Aligning Discourse and Argumentation Structures using Subtrees and Redescription Mining

Laurine Huber, Yannick Toussaint, Charlotte Roze, Mathilde Dargnat, Chloé BraudACL 2019

TL;DR

This paper introduces a novel method for aligning discourse and argumentation structures in text by treating them as distinct graph representations. By applying subtree mining and redescription mining, the authors uncover systematic correspondences and differences between these two linguistic formalisms, offering a data-driven approach to understanding their expressive power.

What Problem Does This Paper Address?

Understanding how texts are structured is crucial for natural language processing. Two prominent ways to represent text structure are:

  • Argumentation Structures (ARG): These focus on how arguments are built, detailing relations like "support" and "attack" between claims and premises.
  • Discourse Structures (RST): These capture the coherence of a text by linking segments with semantico-pragmatic relations (e.g., "reason," "concession," "elaboration").

While both formalisms aim to describe text organization, their relationship and how they map onto each other remain an active area of research. Previous work often relied on manual mappings or focused on aligning individual relations, which can miss complex correspondences. This paper addresses the challenge of systematically and automatically comparing these two distinct structural representations at a more granular level – that of subtrees – to reveal deeper insights into their alignment and differences.

Methods

The authors propose a three-step methodology:

  1. Encoding Representations into Trees: Both ARG and RST annotations from the corpus are converted into distinct tree structures. The central claim/nucleus is designated as the root node. Edge labels represent the argumentative or discursive relations. Crucially, they adapt ARG representations to target the premise of an undercutted relation, ensuring a more consistent tree structure.
  2. Creating Two Views (Subtree Extraction): For each text, subtrees are extracted from its RST tree and its ARG tree separately. The gSpan algorithm is used to find frequent subtrees (those appearing in at least two texts) within each set of trees. Each unique subtree then becomes a boolean feature, indicating its presence in a given text's RST or ARG representation.
  3. Redescription Mining: This is the core of the alignment process. Redescription mining aims to find two distinct characterizations (queries) of the same set of objects (texts). In this context, one query is built over the RST subtree features, and the other over the ARG subtree features. The goal is to find pairs of queries (q1 for ARG, q2 for RST) where the set of texts satisfying q1 is very similar to the set of texts satisfying q2. This similarity is measured by the Jaccard index. The ReReMi algorithm is employed, using atomic updates and allowing conjunctions and disjunctions within queries, with a length restriction of 4 features per query.

Datasets

The study utilizes the arg-microtexts-multilayer corpus. This corpus contains 112 short texts, each annotated with three types of structures: RST trees, SDRT graphs, and ARG structures. For this study, the authors focus on the RST and ARG annotations. The corpus is described as containing texts answering controversial issues, such as "Should Germany introduce the death penalty?".

Key Insights and Ideas

  • Subtree-level Alignment is Powerful: Moving beyond aligning individual relations, comparing subtrees allows for the discovery of more complex and nuanced correspondences between discourse and argumentation structures.
  • Systematic Comparison: Redescription mining provides an automatic and systematic way to compare formalisms, overcoming the limitations of manual hypothesis-driven approaches.
  • Granularity Differences: The findings suggest that discourse structures (RST) are often more fine-grained than argumentation structures (ARG). This is evidenced by disjunctions in the RST query of a redescription, indicating that a single ARG structure can correspond to multiple RST structures.
  • Structural Depth Matters: Deeper or larger subtrees in one representation tend to align with deeper or larger subtrees in the other, suggesting a correlation in structural complexity.
  • Potential for Theory Bridging: The method offers a way to build bridges between different theoretical frameworks and to better understand the expressive power of each formalism.

Key Results

The paper reports on 31 redescriptions found over the corpus.

  • Rd1: The redescription with the highest Jaccard index (0.691) involved 54 texts. It highlighted a common ARG structure (feature a57) aligning with a disjunction of RST structures (r123 V r65 V r125). This suggests that a single argumentative pattern can be expressed through different discourse organizations.
  • Rd2 & Rd3: These are presented as specializations of Rd1, further refining the observed alignments. Rd3, in particular, showed an RST structure (r125) aligning with a disjunction of ARG structures (a59 or a23), again pointing to flexibility in how argumentation is realized in discourse.
  • Relation Mapping: The study observed that a single ARG relation (like "support") can map to various RST relations (e.g., "reason," "concession," "motivation," "list," "conjunction"). This confirms previous findings about the varied ways arguments are signaled in discourse.

Limitations and Future Work

  • Tree Representation Omission: A significant limitation is that the current tree representation omits the positional information of text segments. This means that aligned features (subtrees) might not correspond to the same part of the original text, leading to potential misalignments.
  • Query Length Restriction: The restriction of query length to 4 features in redescription mining limits the emergence of more complex or deeper subtrees.
  • Minimum Support Threshold: The gSpan algorithm's minimum support threshold (set to 2) means that very rare subtrees, even if interesting, are not captured.
  • Future Work:
    • Parametrization of ReRemi: Exploring different parameters for the redescription mining algorithm, including restricting ARG queries to atomic ones or increasing query length.
  • Including Text Segments: Incorporating links to text segments in the tree representations to enable more precise alignments.
  • Extending to Other Formalisms: Applying the methodology to other discourse and argumentation formalisms, such as SDRT.
  • Grouping Substructures: Using the approach to group substructures within a single theory.

Summary updated

BibTeX
@inproceedings{huber2019aligning,
  author = {Laurine Huber and Yannick Toussaint and Charlotte Roze and Mathilde Dargnat and Chloé Braud},
  title = {{Aligning Discourse and Argumentation Structures using Subtrees and Redescription Mining}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure