Introduction
In natural language processing, a significant challenge lies in how machines construct a coherent understanding of events as they unfold across a discourse. Information regarding participants, actions, and contexts is rarely confined to a single sentence. For instance, in the sequence, "Pete bought a car. The salesman was a real jerk," a human reader immediately identifies the "salesman" as the source of the "buy" event. However, traditional computational models often struggle with this connection because there is no explicit syntactic link between the two sentences.
This paper proposes an extension to Discourse Representation Theory (DRT) that enables incremental event building. By treating "open" thematic roles—such as an unstated seller in a purchase event—as active anchors for future information, the authors provide a formal framework for resolving complex references that span multiple sentences.
The Three-Tiered DRT Framework
The authors build upon the foundations of DRT, a framework designed to model how meaning accumulates in a sequence of sentences. They introduce a novel three-tier structure for Discourse Representations (DRs) to manage entities and their relationships:
- Available Referents (AR): This tier contains a list of entities (people, objects, events) that have already been explicitly identified and are available for future reference.
- Unbound Referents (UR): This serves as a "buffer" for items that require a referent but have not yet found one. Crucially, the authors place implicit event roles here. When a "buy" event is introduced, roles such as "Source" (the seller) and "Exchange" (the price) are added to the UR tier, even if they are not explicitly mentioned in the text.
- Constraints on Referents (CR): This tier holds the logical conditions and properties defining the entities (e.g., "X is a car," "E is a buying event").
By placing unfilled roles in the UR tier, the system creates a "hook" that subsequent noun phrases can latch onto. When a new entity appears, the system searches the UR tier to determine if it fits an open role of a previously established event.
Mechanisms of Discourse Cohesion
The research highlights several key insights into how these open roles facilitate cohesion:
Open Roles as Cohesion Devices
Unfilled thematic roles are not merely missing data; they are active participants in the discourse. They tie sentences together by "attracting" new participants introduced later in the text. This mechanism allows the model to bridge the gap between disparate sentences without relying solely on syntactic control.
Definite NPs vs. Pronouns
A significant finding of the study is the differing behavior of definite noun phrases (NPs) and pronouns. Definite NPs (e.g., "the book," "the salesman") are particularly effective at filling open roles because they often focus on sub-components of an event. In contrast, pronouns (e.g., "he," "it") generally require a previously mentioned, "phonologically realized" entity to point to. For example, while "I was attacked. The man was huge" is perfectly coherent, "I was attacked. He was huge" feels slightly disconnected because the "man" was never explicitly introduced as a referent before the pronoun was used.
The Impact of Nominalization
The authors also observe that nominalization—turning a verb into a noun (e.g., "to investigate" vs. "an investigation")—alters the accessibility of an event's roles. Nominalized events appear to make it easier for subsequent sentences to refer back to internal details, providing a more stable anchor for role-filling.
Results and Prototype Evaluation
The paper presents a qualitative prototype model for event building that successfully demonstrates a unified mechanism for handling both "obligatory control" (grammar-based links) and "pragmatic control" (links based on world knowledge).
The model's efficacy is demonstrated through the resolution of "hard cases." For example, in a dialogue where a nurse provides advice to a patient's mother, the system can infer that the "Agent" of a "running" event is the patient, even if the advice was spoken to the mother. By following a chain of thematic role assignments, the system maintains a coherent representation of the event participants across the conversation.
Limitations and Future Directions
While the proposed framework offers a robust logic for event building, several challenges remain:
- Lexical Requirements: The system requires a comprehensive "lexicon" of events. It must "know" that a "buy" event expects a seller and a price. Constructing such a library is a significant undertaking.
- Role Predictability: Determining which roles are "highly expected" enough to be posted to the UR tier is an open question. Over-posting could lead to computational inefficiency, while under-posting might result in missed connections.
- Pragmatic Complexity: While the model handles some inference, complex human communication involving deep layers of social context still requires more sophisticated pragmatic reasoning than the current DRT extension provides.
Conclusion
By extending DRT with a dedicated tier for unbound referents, this research provides a path toward more human-like discourse processing. Treating implicit roles as active hooks allows for a more fluid and incremental construction of meaning, bridging the gap between sentence-level syntax and discourse-level semantics.
Summary updated