Filtering Non-Asserted Events via Modality Analysis

40 concepts26 papers1 proposal

When a natural language processing model encounters the sentences "I caught the flu," "I might have the flu," and "If you get the flu, stay home," a naive keyword or action-verb classifier flags an identical event trigger: flu/catch. However, from a real-world utility perspective, only the first sentence describes a factual, grounded event. The others describe possibilities, conditions, or hypothetical scenarios.

This distinction is known as an event's realis status or factuality. As frameworks like the TAC KBP Event Nugget tracks have demonstrated (Mitamura et al.; Liu et al.; Feng et al.), simply identifying what type of event is mentioned is insufficient for deep semantic understanding. To build reliable downstream applications, models must determine whether the event actually occurred in the real world. Without this capability, information extraction pipelines become overwhelmed by speculative noise, rumors, and unfulfilled intentions.

The Anatomy of Modality Filtering

How do humans instantly know that "might have" isn't a fact? We rely on a dense web of linguistic cues: grammatical tense, mood, auxiliary verbs, and syntactic structures. Modality filtering formalizes these cues into computational layers to isolate asserted occurrences from non-grounded events.

1. Grammatical and Lexical Markers

Systems can leverage explicit grammatical features to separate lived experiences from hypothetical statements. For instance, Park et al. demonstrated that prioritizing the indicative mood and past/present tenses while explicitly filtering out modal verbs like should or must isolates actual personal experiences in weblogs. Similarly, in social media streams, Li et al. and Akbari et al. showed that combining verb tense with first-person pronouns and conditional auxiliary verbs effectively filters out user wishes, future plans, or third-party discussions, leaving only realized personal milestones or wellness behaviors.

2. Predicate-Argument Structures

Beyond isolated words, the structural relationship between a verb and its arguments dictates factuality. Kitagawa et al. utilized deep modality analysis on Twitter streams to evaluate predicate-argument structures. By mapping how verbs modify nouns within a sentence structure, their system classifies events into distinct realis categories (e.g., happened, scheduled, or hypothetical), turning disease tracking from a noisy keyword-matching problem into a precise factuality assessment.

3. Implicature and Verb Signatures

Certain verbs inherently project or cancel the reality of their complement clauses. For example, "managed to escape" implies the escape happened, while "failed to escape" implies it did not. de Marneffe et al. and Klenner et al. leveraged these truth-conditional implications and implicature signatures to determine if an event was realized, prevented, or merely reported.

The Gatekeeper Architecture

Why does separating factuality from event extraction work so well? The core intuition lies in an architectural separation of concerns. Instead of forcing a single neural network to simultaneously learn semantic event types (e.g., Marry, Die, Infect) and their complex modal contexts, pipelines implement factuality detection as an explicit gatekeeper.

[ Raw Text Input ] 
        │
        ▼
┌────────────────────────────────┐
│   Factuality Gatekeeper Layer  │ ──► [Filter Out: Wishes, Conditionals,
└────────────────────────────────┘      Hypotheticals, Speculations]
        │
        ▼ (Only Factual Assertions)
┌────────────────────────────────┐
│ Downstream Extraction Engines  │ ──► Precise Timelines, Negation Scopes,
└────────────────────────────────┘      & Sentiment Inferences

Velldal et al. demonstrated the power of this approach by introducing a binary classification stage that filters out non-factual linguistic contexts (like conditionals, imperatives, or wishes) before deep event extraction or negation resolution takes place. When a system attempts to resolve negation without this gatekeeper, it frequently misclassifies hypothetical premises as negated facts (Lapponi et al.). By isolating the factual assertions first, downstream models only operate on grounded text, drastically boosting precision.

Why Factuality is Essential for Downstream Logic

Isolating asserted events isn't just about cleaning up data; it is a fundamental prerequisite for high-level semantic reasoning across several domains:

  • Temporal and Causal Reasoning: If a model cannot distinguish between a factual event and a hypothetical one, it will attempt to construct illogical links. As Wang et al. point out, establishing temporal or causal relationships between events breaks down entirely if event modality is ignored. For timeline generation, Kessler et al. and Batista-Navarro showed that restricting annotations and extractions strictly to confirmed historical milestones prevents timelines from being cluttered with unfulfilled plans or speculative future dates.
  • Sentiment and Stance Inference: Our emotional and logical reactions to text depend entirely on whether an event actually happened. Klenner et al. showed that sentiment reasoning is gated by factuality: an entity only truly becomes a beneficiary or a victim if the positive or negative event is confirmed to have occurred. If a company "hopes to increase profits," the sentiment does not transfer to the entity in the same way as "increased profits."
  • Fair Evaluation: Finally, modality filtering fixes how we measure NLP progress. Cai et al. revealed that classic event extraction benchmarks often penalize zero-shot models for ignoring non-asserted events (like rumors or threats) that are irrelevant to real-world applications. Filtering evaluation datasets to focus exclusively on grounded, asserted events aligns performance metrics with actual utility.

Key Foundational Work

Detecting Experiences from Weblogs (2010) — contributed 2 concept elements to this direction.
Finding Contradictions in Text (2008) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 26 papers
  2. Version 1 · 2026-10-04 · 46 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_linguistic_modality_filtering_bec7eacd,
  title = {linguistic modality filtering},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/linguistic-modality-filtering/rdr_bec7eacdc39c41acbff2251e3e120b20},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_linguistic_modality_filtering_bec7eacd_v2,
  title = {linguistic modality filtering},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/linguistic-modality-filtering/rdr_bec7eacdc39c41acbff2251e3e120b20/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake