VerbNet Representations: Subevent Semantics for Transfer Verbs

Susan Windisch Brown, Julia Bonn, James Gung, Annie Zaenen, James Pustejovsky, Martha PalmerACL 2019

TL;DR

This paper introduces a significantly revised version of VerbNet, an English lexical resource, featuring enhanced semantic representations for verbs. These new representations offer a more granular and dynamic view of event sub-structures, particularly for transfer verbs, enabling more precise computer planning and reasoning from natural language.

What Problem Does This Paper Address?

While deep learning has advanced many NLP tasks, precisely tracking event sequences and participants in discourse remains a challenge. This is crucial for applications requiring nuanced understanding, such as robotics or human-avatar interactions, where inferring implicit information is necessary. Existing lexical resources, like VerbNet, provide valuable semantic information, but previous representations lacked the granularity to capture the dynamic changes in participant states and relations over time, especially for complex events like transfers of possession or information. The paper aims to bridge this gap by providing more explicit, computationally tractable representations that capture these dynamic event structures.

Methods

The core of this work is the revision of VerbNet's semantic representations, drawing inspiration from the Generative Lexicon's dynamic event structure. Key methodological shifts include:

  • Subevent Structure: Moving beyond a simple Start-During-End model, the new representations allow for an indefinite number of subevents, enabling precise tracking of participants across time. This is achieved through predicates that explicitly indicate temporal and causal relations between subevents (e.g., cause(ei, ej), co-temporal(ei, ej)).
  • Dynamic Event Model: The representations are aligned with the Generative Lexicon's Dynamic Event Model, which explicitly labels transitions and the attributes being modified (e.g., location, state).
  • Explicit Opposition: Negated predicates are used to explicitly represent oppositions inherent in change events (e.g., alive(e1, Patient), ¬alive(e2, Patient) for death).
  • New Predicates: New core predicates like has_possession(e, [slot-1], [slot-2]) and transfer(e, [slot-1], [slot-2], [slot-3]) for possession, and has_information(e, [slot-1], [slot-2]) and transfer_info(e, [slot-1], [slot-2], [slot-3]) for information, are introduced. These predicates have more specific argument slots and formal definitions that apply across classes.
  • Refined Role Handling: The representations now allow for more flexible assignment of thematic roles (e.g., Agent can function as Source or Recipient) by letting the Agent role stand in for whichever role it overlaps, eliminating the need for redundant predicates.
  • Automatic Parsing: The paper also announces the release of a semantic parser that predicts these updated VerbNet representations from natural language input, facilitating immediate use.

Datasets

The paper describes the VerbNet resource itself, which has been expanded to include thousands of verb senses. The authors do not explicitly mention training or evaluation datasets for the parser in this paper, but refer to previous work and existing corpora for PropBank semantic role labeling (e.g., OntoNotes, English Web TreeBank, BOLT corpus).

Key Insights and Ideas

  • Granularity is Key for Dynamics: The most significant insight is that capturing the dynamic nature of events, especially transfers, requires a much finer-grained representation of subevents and participant states over time than previously available.
  • Explicit is Better than Implicit: Computers need explicit statements to track changes. For instance, inferring that an entity is no longer at its initial location after motion begins is not automatic for machines; the new representations make this explicit.
  • Unified Framework for Transfer: The paper demonstrates how a consistent representational framework can be applied to both "transfer of possession" and "transfer of information" events, highlighting their structural similarities while accommodating their distinct semantic nuances.
  • Agent Flexibility: Realizing that the "Agent" role can often subsume other roles like "Source" or "Recipient" simplifies representations and better reflects natural language usage.
  • The Power of Opposition: Explicitly representing oppositions (e.g., alive vs. not alive) is crucial for accurately modeling change-of-state events.

Key Results

The paper announces the release of a new, significantly revised version of VerbNet with enhanced semantic representations. While quantitative results for the parser are not detailed, the primary outcome is the availability of a more expressive and consistent semantic resource. The paper provides numerous examples illustrating how these new representations can capture complex event structures for transfer verbs, offering greater clarity and flexibility compared to previous models.

Limitations and Future Work

  • Parser Sophistication: The authors acknowledge that their automatic VerbNet parser, while functional, represents a "simple baseline approach." They suggest that a more sophisticated, unified approach to VN semantic parsing is a direction for future work.
  • Figurative Language: While the paper focuses on literal uses of verbs, it notes that transformations or re-interpretations of the semantic representations could potentially be used for figurative language.
  • Shared Ownership: The new representations are noted to allow for coverage of shared ownership of possessions, should VerbNet expand in that direction.
  • Temporal Sequencing in Two-Way Transfers: For two-way transfers of information, the subevent ordering is changed so that the state resulting from one transfer_info occurs before the next begins, with the repeated_sequence predicate used to capture this turn-taking. This is a specific handling for these complex interactions.

Summary updated

BibTeX
@inproceedings{brown2019verbnet,
  author = {Susan Windisch Brown and Julia Bonn and James Gung and Annie Zaenen and James Pustejovsky and Martha Palmer},
  title = {{VerbNet Representations: Subevent Semantics for Transfer Verbs}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure