ADROIT: Automatic Discourse Relation Organizer of Internet-based Text

A. S. M. Mahbub Morshed, Mitsuru IshizukaAAAI 2008

Introduction

Understanding the logical flow of human language—how one sentence provides a reason, a contrast, or an elaboration for another—is a fundamental challenge in natural language processing. This field, known as discourse analysis, typically utilizes Rhetorical Structure Theory (RST) to map these connections. While existing discourse parsers perform well on highly structured, professionally edited text like news articles, they often struggle with the "semi-coherent" nature of the internet, including blog posts, web pages, and forum threads.

ADROIT (Automatic Discourse Relation Organizer of Internet-based Text) is an automated system designed to address this gap. It identifies rhetorical relationships between sentences in messy web content and organizes them into structured discourse trees, even when explicit "cue phrases" (like "because" or "however") are absent.

System Architecture

The ADROIT system (version 1.0) operates through a modular three-stage pipeline that transforms raw HTML documents into structured XML discourse trees.

The ADROIT system architecture, illustrating the pipeline from raw HTML to structured XML discourse trees.

1. Discourse Segmenter (DS)

The process begins by breaking raw text into Elementary Discourse Units (EDUs), which are the smallest building blocks of a claim. Because nearly 60% of segment boundaries lack punctuation or explicit cue phrases, the DS module utilizes a Support Vector Machine (SVM) trained on deep syntactic features, including Part-of-Speech (POS) tags and syntactic information, rather than relying on shallow pattern recognition.

2. Core Relation Recognizer (RR)

The RR module identifies how these segments relate to one another. To improve robustness and reduce computational noise, the authors collapsed 110 complex relations from standard linguistic corpora into 10 coarse-grained classes. This module leverages syntactic cues, temporal relations, and semantic similarity derived from external knowledge sources like WordNet and domain-specific ontologies.

3. Discourse Organizer (DO)

The final stage acts as the system's "architect," building the tree structure. To manage the massive search space of potential connections in long-form web text, it employs "block comparison" (using tf-idf and correlation coefficients) to identify relationships between widely separated segments. Notably, ADROIT does not force the entire text into a single, massive tree; if a text is semi-coherent, it can produce multiple smaller trees (a "forest"), which more accurately reflects the disjointed nature of many web documents.

Evaluation and Key Results

The authors utilized the RST Discourse Treebank (RST-DT) corpus to train their models and benchmarked the system against SynDS, a state-of-the-art segmenter.

  • Competitive Segmentation: The Discourse Segmenter performed on par with SynDS, demonstrating the effectiveness of the SVM-based syntactic approach.
  • Long-Distance Cohesion: The block comparison method successfully identified relationships between segments far apart in the text, a common failure point for automated parsers.
  • Practical Application: In walk-through tests using the RST-DT corpus, ADROIT correctly segmented complex sentences and accurately identified "Elaboration" and "Temporal" relations, producing XML output that aligned with human-annotated standards.

Conclusion and Future Work

ADROIT represents a significant step toward making discourse analysis practical for the vast, unorganized data of the internet. By prioritizing coarse-grained relations and flexible tree structures, the system provides a more realistic model for web-based text. Future development aims to integrate the ADROIT engine into dialogue generation systems, allowing AI to synthesize information from the web into more natural, rhetorically sound conversational formats.

Summary updated

BibTeX
@inproceedings{morshed2008adroit,
  author = {A. S. M. Mahbub Morshed and Mitsuru Ishizuka},
  title = {{ADROIT: Automatic Discourse Relation Organizer of Internet-based Text}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
}
Figure