Introduction
Traditional lexical resources, such as WordNet, are effective at identifying synonyms (e.g., "car" and "automobile") and hypernyms. However, they struggle to capture referential variation—the phenomenon where different expressions are used to refer to the same entity or event within a specific narrative context. For instance, in a news report, a person might be referred to as a "suspect," then a "gunman," and eventually a "convicted murderer." While these terms are not synonyms in a dictionary sense, they are coreferential within the discourse. ReferenceNet is proposed as a semantic-pragmatic extension to WordNet designed to map these shifting roles and perspectives.
Methodology: The ReferenceSet
The core unit of ReferenceNet is the ReferenceSet, a collection of words and synsets used to refer to the same individual or event within a specific topical domain, such as gun violence. The authors explore two methodologies for constructing these sets:
- Text-to-Data (Bottom-Up): This approach utilizes semantic parsing and coreference resolution on raw text. While intuitive, it is computationally expensive and prone to "drifting," where distinct events are erroneously merged.
- Data-to-Text (Top-Down): This approach grounds text in structured databases of real-world incidents (a "microworld"). By matching news articles to specific incident reports (e.g., from the Gun Violence Archive), the system can reliably extract various referential expressions for the same "ground truth" entity.
The system employs the NewsReader pipeline, incorporating Word Sense Disambiguation (WSD) and WordNet similarity measures to aggregate mentions into dominant synsets.
Experimental Setup
The study focused on the domain of gun violence, utilizing structured data from the Gun Violence Archive (GVA), Fire Incident Reports, and the Aviation Safety Network. Evaluation was conducted on a manually annotated corpus of 38 news articles covering 20 specific incidents from the GVA, serving as a gold standard for identifying reference relations.
Results and Analysis
The performance of the system was evaluated in two settings: without-i (standard text processing) and with-i (incident-aware/data-to-text).
As shown in Table 2, the incident-aware approach (with-i) provided a significant boost in precision. For mentions, precision increased from 35.57% to 59.80%. This suggests that grounding text in structured data effectively mitigates the "drift" associated with purely bottom-up methods. However, recall remained relatively low, hovering around 20% for mentions.
Error analysis revealed that the low recall is primarily due to the limitations of current predicate detectors. These tools often fail to identify non-verb references, such as adjectives (fatal, deadly) or nouns (tragedy, incident), which are frequently used by human authors to refer to events.
Conclusion and Future Work
ReferenceNet demonstrates that reference is a pragmatic phenomenon that often cuts across standard semantic hierarchies. While the "Data-to-Text" method significantly improves precision, future work must address the recall gap by developing better methods to capture "hidden" predicates in adjectives and nouns. Additionally, the authors plan to expand ReferenceNet to include entity-level references (people and places) by leveraging semantic matches of noun phrases and roles across diverse topical domains.
Summary updated