Modeling Event Modality and Realis Status

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 246 papers66 conceptsfrozen 2026-10-04

Traditional Information Extraction (IE) systems often suffer from "tunnel vision." When a model encounters a sentence like "The company might file for bankruptcy," it often extracts a Bankruptcy event, ignoring the crucial word "might." This failure to distinguish between a realized event and a mere possibility leads to "speculative noise" in timelines and knowledge bases.

To solve this, researchers have moved beyond simple event detection to model modality and realis status—the linguistic features that signal whether an event actually happened, is merely planned, or is explicitly denied.

The Realis Spectrum

The core of this research direction is the classification of "event nuggets" into semantic categories that describe their relationship to reality. As established in the Overview of TAC KBP 2015 Event Nugget Track by Teruko Mitamura and Zhengzhong Liu, events are typically categorized into three "realis" values:

  1. Actual: The event occurred in the real world (e.g., "The protest happened yesterday").
  2. Hypothetical: The event is a possibility, a wish, or a future plan (e.g., "They may protest").
  3. Generic: The event is a general statement or a recurring phenomenon rather than a specific instance (e.g., "Protests are common in summer").

Determining this status is often harder than identifying the event itself. Vincent Ng and the UTD team demonstrated in UTD’s Event Nugget Detection and Coreference System at KBP 2015 that while models can identify event triggers using standard sequence labeling, they require specialized classifiers (like SVMs or deep neural networks) to interpret the nuanced modal context that defines an event's reality.

Identifying Factive and Modal Triggers

Why do humans find this distinction easy? We look for specific linguistic "triggers." Marie-Catherine de Marneffe, Anna N. Rafferty, and Christopher D Manning argue in Finding Contradictions in Text that a system must evaluate the truth-conditional implications of verbs and modifiers.

For instance, "factive" verbs like know or realize imply the truth of their complement, while "non-factive" verbs like believe or hope do not. Similarly, modal operators like "should," "must," or "might" shift an event from the realm of fact to the realm of possibility. Ellen Riloff and Janyce Wiebe showed that by first classifying sentences as subjective or objective, systems can filter out metaphorical or hypothetical language, significantly improving the precision of extracted facts (Exploiting Subjectivity Classification to Improve Information Extraction).

Temporal Status and Intent

The distinction between "what is" and "what will be" is particularly vital for building timelines. In Distinguishing Past, On-going, and Future Events: The EventStatus Corpus, Ruihong Huang and Dan Jurafsky highlight that future-oriented statements—like calls for action or planned gatherings—are often semantically similar to past events but have entirely different implications for real-world monitoring.

This is especially true in social media analysis. Mohammad Akbari found that wellness event detection (e.g., "I went for a run") requires a "factuality filter" to prevent the misclassification of intentions (e.g., "I really should go for a run") as completed behaviors (From Tweets to Wellness: Wellness Event Detection from Twitter Streams).

Why Modality Matters for Reasoning

If a system cannot distinguish between a fact and a rumor, its downstream reasoning will fail. Haoyu Wang 0005 and Dan Roth argue in Event Semantic Classification in Context that modeling these semantic dimensions acts as a logical constraint. For example, a "negated" event cannot cause a "factual" one.

By explicitly modeling event modes, systems can:

  • Reduce Noise: Rémy Kessler and Xavier Tannier used modality analysis to filter out uncertain events when building thematic timelines, ensuring only salient, factual dates were included (Finding Salient Dates for Building Thematic Timelines).
  • Improve Sensitivity: In clinical settings, the NEAR framework by Namrata Nath uses modality detection to distinguish between a "proposed" treatment and a "completed" one, a distinction that is literally a matter of life and death.
  • Enhance NLI: Yibo Hu 0002 and Latifur Khan demonstrated that generating specific hypotheses for different temporal and modal states allows Natural Language Inference (NLI) models to better differentiate between material actions and verbal intentions (Leveraging Codebook Knowledge with NLI and ChatGPT for Zero-Shot Political Relation Classification).

Ultimately, this research direction shifts NLP from merely identifying what is being discussed to understanding the certainty and reality of the discussion itself.

Papers in this version

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 1, frozen text):

@misc{lacuna_modeling_event_modality_and_realis_statu_bec7eacd_v1,
  title = {Modeling Event Modality and Realis Status},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/modeling-event-modality-and-realis-status/rdr_bec7eacdc39c41acbff2251e3e120b20/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}