · p. 13
  • The single-user, client-side implementation does not provide collaborative workflows, shared datasets, versioning, or conflict resolution, and browser memory limits large corpora.Source: §6 · p. 13
  • Basic exports lack interoperability with common DH formats and visualization-state export, while rasterized WebGL limits vector graphics; structured data export offers an alternative.Source: §6 · p. 13
  • Very large corpora can challenge performance during query refinement and uncertainty updates; server-side preprocessing, hierarchical abstractions, and relevance-driven sampling are proposed directions.Source: §6 · p. 13
  • Embedding EntityWeaver in OCR correction, annotation environments, or web-based text repositories could strengthen its role in entity-centric research pipelines.Source: §6 · p. 13

7 Conclusion

EntityWeaver combines document-, mention-, and entity-level representations with filtering, querying, and coordinated close and distant reading. The conclusion presents it as a foundation whose broader evaluation and technical extensions remain future work.

7 Conclusion

  • EntityWeaver integrates document-, mention-, and entity-level representations with flexible filtering and querying and coordinated close and distant reading.Source: §7 · pp. 13–14
  • The domain-expert collaboration highlighted the value of alternative entity candidates, transfer-function uncertainty visualization, and editable graph structures for iterative refinement.Source: §7 · p. 14
  • Further evaluation across scholars, collaborative workflows, additional export formats, and enhanced scalability remain future directions for the prototype.Source: §7 · p. 14
Loading 2609.37329v1…