Details
| Original language | English |
|---|---|
| Article number | 28 |
| Number of pages | 24 |
| Journal | ACM Transactions on Internet Technology |
| Volume | 17 |
| Issue number | 3 |
| Publication status | Published - Jun 2017 |
| Externally published | Yes |
Abstract
The argumentative structure of texts is increasingly exploited for analysis tasks, for example, for stance classification or the assessment of argumentation quality. Most existing approaches, however, model only the local structure of single arguments. This article considers the question of how to capture the global discourse-level structure of a text for argumentation-related analyses. In particular, we propose to model the global structure as a flow of "task-related rhetorical moves," such as discourse functions or aspectbased sentiment. By comparing the flow of a text to a set of common flow patterns, we map the text into the feature space of global structures, thus capturing its discourse-level argumentation. We show how to identify different types of flow patterns, and we provide evidence that they generalize well across different domains of texts. In our evaluation for two analysis tasks, the classification of review sentiment and the scoring of essay organization, the features derived from flow patterns prove both effective and more robust than strong baselines. We conclude with a discussion of the universality of modeling flow for discourse-level argumentation analysis.
Keywords
- Argumentation, Discourse-level structure, Sentiment flow
ASJC Scopus subject areas
- Computer Science(all)
- Computer Networks and Communications
Cite this
- Standard
- Harvard
- Apa
- Vancouver
- BibTeX
- RIS
In: ACM Transactions on Internet Technology, Vol. 17, No. 3, 28, 06.2017.
Research output: Contribution to journal › Article › Research › peer review
}
TY - JOUR
T1 - A Universal Model for Discourse-Level Argumentation Analysis
AU - Wachsmuth, Henning
AU - Stein, Benno
PY - 2017/6
Y1 - 2017/6
N2 - The argumentative structure of texts is increasingly exploited for analysis tasks, for example, for stance classification or the assessment of argumentation quality. Most existing approaches, however, model only the local structure of single arguments. This article considers the question of how to capture the global discourse-level structure of a text for argumentation-related analyses. In particular, we propose to model the global structure as a flow of "task-related rhetorical moves," such as discourse functions or aspectbased sentiment. By comparing the flow of a text to a set of common flow patterns, we map the text into the feature space of global structures, thus capturing its discourse-level argumentation. We show how to identify different types of flow patterns, and we provide evidence that they generalize well across different domains of texts. In our evaluation for two analysis tasks, the classification of review sentiment and the scoring of essay organization, the features derived from flow patterns prove both effective and more robust than strong baselines. We conclude with a discussion of the universality of modeling flow for discourse-level argumentation analysis.
AB - The argumentative structure of texts is increasingly exploited for analysis tasks, for example, for stance classification or the assessment of argumentation quality. Most existing approaches, however, model only the local structure of single arguments. This article considers the question of how to capture the global discourse-level structure of a text for argumentation-related analyses. In particular, we propose to model the global structure as a flow of "task-related rhetorical moves," such as discourse functions or aspectbased sentiment. By comparing the flow of a text to a set of common flow patterns, we map the text into the feature space of global structures, thus capturing its discourse-level argumentation. We show how to identify different types of flow patterns, and we provide evidence that they generalize well across different domains of texts. In our evaluation for two analysis tasks, the classification of review sentiment and the scoring of essay organization, the features derived from flow patterns prove both effective and more robust than strong baselines. We conclude with a discussion of the universality of modeling flow for discourse-level argumentation analysis.
KW - Argumentation
KW - Discourse-level structure
KW - Sentiment flow
UR - http://www.scopus.com/inward/record.url?scp=85058037852&partnerID=8YFLogxK
U2 - 10.1145/2957757
DO - 10.1145/2957757
M3 - Article
AN - SCOPUS:85058037852
VL - 17
JO - ACM Transactions on Internet Technology
JF - ACM Transactions on Internet Technology
SN - 1533-5399
IS - 3
M1 - 28
ER -