Introduction
Researchers have introduced WebDP, a novel benchmark and task designed to map the "discourse structure" of web pages. Unlike traditional text analysis, WebDP focuses on how different blocks of content semantically relate to one another in the complex, multi-topic, and hierarchical environment of the modern web. By moving beyond superficial layout parsing, this work provides a framework and a manually annotated dataset (WEBDOCS) to help AI understand the deep semantic organization of semi-structured documents.
The Challenge: Beyond Layout and Linear Text
Most existing discourse analysis—the study of how sentences or segments form a coherent whole—has focused on plain, linear text such as news articles or essays. Conversely, "Document Intelligence" research often prioritizes visual layout, focusing on where a box is located on a screen or its specific HTML tag.
However, web documents are "semi-structured" and "free-styled." They frequently contain multiple topics (e.g., a news roundup) and complex hierarchies. Crucially, the visual layout or HTML structure does not always align with the actual semantic meaning. Two blocks might look visually identical but belong to entirely different topics. Current models struggle to bridge the gap between "where a block is" and "what it means in the context of the whole page."
As shown in the comparison above, WebDP fills a critical gap. While traditional Discourse Parsing (like RST or PDTB) handles semantics in short, single-topic texts, and Document Intelligence handles complex layouts without deep semantic modeling, WebDP addresses long web documents with multiple topics using physical blocks as the basic unit of analysis.
The WebDP Framework
The authors propose a new schema for Web Document Discourse Parsing (WebDP). Instead of analyzing individual clauses, they treat "blocks" (paragraphs, headings, or captions) as Elementary Discourse Units (EDUs). Their representation utilizes two main types of rhetorical relations:
- Subordinating Relations: These describe hierarchical "parent-child" connections, such as a paragraph elaborating on a heading or an attribute like a source line for an image.
- Coordinating Relations: These describe "sibling" connections at the same level, such as a list of items or a narration of events.
A standout feature is the "Break" relation, specifically designed to handle web pages with multiple independent topics. If two adjacent blocks have no semantic connection, they are linked by a "Break," allowing the model to recognize topic shifts and maintain a "forest" of discourse trees rather than forcing a single root.
As illustrated in the figure above, the visual layout often suggests a simple hierarchy (left) that fails to capture the actual semantic flow. The WebDP structure (right) correctly identifies when content is coherent (e.g., narration) or incoherent (requiring a "Break" relation), providing a much more accurate map of the document's meaning.
The WEBDOCS Dataset
To evaluate this framework, the researchers introduced WEBDOCS, a manually annotated dataset of 300 web documents collected from WeChat Official Accounts.
- Scale: The dataset contains 14,101 EDUs, averaging approximately 47 blocks per document.
- Complexity: These documents are significantly longer and more complex than traditional discourse datasets, which usually average only 10–20 units per document.
- Language: While the current version is in Chinese, the underlying framework is language-agnostic and applicable to any web-based content.
Key Insights and Results
The study yielded several critical findings regarding the nature of web discourse:
- Layout is a Liar: Superficial visual structures often fail to represent underlying discourse. Models relying solely on "boxes" miss deep semantic links.
- Global Context Matters: Ablation studies showed that "Global Context Encoders" are vital. Because web documents are long, models need a "bird's-eye view" to understand how a block at the bottom might relate to a heading at the top.
- XPath is a Powerful Signal: Including the HTML path (the block's "address" in the code) significantly boosted performance, proving that while layout isn't everything, it provides essential hints for discourse.
In testing state-of-the-art parsing models (like SSAGNN and DAMT), the best models achieved an Unlabeled Attachment Score (UAS) of ~63% and a Labeled Attachment Score (LAS) of ~55%. In contrast, the human baseline is ~88% UAS and ~81% LAS. This massive gap indicates that while the task is feasible, current AI still lacks the "exquisite semantic discrimination" needed to truly understand web discourse.
Limitations and Future Work
The authors acknowledge several hurdles, including the labor-intensive nature of annotating such complex structures and the current focus on news-style reports. Future work aims to expand the dataset to other domains like e-commerce or forums, explore multi-lingual support, and develop more data-efficient models that can learn without requiring thousands of manual annotations.
Summary updated