Abstract
Modality is the linguistic ability to describe events with added information such as how desirable, plausible, or feasible they are. Modality is important for many NLP downstream tasks such as the detection of hedging, uncertainty, speculation, and more. Previous studies that address modality detection in NLP often restrict modal expressions to a closed syntactic class, and the modal sense labels are vastly different across different studies, lacking an accepted standard. Furthermore, these senses are often analyzed independently of the events that they modify. This work builds on the theoretical foundations of the Georgetown Gradable Modal Expressions (GME) work by Rubinstein et al. (2013) to propose an event-based modality detection task where modal expressions can be words of any syntactic class and sense labels are drawn from a comprehensive taxonomy which harmonizes the modal concepts contributed by the different studies. We present experiments on the GME corpus aiming to detect and classify fine-grained modal concepts and associate them with their modified events. We show that detecting and classifying modal expressions is not only feasible, but also improves the detection of modal events in their own right.
Introduction
In linguistics, modality serves as a fundamental mechanism for articulating alternative states of the world. It distinguishes factual assertions (e.g., "I am writing a paper") from non-actualized possibilities, obligations, or desires (e.g., "I ought to write a paper"). Despite its prevalence, Natural Language Processing (NLP) has historically struggled with robust modality detection due to three primary limitations:
- Syntactic Narrowness: Traditional systems often focus exclusively on a "closed class" of auxiliary verbs (e.g., can, might, should), ignoring modal triggers that appear as nouns, adjectives, or adverbs.
- Taxonomic Inconsistency: The field lacks a standardized taxonomy for modal senses, leading to fragmented studies that conflate distinct concepts like "permission" and "requirement."
- Event Disconnection: Modality is frequently analyzed in isolation, failing to explicitly link the modal "trigger" to the specific "event" it modifies.
This paper addresses these gaps by proposing a comprehensive framework for event-based modality detection, grounding modal expressions in a unified taxonomy and linking them directly to event heads.
Methodology and Taxonomy
The core of this work is the definition of event-based modality detection as a task involving the identification of a modal trigger, its specific sense, and the "head" of the event it modifies.
A Unified Taxonomy
To resolve the issue of label inconsistency, the authors introduce a hierarchical taxonomy that harmonizes existing linguistic theories. As illustrated in Figure 1, modality is divided into two primary branches:
- Priority: Pertains to events that are prioritized, required, or desired (e.g., Rules/Norms, Desires/Wishes, Plans/Goals).
- Plausibility: Pertains to the likelihood or feasibility of events (e.g., Knowledge/Epistemic, State of the World, or Agent Ability).
Dataset and Modeling
The study utilizes the Georgetown Gradable Modal Expressions (GME) corpus, which consists of over 11,000 sentences from news articles. Unlike datasets focused solely on uncertainty, GME provides expert-annotated fine-grained modal properties. The authors adapted this corpus by converting "event spans" into "event heads"—typically the primary verb or noun of the action—to facilitate more consistent model learning.
The task is modeled as a sequence labeling problem using RoBERTa. The experiments explore various context levels, ranging from simple baselines to models that integrate information about both the trigger and its associated event head.
Insights: The Diversity of Modality
A key finding of this research is that modality is ubiquitous, appearing in 48% of the sentences in the GME corpus. Crucially, the study confirms that restricting detection to auxiliary verbs is insufficient for capturing the full scope of modal language.
As shown in Figure 3, modal triggers are distributed across various syntactic categories. While verbs account for 37% of triggers, a significant 38% are comprised of nouns, adjectives, and adverbs. This diversity underscores the necessity of an open-class approach to modality detection. Furthermore, approximately 25% of triggers are ambiguous (e.g., "can" denoting either physical ability or social permission), making context—specifically the identity of the event head—critical for accurate classification.
Experimental Results
The experimental results demonstrate the feasibility of the proposed task:
- Trigger Detection: RoBERTa achieved an F1 score of 73.2 for detecting modal triggers across all parts of speech, significantly outperforming the majority-vote baseline of 68.2.
- Sense Classification: In an "Oracle" setting where the trigger is known, the model effectively distinguished between the high-level "Priority" and "Plausibility" categories, reaching a 90.7 F1 score.
- Fine-Grained Challenges: Performance naturally decreases with increased granularity. Distinguishing between closely related sub-categories like "Plans" and "Desires" remains difficult, with a fine-grained F1 of 58.14.
- Mutual Reinforcement: An important observation is that detecting the event head becomes easier when the model is aware of the modal trigger, suggesting that these two components of the task are mutually reinforcing.
Conclusion and Future Work
This work provides a linguistically grounded and computationally robust framework for navigating the "possible, plausible, and desirable" in text. While data sparsity in certain categories remains a challenge, the transition to event-based detection marks a significant step forward for NLU. Future research may focus on extracting full "prejacents" (entire modified clauses) and integrating this modality layer into broader semantic representations like Abstract Meaning Representation (AMR).
Summary updated