TL;DR
This paper argues that "attribution" in text isn't a single, distinct rhetorical relation but rather a construct that can be realized by various existing Rhetorical Structure Theory (RST) relations, depending on the writer's intended effect. By analyzing attributions through the lens of intended effect, the paper proposes a more nuanced understanding of how reported speech and thoughts function in discourse.
What Problem Does This Paper Address?
Rhetorical Structure Theory (RST) is a framework for analyzing the functional organization of text, focusing on how different parts of a text work together to achieve a writer's intended effect. A long-standing debate within RST concerns the status of "attribution" – how to classify statements that report what someone else said or thought. While many studies use attributional structures, there's no consensus on whether "ATTRIBUTION" is a fundamental RST relation itself. Previous attempts to define it have struggled with issues like what constitutes a "discourse unit" for attribution, which part is the nucleus (main point) and which is the satellite (support), and how to identify the specific relation. This paper aims to resolve these issues by re-examining attributions through the core RST principle of "intended effect."
Methods
The paper employs a classical RST analysis, emphasizing the writer's intended effect as the primary determinant for identifying relational structures. This approach prioritizes the functional purpose of the text over purely grammatical or syntactical features. The author revisits existing RST analyses of texts containing attributions and provides original analyses to support the claim that attributions are not a single relation but instances of other RST relations. The analysis focuses on three core issues:
- Discourse Units Issue: Can the components of an attribution (predicate and attributed material) be considered distinct discourse units?
- Nuclearity Issue: Which part of an attribution (predicate or attributed material) is the nucleus and which is the satellite?
- Relation Identification Issue: What specific RST relation(s) are at play when attributions occur?
Datasets
The paper analyzes existing RST treebanks and examples from various studies, including:
- Carlson and Marcu (2001)
- Redeker and Egg (2006)
- Taboada and Hay (2008)
- Das and Taboada (2013)
The author also provides several original analyses of texts. No new datasets were introduced.
Key Insights and Ideas
- Intended Effect is Key: The paper strongly advocates for "intended effect" as the crucial factor in analyzing attributions, arguing that it has been largely overlooked in previous discussions.
- Attribution is Not a Single Relation: The central claim is that "ATTRIBUTION" is not a distinct RST relation. Instead, attributions are constructs that can realize various RST relations like JUSTIFY, ELABORATION, EVIDENCE, EVALUATION, INTERPRETATION, and CAUSE, depending on the writer's goal.
- Nuclearity Varies: The nuclearity of an attribution (which part is the nucleus) is not fixed. It depends on which part carries the more salient intended effect. For example, in EVALUATION or EVIDENCE, the attributed material is often the nucleus, while in ELABORATION or CAUSE, the attribution predicate might be the nucleus.
- Discourse Units are Functional: The paper argues that if attributional components participate in discourse relations, they are discourse units by definition, challenging purely syntactical grounds for exclusion.
- "Faux Concession": A particularly interesting insight is the use of cognitive states (e.g., "I think") to moderate claims, not to genuinely concede uncertainty, but to soften delivery and enhance acceptance. This is termed "faux concession" and can be blended with other relations like JUSTIFY.
Key Results
The paper demonstrates through analysis that attributions can function as:
- JUSTIFY: Providing authority or credibility for a claim (e.g., citing financial analysts for sales figures).
- EVALUATION/INTERPRETATION: Assessing or interpreting a situation (e.g., a surgeon's reaction to a patient's survival).
- CAUSE: The attributed material being the cause of a cognitive state (e.g., a low ranking causing outrage among experts).
- ELABORATION: Providing additional detail about the activity in the attribution predicate (e.g., reporting a politician's statement on a bill).
- EVIDENCE: Increasing the reader's belief in the nucleus (e.g., a promise from a specific type of driver lending credibility).
- ATTRIBUTION (as a textual relation): In some cases, attributions serve a purely organizational function, providing orienting information without a strong specific rhetorical purpose beyond reporting.
The analysis resolves the nuclearity discrepancy between Carlson and Marcu (nucleus = attributed material) and Redeker and Egg (nucleus = attribution predicate) by showing that nuclearity is context-dependent and tied to intended effect.
Limitations and Future Work
The author acknowledges that the identified relations are not exhaustive and further analysis may uncover more. The paper suggests that the criteria used by some researchers (like Carlson and Marcu) for excluding certain constructs (e.g., infinitival complements, passive constructions) might be too restrictive and rhetorically arbitrary. Future work could explore how context helps determine the specific relation in ambiguous cases and further investigate the "faux concession" phenomenon. The paper also touches on the potential for this analysis to lead to intraclausal relations, a direction already being explored in the field.
Summary updated