自然语言界面数据可视化（英）-20页_16mb

Summary of "Towards Natural Language Interfaces for Data Visualization: A Survey"

Core Content

This survey provides a comprehensive overview of the development and current state of Visualization-oriented Natural Language Interfaces (V-NLIs) over the past two decades. It explores how these interfaces can serve as a complementary input modality to traditional WIMP (Windows, Icons, Menus, and Pointer) interaction in data visualization, enabling users to express their analytical tasks through natural language rather than through complex interface operations. The paper outlines a classification framework based on the classic information visualization pipeline, with the addition of a V-NLI layer, and discusses the seven key stages involved in the process: Query Understanding, Data Transformation, Visual Mapping, View Transformation, Human Interaction, Context Management, and Presentation.

Main Points

• User Experience Enhancement: V-NLIs offer a more convenient, intuitive, and humanistic way for users to interact with data visualization tools, especially for novices and people with visual impairments.
• Development Timeline: V-NLIs have evolved through three phases: infancy (2001), development (2010s), and outbreak (2018–2021), marked by significant advancements in NLP and increased commercial adoption.
• Key Stages in V-NLI Pipeline:
  • Query Understanding: Involves semantic and syntax analysis to extract data attributes and analytic tasks. It also addresses underspecified utterances.
  • Data Transformation: Converts raw data into structured formats, often involving aggregation and pivoting.
  • Visual Mapping: Maps extracted information to visual structures using spatial substrate, graphical elements, and properties.
  • View Transformation: Transforms visual structures into views for user interaction, though this stage is less commonly addressed in V-NLIs.
  • Human Interaction: Enables users to interact with the visualization interface, which feeds back into the pipeline.
  • Context Management: Facilitates conversation with the system by considering the current visualization state and previous user inputs.
  • Presentation: Focuses on displaying the final visualization, often directly from natural language input.

Key Information

• NLP Technologies: The paper reviews various NLP toolkits such as CoreNLP, NLTK, OpenNLP, SpaCy, Stanza, Flair, and GoogleNLP, highlighting their capabilities in tasks like tokenization, parsing, and named entity recognition.
• Survey Methodology: The authors conducted an exhaustive review of 55 V-NLI-related papers from 2000 to 2021, spanning journals and conferences in VIS, HCI, NLP, and DMM. They also collected 283 related works across multiple topics.
• V-NLI Systems: A list of representative systems is provided, including Cox et al. [41], Articulate [241], DataTone [64], Eviza [207], FlowSense [280], InChorus [229], Ask Data [2], Data@Hand [278], and NL4DV [174], among others. These systems vary in the NLP technologies used, supported visualization types, and recommendation algorithms.
• Future Directions: The paper highlights several promising research areas, including the integration of pretrained language models, multimodal interfaces, and improving ambiguity management in natural language queries.

Classification Overview

The paper extends the classic information visualization pipeline with a V-NLI layer, enabling a structured categorization of existing systems. The classification dimensions are based on the seven stages of the pipeline and help in organizing the research landscape, identifying knowledge gaps, and guiding future work in the field.

Conclusion

The survey emphasizes the importance of V-NLIs in making data visualization more accessible and intuitive. It calls for a systematic review of the literature to better understand the challenges and opportunities in this growing area of research.